Listen to the session or watch below

The US has spent billions building a “virtual wall” of surveillance towers along its southern border over the past 25 years, promising they will help detect and apprehend border crossers and save lives. But a groundbreaking investigation by MIT Technology Review has documented over a thousand people who moved through areas watched by these towers without being reached or apprehended, and who ultimately died there. Some were even under the watch of newly installed AI-powered towers designed to spot people automatically.

Our findings reveal a humanitarian crisis more visible than previously known, and repeated failures of the virtual wall’s basic security promise. Watch a conversation examining the failures of border surveillance technology and uncovering the stories of the people who die in the borderlands.

Speakers: Mat Honan, editor in chief, James O’Donnell, senior AI reporter, and Eileen Guo, senior features and investigations reporter

Recorded on September 28, 2026

Related Stories

• The US spent billions on border surveillance. Why can’t it catch people before they die?
• How we made the first comprehensive map of deaths along the US border’s “virtual wall”
• 4 ways to address the failures we found along the US border’s “virtual wall”

It’s been a busy few months for AI hype. At the end of April, Anthropic claimed that its model Claude Mythos is better at finding software vulnerabilities than most security experts. Then we had the OpenAI–Hugging Face hacking incident, after which Anthropic (proudly) and Meta (reluctantly) disclosed similar incidents involving their models.

This was followed by Anthropic’s claim that one of its models had made a mathematical breakthrough; soon OpenAI claimed a mathematical breakthrough of its own. Most recently, Anthropic engineer Jacob Coxon went viral announcing his departure from the company, claiming that it and OpenAI are “racing straight towards self-improving superintelligence and gambling with our lives.”

Each of these events was mostly covered breathlessly by the press, often repeating the companies’ anthropomorphizing framings—which are designed to portray their software is not only powerful but incipient “artificial general intelligence.” So what is really going on? Are we witnessing a massive, civilization-changing set of technological breakthroughs, or is this marketing?

In all these incidents, massive fanfare from the companies (presented as mea culpas in illicit hacking cases) is accompanied by intense press coverage. Once there is time for experts in the relevant fields to examine what happened, a very different story emerges, but one that gets less media attention.

Regarding the “hacking” incidents, cybersecurity experts say the story is more about OpenAI’s negligence and failure to adopt basic, established security practices than about “models gone rogue” or “AI agents creating civilizations.” 

As for the mathematical results, mathematicians who were initially “stunned” by OpenAI’s press release saying that its latest chatbot, Astra, solved problems that “have been open and seen no progress on the main result for at least a decade” later realized that the results weren’t as “novel as first appeared.” Since then, mathematicians have accused the company of research misconduct and plagiarism, and they’ve reiterated that Astra didn’t make a “profound intellectual leap.” Just weeks later, OpenAI claimed its own mathematical breakthrough. Two days before, Tristan Buckmaster, a math professor at New York University’s Courant Institute, published a bombshell statement suggesting that OpenAI had stolen other people’s work and improperly attributed it. 

Claims of incipient, dangerous superintelligence are not based in good scientific or engineering practice. Rather, they are narratives based in ideologies of transhumanism, eugenics, and wishful thinking about imagined future digital humans.

It’s worth thinking about why there is so much attention on computer programming and math as fields in which to apply large language models and related technology. Not only are they often elevated as the pinnacle of human intellectual achievement, but they involve problems where answers, once suggested, can be verified. The former property helps AI hype mongers sell the idea that they are building everything machines. The latter makes math and coding problems easier to tune systems for, since system output (sequences of likely words or pieces of computer code) can be evaluated without having to pay data workers to look at and annotate each one.

Mathematicians in particular have warned against corporations using their field in this way. A statement signed by hundreds of them says there is “currently a strong commercial incentive on the part of the technology industry to overstate the capabilities of their products” and asks policymakers to “consult with experts, including mathematicians, in forming policy decisions rather than relying on press releases or popular reporting of mathematical results.”

We echo this call and note that the illusion of speed and urgency promulgated by the tech companies is also a ploy to misdirect both policymakers and the public. Unfortunately, it sometimes works, such as with Senator Bernie Sanders’s well-meaning but ultimately misguided proposed legislation to prevent the development of “artificial superintelligence.”

Describing them as “superintelligence” or “rogue models” ascribes agency to products rather than to the companies building them. This framing markets these companies’ products as “superhuman” and, at the same time, helps the companies evade accountability for their actions. 

Instead of OpenAI being prosecuted for creating malware that hacked another company, press releases, news outlets, media personalities, and lawmakers refer to “rogue models” as if they acted on their own. Instead of researchers being questioned about their companies’ habit of plagiarizing academics’ work or using customer data to train models without consent, the public’s imagination is redirected to fears about what the future might hold upon the arrival of fictional superintelligent machines.  

The AI industry has even suggested that popular, bipartisan anti-data-center activism is a “distraction” from attempts to regulate the impending, scary, “superhuman” machines these companies are building. According to the AI industry, we should be more worried about a fictional machine god than about the climate catastrophe that these data centers exacerbate, the asthma suffered by those living near them, the rising electricity bills of the public subsidizing them, or the water that is redirected to cooling them. 

We know better than to make decisions based on marketing and better than to capitulate to corporate pressure to make those decisions quickly. Wise decision-making, by policymakers and communities, demands time to hear from independent experts and contextualize corporate claims. The best possible outcome from this summer of hype is that policymakers and the public at large learn to take a breath, hold onto our skepticism, and recognize this kind of hype for what it is the next time it comes around.

Timnit Gebru is executive director of DAIR and author of the forthcoming book Deep Unlearning: The Radicalization of a Tech Idealist, which is available for preorders now and set to publish on February 16. Emily M. Bender is professor of linguistics at the University of Washington and coauthor of The AI Con.

Our 15-month investigation into death and surveillance along the US-Mexico border began with a simple question: Why did so many people die near government surveillance towers meant to help track and apprehend them?

The result of our investigation is the first comprehensive map and analysis of deaths near border surveillance towers. Here’s how we built it, what decisions we made along the way, and what the data can and can’t tell us.

Our data

The investigation relied on knowing where migrants have died and where and when US Customs and Border Protection towers were installed. We analyzed cases dating back to 2015, allowing us to cover different border policies, presidential administrations, and tower technologies.

Migrant deaths

The US-Mexico border has been called the world’s deadliest land border, and Border Patrol estimates that more than 10,000 people have died during crossings since 2000—a figure that’s widely considered an undercount. But the amount of information available publicly on where and when these people died, and who they were, varies widely. Some remains are never discovered, and for those that are, there’s no national protocol for how the records are handled.

We considered a case for further analysis if we could confirm three basic things: whether the person was believed to have been crossing the border, where the remains were found, and roughly when the person died. 

We created our dataset by merging existing ones that had been compiled and shared by other organizations, including No More Deaths, Humane Borders, and the Electronic Frontier Foundation, with new records we obtained ourselves from more than a dozen agencies.

Public records requests in Texas

Texas was the biggest missing link in most existing databases of migrant deaths. Unlike Arizona, it has no initiative to share records online, and unlike New Mexico, it leaves individual counties to handle their own death investigations. Some records for those individual counties existed, compiled by a handful of dedicated researchers, but nothing was comprehensive or current. 

We identified 17 counties in Texas that would be relevant to our investigation: Culberson, Jeff Davis, Presidio, Terrell, Val Verde, Kinney, Maverick, Dimmit, Webb, Zapata, Jim Hogg, Starr, Hidalgo, Cameron, Kenedy, Brooks, and Duval. There are other counties in which migrants have died, but they do not have areas covered by surveillance towers and were therefore excluded.

In most of these counties, locally elected administrative judges called justices of the peace collect the most comprehensive information on local deaths. But records requests submitted to these individual judges can stall for years (indeed, some of our requests to justices of the peace remain unfilled more than a year after we filed them). Reports from justices of the peace sometimes leave out the scene or location details we needed; some justices never even visit the place where a migrant died, instead formally declaring the person deceased over FaceTime with the first responders on scene.

MIT Technology Review instead filed records requests to the sheriffs of these 17 counties beginning in July 2025. These agencies are often first on the scene when a death is discovered, and their reports can have more detailed information on where remains were found. But relying on sheriffs’ records means our analysis undercounts migrant deaths in Texas, as it doesn’t include any deaths handled by local or state police.

After we filed a request, it took periods of near daily calls to the offices to receive the records, if we received them at all. Some were provided with no charge. Other counties charged between $300 and $1,300 to fill our request. Counties warned that their records were incomplete, with an unknown number lost during a move, damaged, destroyed, or misplaced during transitions between sheriffs.

Those delays were compounded because most agencies do not record whether a person is believed to have been crossing the border when they died. That’s despite provisions in some counties that allow officers to track far more granular details about other situations; Zapata County reports have a checkbox to indicate whether jewels were stolen in a burglary, for example, while Cameron County reports have one for whether a burglar entered through a chimney. Without a way to readily identify migrant deaths, offices had to pull records by hand, making the process slower, more costly, and more prone to error.

We received usable records from 14 of these counties (records have yet to be received from Dimmit or Duval, and Maverick’s office charged a per-record fee that was cost prohibitive).

Agencies generally sent us police reports for each individual case, amounting to over 4,000 pages of records in total. Some were handwritten documents. 

Except for the reports from Kenedy, Webb, and Hidalgo Counties, we pulled out the relevant information by hand. For those three counties, which sent large volumes of records, we used Anthropic’s Claude, accessed via API, to inspect each case report and pull out coordinates of the spots where remains were found; then we checked batches of those cases by hand to verify the AI’s accuracy. Agencies sent us records on a rolling basis from July 2025 to February 2026, and most did not specify the date through which their records were current.

We generally trusted that agencies sent us cases they believed involved migrant deaths, since they had the most information about each case. But some records clearly suggested otherwise (indicating, for example, that someone died at home), and we excluded those cases.

Several hundred cases were removed from our analysis because they didn’t include enough information for us to be certain where the remains were found. In fewer than 100 cases, coordinates were not listed but the description was specific enough for us to locate a point within 0.25 miles of where they were found—so we kept those in. 

Compared with agencies in other states along the border, those in Texas provide far fewer details regarding how long ago someone may have died before remains were found or how far the decomposition of those remains had progressed. Some reports included this information, but many didn’t. While autopsy records sometimes have more details, many migrant autopsies in south Texas are done through the Webb County Medical Examiner, which charges a fee for each report. That was cost prohibitive given the scale of our analysis.

MIT Technology Review consulted with Greg Hess, director of Arizona’s Pima County Office of the Medical Examiner, on a framework for translating descriptions of the scenes, bodies, and causes of death available in many of these records into rough estimates for when a person may have died (more detail on this process is provided below).

In total, we included more than 1,500 cases from our Texas records requests in our final analysis.

Accessing Arizona records directly

The Pima County Office of the Medical Examiner (PCOME) in Arizona handles death investigations for all border counties in the state except Yuma (we sourced Yuma’s records from the organization No More Deaths, as detailed below).

The office determines which cases it believes to be border crossers. It publishes a data portal for these migrant deaths that’s updated monthly and shares its data with the organization Humane Borders, which then publishes it in a map and database. We included the data from 2015 through April 2026.

Humane Borders notes for each case how precise the location description is. We included only cases with GPS coordinates precise to within 300 feet. Collecting these coordinates has been standard practice since around 2012. The records also include an estimate of when the individual died, ranging from less than a day to at least six to eight months before remains were found.

In total, we included more than 1,700 cases from Humane Borders in our analysis.

Data collected by No More Deaths

To obtain data on deaths in California, New Mexico, Arizona’s Yuma County, and El Paso and Hudspeth Counties in Texas, we turned to the nonprofit organization No More Deaths, which has tracked migrant deaths since 2004. It publishes information about its records requests and its methodology. 

Many of the cases in its database include GPS coordinates, sometimes with notes that the location is only approximate—for example, accurate to within half a mile or one mile. We included only cases where the location was known to within a quarter-mile.

No More Death’s database includes, when available, information on the level of decomposition in which each set of remains was found. As with cases in Texas, MIT Technology Review consulted with PCOME’s Greg Hess on a framework for translating this information into rough estimates for when a person may have died. 

Here’s where the No More Deaths data used in our analysis comes from:

• In California, No More Deaths requests records from the San Diego County medical examiner and the Imperial County coroner. Its latest data for Imperial County is from December 2025, and its latest for San Diego County is from October 2025. 
• In Yuma County, Arizona, No More Deaths requests records from the Yuma County medical examiner. The latest data is from September 2025.
• In New Mexico, the organization requests data from the state’s Office of the Medical Investigator. Unlike most states, New Mexico has a statewide medical examiner, allowing No More Deaths to obtain migrant death data for the entire state. The latest data is from July 2025.
• For Hudspeth County, Texas, No More Deaths requests records from justices of the peace in Districts 1 and 2. The latest data is from December 2023.
• For El Paso County, Texas, records were from the El Paso County Office of the Medical Examiner. The latest data is from September 2025.

In total, we included nearly 1,000 cases from No More Deaths in our analysis. 

Total cases reviewed

When accounting for all our data sources and excluding deaths without specific enough records to determine where remains were found, we analyzed a total of nearly 4,000 cases. 

Other notes 

We generally assume that a person died where their remains were found. That may not necessarily be true: Remains can be moved by other people, flowing water, or animals. But medical examiners and researchers we consulted said such cases are rare.

We excluded cases in which the virtual wall did not appear relevant. For example, we removed cases involving people who died inside Border Patrol stations or in car crashes during pursuits. 

The deaths we analyzed reflect a variety of causes that may seem at first like very different sorts of surveillance failure. If someone slowly died of dehydration within view of a camera, for example, agents would have had far more time to intervene than they would have in a case where someone drowned quickly in a river or canal. But we included all these deaths because, in each one, the virtual wall could have prompted a response from agents.

Surveillance towers

Including a tower in our analysis required knowing three things: its location, its type (so that we could judge how far it is supposed to see), and the time it was installed. 

Locations

The Electronic Frontier Foundation has been tracking and mapping CBP surveillance towers since 2023, using on-the-ground reporting, satellite imagery, and government documents obtained through public records requests. Its tally is an undercount; government records suggest that about 800 towers are currently deployed, while EFF has logged about 600. Our analysis therefore misses deaths that occurred near towers not yet catalogued by EFF.

MIT Technology Review worked closely with EFF’s director of investigations, Dave Maass, throughout this project. Our tower data began with the database published by EFF in April 2026, but that map did not include towers that have been removed. We consulted EFF for the details on these former towers and added them to our database. We generally use EFF’s reference numbers for individual towers, but we changed some to ensure that each tower has a unique identifier.

The locations are identified by GPS coordinates. We verified each tower’s coordinates by analyzing satellite imagery.

Tower types

Our analysis focused on towers of three main types. Certain towers mapped by EFF don’t fall into any of those categories, so we excluded them because it’s difficult to obtain information about how far they are supposed to be able to see. This includes some towers that are installed at Border Patrol checkpoints or stations.

CREATIVE COMMONS

CENGIZ YAR FOR MITTR

To know how far the main tower types could see, we consulted materials from tech companies and EFF’s research, led our own review of government documents, and interviewed former Border Patrol agents and officials. The distances described are consistent across these sources. But they are advertised estimates, not contractual guarantees: Some government records redact more specific information about a tower’s surveillance range in particular locations, citing security concerns.

When towers were present

To know if a tower in place now was present when someone died, it’s essential to know not just where it’s located but also when it was put there. To figure this out, we used multiple sources of satellite imagery. 

Maass, at EFF, had previously analyzed some of the towers to determine when they first appeared. MIT Technology Review checked satellite imagery for these towers to confirm these timelines, and then checked for all the remaining towers. For each tower, we logged the date that it first appeared and, if it eventually disappeared, the last date it was present. We spent months meticulously checking when hundreds of towers appeared in multiple sources of satellite imagery to confirm there were no contradictions. 

This is easier for certain tower types than others; towers from Anduril, for example, tend to go up where there was no prior infrastructure, and it’s fairly easy to see how a patch of empty desert gave way to the new tower and its trademark solar panels. Other towers, like IFTs, were often built on or near existing structures. For these, we leaned more closely on EFF’s expertise in analyzing the imagery. 

When imagery wasn’t sufficient, we used Google Street View, government documents, or photographs collected by EFF to verify when towers went up.

One limitation is that historical imagery of towers along the border is not available for every day in the past. Especially for older towers or those in remote locations, images may have been taken months or even a year apart. A tower visible in May and again in December could theoretically have been removed and reinstalled in the intervening months. But most tower systems are permanent structures. And multiple Border Patrol agents told us that even towers that are meant to be mobile, like autonomous towers, are in practice very rarely taken down or moved (consistent with this, we observed that semipermanent fences were erected around the sites of many Anduril towers).

Another limitation is that even if a tower is visible at a particular time, that does not mean it was online. We can’t verify that it was functional, receiving power, or transmitting data. Our analysis therefore establishes when a tower was present, not whether it was operational.

Heights

We conducted a topographical analysis, detailed below, to see whether a tower’s view of a given person might have been blocked by the terrain. This required estimating how tall each tower is. 

We included a minimum and maximum height for each tower, based on research from EFF. In select cases where the surveillance system was not on a typical tower but mounted to an existing structure, like a water tower or building, we used satellite imagery to estimate its height. The higher a tower is, the farther it can see.

Our analysis

With the data we collected, we could begin to see which deaths happened in range of the virtual wall. But for a given match between a set of remains and a tower to count, it needed to satisfy two criteria:

1) Distance: The death must have happened within the published surveillance range of the nearby camera.

2) Timing: The tower had to be present when the person was estimated to have died.

Distance

Minho Kim, an outside collaborator who is now a postdoctoral researcher at Stanford with a focus on wildfire mapping, led the distance analysis. His software ingested our database of human remains and our database of towers. Then it created a row for each time a set of remains was found near a given tower, along with the distance between the two.

We then checked whether, for each match, the tower was actually present when the person died.

Timing

Sometimes, as with the Arizona cases, an estimate of when someone died is given in the original data; the medical examiner may note that a person died more than three weeks but less than five weeks before remains were found, for example. 

In other cases, we used the level of decomposition, the cause of death, or both to calculate a window of time in which the person would likely have died. Finally, we compared this window with the time when we determined the tower was built. 

This step took us many months and involved reviewing original police reports for hundreds of cases, inspecting photos of bodies, cross-referencing cases with news articles, interviewing medical examiners, and researching decomposition timelines. The resulting death windows are approximate, but they are the best estimates we could make from the information available in each case.

We assigned different confidence levels to cases based on how these dates lined up. If a person’s entire estimated death window fell well after a tower was first observed, we felt confident the tower was present when they died. If any part of the death window fell before the tower was first observed, we excluded the case.

Many cases involve skeletal remains. Remains decompose differently depending on location and season, but when records didn’t provide a specific estimate, we conservatively estimated that skeletal remains had been there for at least six months. For these and other cases without a specific estimate of when the person died, we used three years as a conservative cutoff: If a tower had been present for at least three years before the remains were found, we included the case, judging it unlikely that the person had died before the tower was installed and remained undiscovered that long. If the tower had been present for less than three years, we excluded it.

Viewshed

It is possible that someone who died within range of a tower could have been blocked from the tower’s view by terrain, vegetation, or buildings.

We estimated the impact of this issue with a technique called topographical analysis.

Kim, our collaborator, led this part of the analysis. Here’s how it works: Imagine a map of the border region divided into a grid, with each square having one value for that area’s elevation. Take a given match between a set of remains and a tower and draw a line between the tower’s estimated height and an estimated height of 5’ 6″ for the person who died. If the terrain along that path rises high enough to block the line between the two points, we estimate the tower’s view as obstructed. If it doesn’t, we estimate the view as clear.

It’s not perfect. In the elevation data we used, each square represents an area 30 meters wide, or about 98 feet, and is assigned a single elevation. That means smaller changes in the terrain can be lost, particularly in steep or rugged areas where elevation changes quickly. Our analysis is therefore less about re-creating exactly what a tower could see than providing a rough estimate of whether terrain may have blocked a tower’s view of the spot where someone died. 

We include results on our map whether a clear line of sight was estimated or not. Even if the location where someone died was blocked from a tower’s view, the person may have passed through areas visible to the tower before reaching that spot.

It’s much harder to estimate whether a line of sight was blocked by vegetation. We are less concerned about this limitation because virtual-wall towers use thermal imaging, which can detect body heat through light vegetation. Dense vegetation, however, can still obstruct their view.

Buildings pose a greater challenge. We could not reliably account for their locations and heights. Instead, we reviewed each tower’s surroundings and flagged towers located in or near urban areas. These notes warn readers that buildings may have limited what those towers could see.

Beyond the data 

The most fundamental question in each of these cases, and for the families of anyone who died, is whether the US government saw someone attempting to cross into the country and never intercepted them—or never saw them at all. 

This is all but impossible to answer with records or data—Border Patrol doesn’t disclose which alerts came in surrounding any given migrant death, and footage isn’t retained unless an internal investigation is triggered, which doesn’t happen unless a migrant dies while in custody.

MIT Technology Review instead conducted dozens of interviews with people who worked for Border Patrol, Customs and Border Protection, or the Department of Homeland Security or who served as presidential advisors. These conversations yielded previously unreported details about how agents use surveillance technology, turned up instances where it has underperformed, and underscored how seldom its effectiveness is evaluated internally.

We also sought out friends and family members of people who died near surveillance towers, using additional records, social media, and other research tools to find them. We approached these interviews using best practices for trauma-informed reporting.

Beginning in September 2025, MIT Technology Review requested interviews with current Border Patrol agents and leadership at CBP, and we asked to visit the border to understand the strengths and weaknesses of the latest autonomous tower program. Following delays, which were in part related to DHS funding lapses, in July 2026 we were told our media request had been approved at the local level in Big Bend, Texas. We wanted to visit this area because Border Patrol is expanding the number of Anduril towers there. But when the request went up to the DHS headquarters for a final sign-off, the agency denied it. Before publishing our investigation, we sent detailed questions to CBP, Anduril, General Dynamics and Elbit. CBP and Anduril responded, but did not substantively address our questions. General Dynamics referred us to CBP. Elbit did not respond to our questions.

MIT Technology Review thanks the following for their assistance with this project:

Minho Kim, Stanford University

Greg Hess, Pima County Office of the Medical Examiner 

Stephanie Leutert, University of Texas at Austin

Don White, Brooks County Sheriff’s Office

MIT Technology Review today published our investigation into how many people have died near the “virtual wall” of surveillance towers that the US government has installed along the US-Mexico border.

We found cases of people who walked undetected through areas surveilled by advanced, AI-enabled towers and later died nearby, where their bodies remained unnoticed for weeks or months. These cases reflect systemic flaws in the virtual wall: broken towers, algorithms that failed to detect people, and agents who simply did not respond to alerts.

Each of these deaths should have raised alarms inside US Customs and Border Protection (CBP) about the flaws in the agency’s surveillance network. This is particularly true for cases near the latest towers built by the defense company Anduril, which continues to win major contracts and is poised to benefit from the windfall of federal money slated for border security. The US is set to spend $1 billion to triple the size of the virtual wall by 2034.

Here are four ideas for what CBP should do next to address these failures. 

1. Conduct a comprehensive audit of deaths near the virtual wall

“If I were the federal government, I would do an audit, because this feels like it’s not working.” 

That was New Mexico state representative Sarah Silva’s reaction when MIT Technology Review showed her that in 2024 alone, more than 18 people had died near surveillance towers in just one small stretch of desert near her district.

An audit is necessary because, despite poring over thousands of pages of police reports and other documents about surveillance towers, we have only limited knowledge of how many towers are out there and where they’re located. 

CBP does not publish a comprehensive map of its virtual wall. We instead relied on mapping work led by the nonprofit Electronic Frontier Foundation and then used satellite imagery to verify where specific towers were and approximately when they were installed. 

We analyzed lots of towers—nearly 600. But that’s still far less than the 800 the government says operate today. And our estimates of when they went up are inexact. That means our tally of more than 1,050 people who have died near them is an undercount. And we don’t have a complete picture of where along the border the problem is worst.

An internal audit would not have such limitations. In theory, CBP tracks when each tower went up and when it was active. It has had the ability to audit their effectiveness—while quantifying how many people have died near them—since at least 2015, when it began installing the types of towers included in our analysis. But sources told us it has never done such an audit. Our findings make the need for one more urgent.

2. Fix the way the agency tracks migrant deaths and remains

If CBP is to analyze instances where people have died near its surveillance technology, it needs to know when and where those deaths happen. And the current system for tracking that data is beyond broken.

As we note in our methodology, there is no national agency that collects and publishes all available records on migrant deaths. We sifted through thousands of pages of records from dozens of agencies in Texas alone, and our picture is still incomplete owing to the lack of a shared protocol for documenting these deaths in public records. 

Border Patrol has had its own internal system, called the Border Safety Initiative Tracking System (BSITS), in place since 2007. But it wasn’t until 2021 that national rules were established for how different sectors should enter cases into the system. The next year, the Government Accountability Office (GAO) found that many cases still weren’t being tracked in this system, especially deaths that weren’t discovered by Border Patrol. In 2023, the GAO reported that Border Patrol has taken steps to fix that, but outside researchers consistently find the agency’s estimates to be undercounts even today.

In addition to tracking migrant deaths, CBP launched a program in 2017 to prevent them, called the Missing Migrant Program. In April 2025, the GAO found the program was improving the way it collected and tracked data about events like 911 calls and rescues. The problem, it said, was that there was no plan in place to use the data to measure whether the program was actually increasing rescues or reducing deaths, which are its primary goals. It made two recommendations for solutions, neither of which has been implemented.

3. Track cases when surveillance technology leads to an apprehension

In 2014, when the virtual wall was much smaller, the GAO suggested that Border Patrol agents log when a piece of technology helped them apprehend someone who was trying to enter the US illegally. It would take agents some extra time, but logging these assists would help Congress understand how often various surveillance technologies were leading to actual apprehensions. Once this data was collected, GAO added, the agency should measure whether its surveillance technologies were working. CBP agreed, and it started requiring agents to enter such information.

Then, in 2017, the GAO reported that the data being collected was inaccurate and unreliable; it found agents in Texas, for example, who were crediting a specific type of tower in apprehensions hundreds of times, even though no such towers existed anywhere in the state. Border Patrol again said it would improve; it agreed to train agents on the importance of collecting such information.

But years later, agents told us, it’s still not a priority. One former CBP official described the mandate as “spitting in the wind, to be honest,” and said the messy reality of apprehensions doesn’t translate into neat metrics. (The official, who requested anonymity, served during the Biden administration and wasn’t authorized to discuss sensitive issues.) 

And in August 2021, more than seven years after it agreed to do so, CBP still had not produced any analysis on the towers’ effectiveness (it finished one in 2022, the GAO notes, but the results are not public).

That means the virtual wall is growing faster than anyone is assessing its impact.

4. Investigate deaths as potential surveillance failures

CBP should also examine new deaths as they occur and determine whether they happened within sight of a tower or other surveillance technology. And for those that did, the agency should determine why those people weren’t spotted sooner.

Several agents and officials told us that if Border Patrol agents come across the body of someone presumed to have died while crossing into the US, there’s no requirement to check whether that person died near surveillance infrastructure and no process for doing so. At a minimum, that means potentially valuable evidence about the death goes unexamined. It also means that migrant deaths are not routinely treated as possible failures of the virtual wall.

CBP should automatically check every migrant death against its internal database of towers and investigate those that occurred within surveillance range to determine what the system detected and whether agents responded appropriately. Agents already log GPS coordinates when they apprehend a group or encounter footprints, among other events.

Without that check, law enforcement, medical investigators, and families may never learn that someone died near a surveillance tower—and that footage of the moments leading up to the death may exist. CBP typically deletes surveillance footage from the towers after 30 days unless an internal review is opened, which usually happens only for deaths in custody. At a minimum, flagging these cases could preserve footage for those investigating the death and give families a chance to request it.

Such reviews could also expose flaws in the virtual wall itself. Consider one location we identified in New Mexico, where several people died despite being in clear sight of an advanced AI tower. They died within the span of a few months, sometimes just hundreds of feet from one another. Investigating these cases might reveal, for example, that a camera repeatedly failed to detect or identify people in a particular area, or that agents were alerted and simply never responded.

Carrying out these postmortem inquiries would be the only way to answer the question that families, researchers, and volunteers raised most frequently in our investigation into the hundreds of migrants who died near towers: Did the cameras see them, and nobody came? Or did nobody see them at all?

When José Morales Bernal crossed the border into the United States on April 8, 2024, the day before his 32nd birthday, it should have triggered a chain of technological alerts and human responses. 

As he walked through the desert in southern New Mexico that morning, he was within range of three surveillance towers. Newly installed by US Customs and Border Protection (CBP), they were built by the defense tech company Anduril and equipped with cameras and AI to automatically detect and track people. They transmit live video to nearby control rooms and can send alerts to the government-issued smartphones held by agents in the area, prompting the closest available to respond.

These AI-enabled towers are meant to give greater visibility across the 1,951-mile southern border, freeing up border agents from having to spend hours staring into video monitors. They were installed in this particular place to spot border crossers before they reached the nearby town of Sunland Park.

If the system worked as intended, Morales should have been apprehended. If he needed medical help, agents were trained to provide it.

That didn’t happen. Despite the nearby surveillance towers, it was employees of the local landfill, rather than Border Patrol, who first spotted Morales that morning. At 1 p.m. the workers saw him again, now lying in the sand. At 4 p.m. the landfill workers saw that he had not moved and called Border Patrol. Agents arrived 45 minutes later. He was dead. When an agent then called 911 to report the body, he said it was “probably one of the migrants crossing through there,” seemingly unaware that Morales had been moving near the agency’s surveillance systems earlier that day. 

Morales had died just 360 feet from the closest surveillance tower. Two more towers stood watch to the east and the west. An autopsy later concluded that he had died of “environmental exposure.” 

The towers surrounding Morales were only the latest addition to the “virtual wall” the government has spent 25 years and billions of dollars building along the entire border, which also includes earlier generations of towers with more basic features, as well as blimps, drones, seismic sensors, and even tunnel-sensing robots. Together, all this technology provides “persistent surveillance” and “situational awareness” to help the Border Patrol quickly and accurately detect people crossing and, crucially, make sure agents are sent to intercept them. CBP has also credited it with saving lives.

But a first-of-its-kind investigation by MIT Technology Review reveals that deaths like Morales’s are startlingly common. We cross-referenced nearly 4,000 locations where human remains were found—drawn from records collected by nonprofit groups like No More Deaths and Humane Borders, as well as hundreds of records we obtained in Texas—with information on nearly 600 towers identified by the Electronic Frontier Foundation. We considered when towers were installed and when each person is estimated to have died, and combined this information with field reporting from the border. The result is the first comprehensive map and analysis of deaths near CBP surveillance towers. 

Our investigation shows that a humanitarian crisis at the border has unfolded in view of the government’s own cameras.

We found more than 1,050 people who died within range of border surveillance towers between 2015 and early 2026. These deaths are not failures of a few towers or technologies: We found deaths within the advertised range of nearly two-thirds of all the towers we analyzed. Our topographical analysis—which assessed the degree to which terrain might block a tower’s view of a particular death and its surveillance area in general—found that some have sight of as little as 10% of their advertised surveillance area, and yet we also found most deaths did not occur in towers’ blind spots. These deaths are not the result of legacy systems, as our estimate found more than 110 people have died within range of modern autonomous towers from Anduril, among the most advanced systems CBP has deployed, since 2021.

The overall picture reveals repeated failures of one of the virtual wall’s basic security functions, as CBP has described it in press releases: to effectively identify and locate migrants entering illegally into the United States. 

The year that Morales died, 18 other people died in that same stretch of desert in range of the three Anduril towers. Five were visible from the same surveillance tower closest to where his body was discovered.

One man, after walking a mile past the border and in range of two AI towers, dragged his 30-year-old brother into the shade when he began having trouble breathing, according to records we obtained from the medical examiner. The two were spotted by Border Patrol only when the man waved down a helicopter for help; by the time agents arrived, his brother was already dead. A 24-year-old woman died near another Anduril tower, where our terrain analysis showed it should have had clear sight of her location. Her body lay unnoticed for weeks; it was decomposed, and blistered by the summer heat, when it was spotted by agents patrolling the area. 

A few miles west, and a short drive from a Border Patrol station, four other AI towers stood watch. Near them, 10 more bodies were found that year, all in locations where at least one tower had a clear view.

In response to a list of questions, an Anduril spokesperson replied that once a tower is delivered, it is operated by CBP, and directed questions about specific incidents to the agency. The response noted that an incident occurring nearby does not mean the tower missed a detection and said actual surveillance ranges vary depending on terrain, physical obstructions, and the boundaries CBP sets for where the tower should look (CBP is able to set virtual boundaries on towers’ views for privacy and other reasons). The Anduril spokesperson also alleged inaccuracies in our reporting, given those boundaries and obstructions, but did not respond to follow-up questions on what was inaccurate.

The most pressing question in any death near the virtual wall is whether authorities knew someone was there and failed to reach them or weren’t aware anyone was crossing at all. Either is a system failure. And the deaths we found capture only the failures that left a trace—we don’t know how many people pass through undetected, or how many bodies remain undiscovered.

Interviews with more than 45 people—including current and former White House advisors and presidential appointees, Border Patrol agents, medical examiners, sheriffs, humanitarian volunteers, and employees of tech companies—showed that both types of failures are occurring: The technology is failing to detect, and agents are failing to respond. Both show the limits of throwing technology at a complex problem. 

The findings reveal previously unreported issues with the virtual wall, even as it continues to enjoy broad political support and a surge in federal spending. Border security hardliners have long seen it as another tool for stopping smugglers moving drugs or people, while others tout it as a cheaper alternative to a physical wall. In 2023, the government estimated that its plans for using the towers, which now number 803, would cost $6.2 billion over their lifespan. With the historic levels of funding it was awarded in 2025, CBP plans to spend $1 billion for 1,497 more towers by 2034.

But our reporting shows that CBP has done little to assess how its towers are working or how many people have died where they keep watch.

Officials who oversaw border security across the last four presidential administrations told us they believed deaths near the so-called virtual wall were either exceedingly rare or nonexistent. But our reporting shows that is not the case. None could point to any comparable analysis the government had ever conducted on its own. Former agents and officials also told us that when someone’s body is found, CBP does not formally investigate whether surveillance should have detected them or, if they were detected, why agents didn’t reach them before they died. 

In response to nearly 30 questions about our findings, which covered multiple generations of technology, Hilton Beckham, CBP’s assistant commissioner for public affairs, said, “Autonomous surveillance towers use artificial intelligence to detect and classify people, vehicles, and animals and alert Border Patrol agents to activity in monitored areas … ASTs complement physical barriers and other border security infrastructure by improving detection and situational awareness between ports of entry. CBP evaluates the technology based on its impact on detection, response coordination, agent safety, and mission outcomes.”

“I’m sure that the cameras and other surveillance assets do deliver … useful intelligence and enhance operational efficacy in some places, under some conditions, in certain circumstances,” says Geoff Boyce, an assistant professor of geography at University College Dublin who has studied surveillance technology used at the US southern border. “I’m also absolutely positive—because this has been the track record—it is not delivering the level of operational support, information, or efficacy that either the companies delivering these infrastructures or the Border Patrol and Department of Homeland Security claim.” 

We also reached out to multiple lawmakers from both parties with a summary of our findings. In response, Delia Ramirez, a Democrat representing Illinois’s 3rd district who sits on the House Homeland Security Committee, said, “AI-powered surveillance technologies are not making us safer. Yet DHS continues to spend millions of taxpayer dollars on these ineffective, negligent technologies, with no commitment to oversight or transparency … It is clear we must terminate CBP’s integrated surveillance tower program and dismantle DHS.”

Building the virtual wall

Since its earliest efforts to police the border with Mexico, the US government has faced the same basic challenge: How do you effectively secure nearly 2,000 miles of remote and rugged terrain? It has increasingly turned to technology for answers.

Gerardo Galvan joined Border Patrol in 1995, as the government was undertaking an unprecedented expansion of border enforcement. In 1993, President Clinton’s first year in office, the agency mobilized huge numbers of agents to guard the country’s urban borders, starting in El Paso, Texas. That pushed more crossings to rural, unpopulated areas—which quickly proved to be far more deadly because of the rugged desert terrain, extreme temperatures, scarcity of water and the long, indirect routes often used by smugglers.

Stopping these remote crossings was a new law enforcement challenge for the agency, especially given the limited technology available at the time. “We had radios,” says Galvan, who would go on to be the head of operations for the El Paso sector—not the cell phones or GPS systems agents have today. They relied on underground sensors left over from the Vietnam War to alert them to movement.

The biggest technology rollout of Galvan’s early career started around 1998. That was when Border Patrol began installing its first video cameras that agents could remotely pan and zoom. They were mounted on metal structures similar to cell-phone towers, 60 to 80 feet tall, or on buildings in urban areas. 

“For the first time, somebody sitting in a room somewhere was able to surveil large swaths of area without being dependent on getting an agent out there with binoculars,” says Matthew Hudak, a former deputy chief of Border Patrol who was working as a frontline agent at the time. “That was a very significant game changer, and for the most part, it was a huge, huge advance.”

This early iteration of the virtual wall was a big deal: In its first eight years, 200 towers were installed along the southern border at a cost of more than $429 million. But the system’s flaws were obvious to anyone who used it.

“There could be … 30 screens on a wall,” says Mark Borkowski, who managed CBP contracting, including the surveillance tower program. Each camera might be capable of seeing three or five miles in any direction. The agent was expected to keep eyes on all that space—upwards of 850 square miles.

“Well, what are the chances after 20 minutes that those Border Patrol agents would see an explosion on one of those screens?” Borkowski says. “About zero.” 

When these cameras were first being put up, the plan was to pair them eventually with a system that would automatically direct them toward places where sensors picked up activity, so agents would know where to look. But by 2005, the DHS inspector general said that still hadn’t materialized, and that illegal activity might “go unnoticed” unless agents were actively watching the cameras. Nonetheless, more towers were installed, forming what is today known as the Remote Video Surveillance System (RVSS). In 2013, the defense giant General Dynamics won a contract worth up to $103 million to overhaul and expand RVSS along the border, becoming the primary contractor behind the system that remains in use today. (That contract eventually grew to a ceiling of $216 million, and the company received another award in 2023 worth up to $135 million.)

General Dynamics would install more cameras, better sensors, and additional towers. But the problem identified nearly 15 years before remained unsolved. It was information overload: One agent told MIT Technology Review about an incident in San Diego in which cameras captured a group of 30 crossing the border and getting picked up in a van, all unnoticed by the agent monitoring those cameras. Agents said the virtual wall was bringing more of the border into view than they could reasonably pay attention or respond to. 

The camera room was also the job nobody wanted. New Border Patrol agents, often in their early 20s, want to be outside in trucks and ATVs, not sitting in a dark room watching screens—even though the expanding virtual wall increasingly required someone to do exactly that. “Someone that’s injured and can’t go out into the field—we’ll send them to the camera room,” says Rafael Reyes, who was in charge of the Border Patrol station in Deming, New Mexico, until 2024. 

Reyes says it was also a job his agents were often too busy to do properly. His station area had about 30 cameras, and the agents responsible for them also had to monitor sensor alerts and radio traffic, run records checks, and coordinate with other agents. “Then if they have some time, they’ll monitor the cameras,” he says. 

Not only that, but many cameras didn’t work. At Reyes’s station in New Mexico, five of the 30 cameras were broken on any given day, he says. Others had limited functionality; they might be capable of looking in just one direction or seeing only in the daytime. Jaime Fierro, a former agent who worked in the Laredo sector in Texas until 2025, said the broken towers were tanking morale among agents and that they brought it up at every staff meeting. “That was like basically having an eye shut for us on the border,” he says. 

They weren’t alone: In 2024, members of the House Committee on Homeland Security wrote that more than 66% of Border Patrol’s first-generation towers were unusable, despite operating budgets of $50 million to $100 million each year. In January 2026, 30% remained broken, one congressional staffer told MIT Technology Review. The RVSS program was estimated in 2020 to have a total cost of $3.7 billion.

Border Patrol has nonetheless publicly credited these cameras with helping to save lives. The agency has published dozens of press releases, dating as far back as 2016, about instances when agents responded to people seen in distress on the remote video cameras. Despite these publicized wins, the bodies began to pile up, undetected, beneath these systems’ watchful gaze.

Near El Cenizo, Texas, there’s a roughly 13-square-mile expanse of mesquite bushes and grasses described by the Webb County Sheriff’s Office as containing sections of Hachar Ranch and Espejo Ranch. It’s surrounded by seven RVSS systems. It was nonetheless a hot spot for fatalities: MIT Technology Review found that 19 people died there, including one man who died less than 500 feet from the road bordering the ranches and half a mile from two schools. In each of these cases, narratives that we obtained via public records requests describe officers’ learning of these people only after their bodies were discovered or through 911 calls reporting them missing, not as a result of the border surveillance technology installed across Hachar Ranch. 

In response to a list of questions about its towers and these specific incidents, General Dynamics referred us to CBP. CBP did not address any of our questions about RVSS towers.  

 

Stories of some of the people that died near RVSS towers

Many deaths near RVSS towers happened in Texas, in the heavily surveilled corridors between the Rio Grande and the nearby highways, where paid smugglers often organize pickups after crossings. Cameras watch many of the spots where people are known to cross, often looking across the river into Mexico to spot them before they reach the water. Drownings are common; we found multiple cases in which local fishermen hooked the bodies of people who had died attempting the crossing.

In one instance, a family of three from Brazil attempted to cross the Rio Grande near Del Rio, Texas, in 2022. The father, Daniel Lenda, was carrying his two-year-old daughter, Eloah, on his shoulder when he fell in the water. The mother, 23-year-old Thais Natali Montenegro Lenda, lifted Eloah out of the water as she watched her husband disappear. 

She placed the girl, now unresponsive, on a rock and went for help. She was just 100 feet from one camera system and a third of a mile from two others, set up precisely to enable faster apprehensions at a popular crossing point. Our topographical analysis shows that all three cameras had a clear view of the place where the family crossed and exited the water. Nonetheless, Thais Natali was not spotted until she waved down an agent on the Del Rio-Acuña bridge. The agent provided CPR to the toddler, who did not survive. Daniel’s body was found days later. 

Val Verde County sheriff Joe Frank Martínez responded to that case (though Border Patrol might discover bodies, local law enforcement is responsible for investigating the deaths). He grew up in the area with his nine siblings and has been sheriff since 2009. Riding with MIT Technology Review through Del Rio in his patrol vehicle, he rattled off the deaths that his office has responded to since he became sheriff. 

He keeps a three-inch-thick binder of these cases in his office. But this case with the toddler still sticks out in his memory. He told MIT Technology Review that when he arrived on scene, he heard from Border Patrol that the tower operator monitoring the camera had seen the father and daughter fall in the river. He never received an answer as to why agents weren’t sent sooner. Martínez asked for the footage from the surveillance cameras to aid in his investigation of what happened. Border Patrol never provided it. (General Dynamics did not respond to our questions about this incident, instead directing us to CBP. The agency did not address questions about this incident either.)

Establishing how many people have died in range of the virtual wall is challenging, but it’s especially so in the case of the RVSS towers. There are different models, some with a shorter range of sight—one to three miles—and others with a longer range of up to 7.5 miles. Former agents and officials told us most towers should see at least five miles, but since no public records exist to confirm the model type of any given tower, we used several possible viewing distances to calculate the number of deaths considered in range.

Another complication is that many remains are never found at all. For those that are, MIT Technology Review could analyze only deaths that had been logged with GPS coordinates or sufficiently precise location descriptions. Many records lack this level of detail, including nearly a fifth of the more than 1,500 cases we obtained from 14 Texas counties. We also needed to determine when someone died, rather than simply when their remains were found, to confirm that a tower was present at the time of death. When official sources did not provide an estimate, we consulted multiple medical examiners on how to make that determination from the available details. And satellite imagery provides only intermittent evidence to suggest when towers arose, not exact proof of when they were functioning. For these reasons, as detailed in our methodology, we tried to be conservative in our analysis.

Even so, the number of deaths estimated to have occurred near at least one of these towers is staggering: somewhere between 500 and 700 since 2015. Even if we look only at towers in rural areas, where their views are far less likely to be blocked by buildings, we still estimate more than 300 people have died in places where the cameras had a line of sight, most less than three miles from a tower. Of the nearly 300 RVSS towers we included in our analysis, more than 250 were within range of a location where someone died. More than 150 people died within a half-mile of one, often across open, empty plains, close enough to see the cameras themselves perched atop towers up to 200 feet tall.

The next watchers

When Borkowski was overseeing the rollout of those first-generation towers at CBP, he became increasingly certain that simply having more remote video cameras was not going to lead to faster responses. The information overload meant the government could watch a large area of the border but not meaningfully see, much less act on, everything within it. The agency’s focus then became solving that overload problem. 

What if, instead of relying on someone to scan 30 screens, the towers could use radar to detect motion, and then point the camera to it automatically? 

In 2011, Border Patrol asked the industry for a new generation of technology that could do just that—initially focusing on Arizona, where the border was busiest. The contract ultimately went to the Israeli defense contractor Elbit, whose tower systems watch Israel’s borders. In 2015, the company started installing versions of these—called integrated fixed towers, or IFTs—that could each see more than five miles.

The IFTs were still being installed when President Trump first took office, in 2017. They didn’t command the same attention as Trump’s campaign promise of a physical wall, but there was still support. Trump’s deputy CBP commissioner, Ronald Vitiello, told Congress in 2018 that the IFT towers automatically detect people with radar and then track them, adding that such surveillance technology is “critical in protecting border areas with short vanishing times, where illicit crossers can quickly evade law enforcement by ‘vanishing’ into border communities.”

There were 55 IFTs built in Arizona under an initial $145 million contract with Elbit and other contracts that followed. The towers collectively watch more than 7,000 square miles of terrain, from the mountainous areas outside Nogales to the plains of the Tohono O’odham Nation Reservation. Arizona is in many ways well suited for these tall towers because much of its landscape is flatter than, say, Southern California. 

But even here, being within a tower’s advertised range does not necessarily mean being within view. There could be terrain, brush, or buildings that block the line of sight. 

MIT Technology Review conducted a topographical analysis to estimate whether each tower had a clear or obstructed view of the place where someone’s remains were found. We estimated the height of each tower and used US elevation maps to model the terrain between it and the remains. The analysis does not account for buildings or vegetation, though all the towers we analyzed use some form of thermal imaging that is designed to see through light brush.

Most of the instances in which someone died in range of a tower happened where we estimate that the camera’s views were not blocked by terrain—a finding that held in both rural and urban areas. And even when someone died in a blind spot, it does not mean they did not pass through a tower’s line of sight. Additionally, these numerous blind spots indicate issues with the tower’s supposedly comprehensive coverage.

We found that nearly 350 people have died in range of Elbit’s towers since 2015. There were 40 people who died within range of four or more IFTs, with some near as many as six different towers. In May 2024, a 19-year-old man died where three IFTs had a clear view, and a year later, in May 2025, a 38-year-old woman died where two IFTs had a similarly clear view. Elbit did not respond to a detailed list of questions about its towers or deaths we found to have occurred near them. 

There’s been a death within range of 52 out of the 53 IFTs we analyzed, but some were particularly notable. There’s a tower located a couple of hours from Tucson on the Tohono O’odham reservation that we estimate has upwards of 80% visibility of the surrounding area. Still, 27 people, ages 19 to 64, have died in its surveillance area just since 2021, along with several others whose remains have not been identified. In 25 of these cases, we estimated that the tower had a clear line of sight to the location where the person died.

Stories of some of the people who died near IFTs

---

The age of autonomy

Around 2018, officials within Border Patrol were again discussing the limits of the towers. The new radar systems were better, but they could only detect movement, not what was moving. This generated false alerts from cattle, tumbleweeds, and other harmless activity. The now decades-old dream of the virtual wall remained unfulfilled. Under pressure from Congress, which was demanding to know what $33 billion in proposed funding for border security over the next decade was going to achieve, Border Patrol once again was hoping a new generation of technology could solve its problems.

That year the head of CBP, Kevin McAleenan, created a unit called the Innovation Team. “He gave them some seed money to go after kind of Silicon Valley innovative technologies and run pilots,” Borkowski says. 

One company they landed on was Anduril, the defense tech startup founded by Oculus VR founder Palmer Luckey. The company was without a product but had funding from Peter Thiel, recruits from the security tech giant Palantir, and a goal of bringing Silicon Valley experimentation to defense and national security technology. 

With grants from the Small Business Innovation Research program, Anduril’s product started taking shape: a 33-foot-tall solar-powered structure with cameras to see and radar to detect movement, all sitting atop a modular metal pole, with solar panels arrayed beneath. 

More important was the component you could not see: computer vision that promised to automatically classify, track, and generate alerts for what came within the system’s view. Anduril’s standard tower couldn’t see as far as the Elbit and General Dynamics models—about 1.75 miles instead of more than five—but, the company argued, it didn’t need to: By automatically identifying and tracking people, the towers could give agents enough warning to respond. (Its newer, extended-range towers, on the other hand, stand 80 feet tall and are advertised as able to detect and track “objects of interest” up to 7.5 miles away.) 

When President Trump left office in 2021, Anduril continued to find support from his successor. Joe Biden had campaigned on the promise of shutting down the project to expand the physical wall, but he saw political reasons to keep investment in tower technology on the table. “It’s less visible—it’s unobtrusive,” a former immigration advisor to Biden recalls, adding that despite some privacy concerns, the towers were mostly uncontroversial. 

When Border Patrol deployed Anduril towers to the Santa Teresa station in New Mexico in 2021, a press release stated, “Because of its accuracy with detection, in many cases this type of technology can and will save migrant lives.” In a permitting request filed in 2024 for towers in California, the government was even more explicit: “If anyone within the viewshed of the towers appears to be in distress, EMTs or first responders will be sent to help.” 

Anduril echoed those claims. The software built into the surveillance towers, the company wrote in a 2021 press release, gets agents “out of communications operations centers and into the field where they can effectuate security and humanitarian responses.”

It wasn’t the first time CBP had discussed a humanitarian side to its duties; in 1998 the agency created BORSTAR, a search and rescue unit, and by August 2023, it had installed some 170 rescue beacons across the border that people attempting a crossing could theoretically use to call for help. But they weren’t effective, says Mario Agundez, a retired Border Patrol supervisor from Arizona who had worked on various rescue efforts; people either avoid them or “die next to them, because there was no way [for Border Patrol] to get to them in time.” (They don’t always work, either; one that MIT Technology Review passed, just next to a dirt road in the Otay Mountain wilderness area in Southern California, had its prominent solar-powered blue light, meant to be visible for miles to people in distress, turned off on a recent night in August. A CBP spokesperson did not comment on why it was off but said, “The solar-powered beacons are maintained and monitored by the responsible Border Patrol station that patrols that area.”)  

Anduril soared through a pilot phase in just two years before its Sentry towers—then the only autonomous surveillance towers CBP was using—moved into an official program of record with the government in 2020. That’s light speed in the world of government procurement, especially since it was Anduril’s first hardware product. By June 2021, CBP had already spent over $98 million (Anduril’s contract for these towers is now worth up to $1.1 billion). 

---

Gerardo Galvan was in charge of the Santa Teresa Border Patrol station—which patrols the area in New Mexico where Morales’s body was later found—when it received its first batch of Anduril surveillance towers in 2021. His agents were worn out; the El Paso sector, home to his station, had seen encounters tick up each month since April 2020. Galvan says the agents were surprised but eager to get the AI towers, and he began planning where to place them. The sites he was considering sat where one kind of landscape abruptly gives way to another. 

To the east of the station is Cristo Rey, a mountain that obscures the view of the city of El Paso. Running north from Cristo Rey is the state’s border with Texas, where the banks of the Rio Grande are flanked on either side by pecan orchards. And west of those orchards are the towns of Sunland Park and Santa Teresa. The last houses in these neighborhoods butt right up against the desert, which sprawls westward without a single gas station for nearly 100 miles.

It is in this stretch—between the US-Mexico border to the south and Highway 9 to the north—where agents from his station aimed to stop migrants. And when Galvan was planning the tower locations, there were lots of crossings; Santa Teresa was becoming the busiest station in the 125,500-square-mile El Paso sector, as heightened border enforcement in Texas drove people west. Along this new route, more people were dying than in previous years. (The traces of that period are still visible years later; MIT Technology Review visited the area with the humanitarian group Battalion Search and Rescue and saw jawbones bleached white by the sun, alongside frayed backpacks and clothes.)

Galvan’s aim was to use the new towers as a stopgap where there was little other border infrastructure. That included a spot near the foot of Mount Cristo Rey, where there was a break in the border fence and people often attempted to cross. Galvan looked at data on previous apprehensions with an eye toward helping agents spot people before they arrived in more residential neighborhoods. Border Patrol struck agreements with private landowners, who generally didn’t mind having Anduril’s low-profile towers on their land (and were compensated via lease agreements). 

When the towers were first set up near the Santa Teresa station in 2021, engineers from Anduril came to fine-tune the algorithms meant to autonomously classify whether what it had detected was a vehicle or a possible border crosser, among other things. A former engineer for Anduril, who spoke on the condition of anonymity to discuss his previous employer, says these algorithms were the main focus of the tower program; Anduril didn’t manufacture the cameras or radar systems itself, so the algorithms were its main value proposition to Border Patrol. 

Galvan says the towers would flag people crossing before agents saw them. But he also saw problems from the start. Agents would apprehend a large group near one of the towers, for example, and check to see if the footage revealed anyone who got away, only to find that the incident hadn’t been captured at all. Anduril’s towers constantly pan around, lingering only on an object of interest. But in these cases, Galvan says, they panned elsewhere, failing to capture the border crossers they were supposed to automatically track.  

Reyes says algorithm mistakes were infrequent but problematic at his station. A group of people was sometimes labeled as cattle, for example, and agents wouldn’t receive an alert. The problem got worse the farther groups were from the camera.

Anduril has touted its technology’s ability to reduce false positives, like cases in which wildlife is mistaken for people. But the company has said little about these false negatives: people the systems fail to detect.

Jaime Fierro, the former agent in the Laredo sector in Texas, describes one incident that shocked him: Agents pursued a car believed to be carrying migrants who had just crossed the border. The car turned around, drove toward the banks of the Rio Grande less than 100 feet from an Anduril tower, and crashed right into the river. Several occupants got out and swam to Mexico while the car floated in the water. Back at the station, Fierro hurried to see what the camera had picked up. Only it never detected the incident at all. There was no footage. “That was a huge, huge issue,” Fierro says. He remembers Anduril coming out to investigate and the problem being sent all the way up to Washington leadership. But Fierro never got an explanation of why the crash was missed. (An Anduril spokesperson did not respond to a question about this incident but said the range of the towers we asked about was limited by physical obstructions and by boundaries established by CBP. )

Several agents estimate that the towers initially missed 10% to 15% of what they should in theory have caught. “To us, 10-15% on a system we spent millions on that was supposed to work—that was alarming,” Fierro says.

Philip Sullivan worked as an agent in the Laredo sector too. When the Anduril program started, he says, he put his hand up to be involved and even went to an Anduril test site for training. When new towers went up, he worked directly with the company on improving its algorithms. He enjoyed the work. But he describes one recurring issue they couldn’t resolve.

Smugglers would put up to a dozen people on a rubber raft, cover them with a tarp, and cross the Rio Grande while using submersible motors to propel it forward. The tarp fooled the algorithm, which detected the motion but attributed it only to an “unidentified object.” That meant no audible alert—the very feature agents relied on, because the Anduril system was supposed to do the watching for them. 

The smugglers repeated the tactic dozens of times. Other smugglers would cover groups with netting or blankets so they would remain undetected while walking through the brush. Only afterwards, having seen the group farther inland and noticed clues about where the people had crossed, would Sullivan review the footage and see the crossings that were missed. He worked with Anduril, sending engineers videos of the crossings to retrain the firm’s algorithms. But the situation had not improved by the time he got promoted to a sector-level job sometime in 2022, he says.

“The whole purpose of their system,” Sullivan says, was the idea that “the cameras are moving around, identifying automatically and tracking and recording and flagging.” Having cameras that could pan on their own worked wonders, he says. But the algorithm had its limits. “You look at it on the screen yourself, and you see the blob and the raft moving across—well, you know that’s people. But to get an AI to identify that is a challenge.”

When the towers were first being rolled out at the Santa Teresa station in 2021 and 2022, Galvan says, similar issues led to a negative feedback loop: An agent would see a tower miss something significant and come to distrust the algorithm. When that agent’s turn came to operate the towers and manage the alerts, they might override the AI altogether and use the camera manually. That would lead to more missed alerts, more distrust.

Despite these shortcomings, the El Paso sector’s relatively flat landscape meant that the system’s cameras had about 80% to 90% visibility, according to our topographical analysis. That was better than in hillier or more mountainous areas, like the Otay Mountain Wilderness to the east of San Diego, where we estimate Anduril’s towers could see just 15% to 25% of their promised coverage area. 

It was in this wilderness, in fact, that some of the first bodies started appearing near Anduril’s towers. During one week in August 2021, two women in their 20s, from the same city in Mexico, died miles apart. Their bodies were found near three different towers that were first observed on satellite imagery between March and July that year. 

One of them, 26-year-old Sarahi Hernández Alfonso, began her journey to cross into the US on August 5, and medical investigators note that she reportedly fainted the next day near the Otay Mountain Wilderness and was left behind. It would be another four days before the Mexican government reported her missing to Border Patrol on August 10, supplying a set of coordinates. Agents went to those coordinates and found her decomposing body. She’d died 1.3 miles from two different Anduril towers, one of which our analysis found had a clear line of sight.

Just the day before, 25-year-old Karina López Antonio was found dead near a third Anduril tower. She had crossed the previous night with her cousin and nephew. That morning, she felt sick, and her nephew looked for a Border Patrol agent to help. When agents returned, she was dead. 

The following October in New Mexico, where the new Anduril towers had been rolled out, agents were patrolling on ATVs when they came across footprints. They followed them until they found the body of 31-year-old César Perea Itzincab. The presence of maggots and the level of decomposition indicated that he had died at least a week earlier, and medical examiners on the scene believe he had dragged himself to the spot where his remains were found. He was 1.5 miles away from an autonomous tower—within Anduril’s advertised range—and should have been visible to the camera, according to our topographical analysis.

We can’t know what those cameras saw as these people died. Footage and data from Anduril’s towers are overwritten every 30 days, and a former official with internal affairs at CBP told us the information wouldn’t be saved for longer unless it was part of an active investigation—which starts only if someone dies in custody. CBP did not respond to questions about specific incidents like this one, and the Sunland Park Police Department closes cases like this when it determines that no crime has been committed. 

The lack of documentation is a missed opportunity, says Amerika Garcia Grewal, co-director of the Frontera Federation, which aims to help rescue migrants in distress—or find and identify their remains. “The surveillance towers along the border could be incredibly useful,” she told us. For example, they could in principle be used to hold Border Patrol agents accountable for whatever actions they did or did not take to locate someone in distress. “What the tool turns out [to do] depends on the person who is holding it,” she says, adding, “I don’t trust the folks that are using them.”

“If I could get that footage,” she says, “then we would go through that, and hopefully bring some answers” to family members and “documented proof of what happened” in their loved ones’ final moments. 

When asked about deaths near Anduril towers, agents often point to faults in the technology. Galvan, for example, says the towers don’t capture how groups move: A large group of 20 people might splinter into smaller groups, especially if agents are pursuing them, and the tower doesn’t keep track of everyone. If someone’s missed, he says, “that person stays behind in the brush, out of sight.” Then they might pass out—which could be a death sentence in the desert. 

And though Border Patrol says the towers can “hand off” surveillance of people from one tower to another, agents on the ground say that people are often missed during these handoffs. On top of that, the algorithms remain imperfect, according to Jason Owens, the chief of Border Patrol from June 2023 to March 2025. “We never really got to the point of ‘Set it and forget it,’” he says.

Agents also say they were just stretched too thin to respond effectively. Around 2022 and 2023, the border saw historically high levels of migration. Title 42, a policy that began under Trump and continued under Biden until May 2023, cited a public health emergency to expel migrants before they could apply for asylum. But the rapid expulsions carried fewer of the repercussions that could normally follow an apprehension, and many people simply tried to cross again.

Meanwhile, asylum claims had been rising for years, while shifts in US immigration policy and in the places migrants where originating from meant more people required lengthy processing rather than being quickly returned across the border or to their home countries. That increasingly tied up agents with people who had already been apprehended, they said, leaving fewer available to respond to surveillance alerts. 

In other words, from the agents’ perspective, the effectiveness of Anduril’s towers was limited by the same issue that had always caused trouble: There just weren’t enough agents.

But the overwhelming demands on agents cannot explain all the deaths our investigation found. In 2024, the border started to get quiet again. By July of that year, the number of Border Patrol encounters nationwide had plummeted from their highs in December 2022. Encounters in the El Paso sector, where the Anduril towers in New Mexico were located, had fallen to nearly one-tenth of their peak. Agents were less tied up than they’d been in years. The technology was supposedly improving, too. 

Galvan had left for a promotion by this time, but he says agents had come to trust the towers more, and they helped train the algorithm by giving a thumbs up or thumbs down if the system identified something correctly or incorrectly. There was also a new smartphone program: Every agent at the station was given an Android device with a map of where other agents were, and it sent alerts from the station about what an Anduril tower detected. 

Despite all this, the deaths near towers continued. There was Morales in April, whose body Border Patrol learned about only from the landfill workers, though it had lain for hours within sight of an Anduril tower. 

Others were discovered by luck. 

On June 20, Border Patrol agents were mistakenly tracking a group of people they thought might be migrants, though they were in fact employees of the same landfill. Those employees told the agents they had come across a body, which turned out to be partially mummified remains of a man in his 40s who died within range of two Anduril towers. 

Three days later, on the 23rd, another: the body of a 21-year-old Guatemalan man discovered by agents while on patrol, 1.3 miles from a tower across clear and open desert. Two days after that, another: An agent was looking for a lost person when he came across the remains of a 56-year-old man—just a football field’s distance from the previous one, with a similarly clear and unobstructed view to a tower.

So if Santa Teresa had gotten quieter—and become something of a showcase for Border Patrol’s newest technology efforts—what was going on?

“Agents review the information and determine the appropriate response—the technology does not make law enforcement decisions,” Beckham, CBP’s assistant commissioner, said. Meanwhile, individual agents described scenarios where they might not prioritize responding to alerts.

Sometimes “hanging back” was a strategy: Agents might observe a group to see what route they would take, or agents might not respond to a small group, in case those people had been sent by smugglers to distract from a larger group coming behind them. Other times, agents might deem the location where a group of migrants had first been spotted to be impractical for an apprehension, and instead wait to intercept them at another location.

Knowing someone’s location didn’t always lead to immediate action. Volunteers describe providing Border Patrol with the coordinates of someone who was lost but alive and then waiting, “sometimes [for] a week,” for agents to reach the location, says Garcia Grewal, from Eagle Pass, Texas. They might be told “Oh, so-and-so went out there and they found remains,” she recalls. “Well, they weren’t remains when we called you. They were alive.”

Mireya Morales, the younger sister of José Morales Bernal, who died by the landfill on the day before his birthday in April 2024, wasn’t aware her brother died near several surveillance towers, or that the government said these towers could save lives. “Well that’s good,” she said of that promise, “but in this case, it didn’t help my brother. And there’s no way to know exactly how things played out.”

“Border Patrol should have found him quickly,” she says, “and they could have done something.” If they did, this journey into the United States—his fourth trip—likely would have ended with his apprehension, detention, and deportation to Mexico. But he would have seen the birthday texts that his sister sent the next day. At least he would have made it home to celebrate his eldest daughter’s quinceañera, which took place earlier this year.

Stories like this are why Iván Chaar López, an assistant professor of American studies at the University of Texas at Austin who leads its Border Tech Lab, says he’d “rather talk about harms” than about “the failure of an algorithm.” He adds, “A system may fail, but humans suffer.” 

Stories of some of the people that died near Anduril’s autonomous surveillance towers

What happens next

Ultimately, any attempt to understand where things are going wrong is hampered by an institutional reluctance to measure the problem.

A complete audit of this virtual wall can only come from Border Patrol. The agency is, in some ways, increasingly equipped to take on that task: Agents told us that each time they apprehend someone, the location is logged with GPS coordinates, as are instances when Border Patrol sees evidence that someone crossed but cannot find them. It also has precise data on when each surveillance tower went up, which alerts came in, and how they were resolved. Having that information is the only way to measure whether the newest towers are truly missing fewer people than the previous generations. 

But researchers and outside oversight agencies say this ocean of data hasn’t translated into a reliable system for measuring the effectiveness of the virtual wall. For example, CBP records the people it misses by logging “gotaways,” a tally of how many times agents see signs of someone who evaded apprehension—footprints, appearances on camera, reports from other people who were apprehended. It’s imperfect. And that creates room for interpretation.

“The goals with border enforcement have always been a moving target,” says Jeremy Slack, a researcher of migration and border issues at the University of Texas at El Paso. “They’re always set up so that it’s a win-win.” If apprehensions go up, for example, CBP says it means agents have gotten more effective at catching people. If apprehensions go down, it means the border is quiet because people are too scared to cross. 

That’s not to mention the conflicting ideas within the agency about what the virtual wall was supposed to accomplish. Borkowski says higher-ups would ask how many fewer agents they could get by with if they built more towers. But supervisors receiving new towers told us they’d often ask for more agents, not fewer, because they now had more activity to respond to that had previously gone unseen. 

And if the goal was for the sight of the towers to deter people from crossing to begin with, an external study from RAND in 2020 was ambiguous: It found that deploying IFTs resulted in lower apprehension levels nearby but said this didn’t mean the towers were actually deterring crossings. (Research by Boyce and colleagues found that earlier surveillance towers in southern Arizona pushed people to seek more difficult terrain out of view.) It leads to a question: Do the deaths near the virtual wall constitute unacceptable surveillance failures, or are they within the range of effectiveness the government deems acceptable? (CBP’s response did not address our questions on how it explains these deaths.)

Against this backdrop, the Government Accountability Office and DHS’s Office of the Inspector General have tried to focus on a narrower question: Is the virtual wall increasing the likelihood of apprehensions? 

The answer has been incomplete since 2014. That’s when the GAO first suggested that whenever agents log their activity in Border Patrol’s database, they should include information about whether or not a piece of technology helped in their apprehension. This could at least do something to show Congress whether the technology is working. Border Patrol began collecting that information, but years of incremental improvements have been followed by repeated findings that the resulting data is unreliable or insufficient to determine how well the technology works. The reality, several people from Border Patrol told MIT Technology Review, is that agents see it as a chore demanded only by bureaucrats who don’t understand the realities of the border. 

It’s “spitting in the wind, to be honest,” a former high-ranking official at DHS during the Biden administration told MIT Technology Review on background. “If we have 10,000 people a day crossing in between the ports of entry, do you think an agent is worried about telling them how they freaking apprehended those people?” Several other officials said that the logging of technology assists had improved, but not consistently enough to help much with analysis. 

And there is no procedure—at the local or national level—to examine individual deaths near the towers or analyze them collectively for surveillance failures. “We never thought about plotting [migrant deaths] to see if the towers were missing things or agents were missing people,” said a former CBP official who evaluated how personnel handled deaths in custody.

“Nobody ever raised the issue,” the official said, adding that doing so could reveal important gaps in the agency’s approach: “If I still worked for CBP, and you brought it up, I would probably send some people to look.”

In response to questions from MIT Technology Review, CBP said that autonomous surveillance towers complement physical barriers by improving detection and situational awareness. Its statement did not address questions about its other technologies, or broader criticisms about how it evaluates the virtual wall.

This lack of measurement has not slowed the enthusiasm for more technology funding. The government spending bill passed in July 2025 awarded $2.7 billion for border security technology alone, and CBP was quick to signal to industry that the money would soon start flowing. “We’ve got an historic investment in infrastructure and technology coming up,” a CBP director told tech companies in an industry webinar that month. In a December 2025 interview, when he was director of homeland security and immigration at the America First Policy Institute, Cooper Smith described the technology investments as “fortifying” the border against a future president who might be, in his estimation, weaker on the border than Trump. (According to LegiStorm, a research organization that focuses on political staffers, Smith now serves in a policy-focused role with CBP.) The number of Border Patrol agents has climbed too, to nearly 21,500 agents as of June 2026—the highest in the agency’s history. 

In the Big Bend region of Texas, surveillance towers are becoming a focus of local politics: Officials and residents across party lines, including the sheriff of Terrell County, are asking for Anduril towers instead of a controversial new border barrier project that’s recently been halted.

CBP has announced a desire to spend $1 billion on nearly 1,500 more towers by 2034. The law now requires all these towers to be equipped with AI. In December 2025, Anduril’s Palmer Luckey said the towers remained the company’s best-selling product, one that gives “basically perfect situational awareness of what’s going on in the area.” Anduril’s towers have been sold abroad—to enforce the UK’s borders, defend US Marine Corps bases in Japan and elsewhere, and serve as drone defenses for an Australian Air Force base. 

But as for CBP, Anduril will no longer be the only game in town; General Dynamics has now released AI towers of its own and received an order for them in June from CBP worth up to $115 million. The Electronic Frontier Foundation—which has tracked surveillance towers since 2022—says some have already been spotted at the border.

This time around, the agency will be spending this money with even less oversight than it had just a couple of years ago. In October 2025, the Department of Homeland Security, which oversees CBP, dissolved its department-level office that oversaw major spending programs.

---

Most of the records on human remains that MIT Technology Review collected were current only through the fall of 2025 or, in a handful of cases, early 2026. Even those dates come with a caveat: Many remains are never found, and others are found months after someone died. The reports can then take months to be released via public record requests. That makes it nearly impossible to keep track of how many people have died near surveillance towers in anything close to real time.

Still, the border today is significantly quieter. The latest figures from DHS show just 128,009 “enforcement encounters” from January to August 2026, compared with more than 2.4 million during the same period in 2024 (“encounters” count each of Border Patrol’s interactions, not individual people).

Even as border crossing rates decrease and DHS speeds ahead with its tower acquisition plans, people are still dying within range of the surveillance towers.   

On September 14 2025, a 30-year-old Mexican woman named Graciela Gómez Hernández died just over 350 yards away from the physical border wall that separated Southern California from Tijuana—and within one mile of an RVSS tower. Earlier that afternoon, she had told her family that she could not walk any further, and her voice notes abruptly stopped. 

Her skeletonized remains were recovered nearly three weeks later, on October 4. Some of her bones were missing, and there was “apparent animal activity … to the ribs,” as the medical examiner’s report read. 

Sara Stroud, an organizer with the Borderlands Relief Collective, a volunteer humanitarian group that leaves water and supplies in a wilderness area frequented by people crossing the border, went to the site of her death a few days later to put up a memorial cross. A Border Patrol helicopter showed up almost immediately, flying low circles above the volunteers. 

Between the helicopter and the RVSS tower that was visible in the background, “I was astounded about how we were being surveilled the whole time,” Stroud says, adding that this was especially stark in contrast to the agency’s often slow—or absent—responses to people who’d died on camera. 

“Are you telling me you didn’t see it or you did see it?” she asks. “Because if you did see it, that’s horrible. And if they can’t, then how do you explain that?”

Additional reporting by Lillian Perlmutter.

This work was supported by a grant from the Tarbell Center for AI Journalism.

She had only walked for a couple of hours, and already she was lost. 

It was early afternoon on Sept. 14, 2025, when 30-year-old Graciela Gómez Hernández crossed the border from the eastern edge of Tijuana into Southern California, sending voice messages to her mother and sister as she walked. 

Gómez Hernández did not tell her family about her plan until the day before. Once she arrived in California, she hoped to work and send money back to her three children and her parents in southern Mexico. Her oldest son, age 11, had asked for a bicycle.  

She either climbed a ladder over the border wall or slipped through a section in steep terrain where fencing has yet to be built. Dozens of game trails cut through the dust on hillsides of sagebrush and manzanita. It was the first segment of a hike through the Otay Mountain Wilderness. 

The eastern edge of San Diego is just a mile away, but countless migrants hike 10 or 20 miles north through the mountains to evade the Border Patrol, emerging in smaller towns and on distant highways with fewer checkpoints.

Between 2 and 3 p.m., her sister said, Gómez Hernández sent a final message to her mother. “I can’t do it anymore,” she whispered into her phone. By the time a forensic crew came to collect her, two weeks later, most of her body had disappeared into the earth. 

She died on her own, but she was not, strictly speaking, alone. She was in one of the most heavily surveilled strips of land in the world.

Before she died, Gómez Hernández told her family in a WhatsApp message that she could see helicopters overhead and white Border Patrol trucks making periodic laps on the road below. And even if none of those agents saw her or knew she was in distress, something else had a clear view. Supervising the whole territory, from a nearby hilltop, was a slender gray surveillance tower. 

That tower, made by General Dynamics and in place since 2019, boasts two sets of electro-optical and infrared cameras with long-range lenses that can track targets 5 to 7 miles away. Gómez Hernández was just one mile away. 

Surveillance towers have proliferated in recent years, as contractors added newer camera systems that use artificial intelligence to distinguish border-crossers from other people and animals, and track their movements, at an overall cost of more than a billion dollars. 

Yet as the cameras scan the terrain, migrants continue to die in plain sight. 

Thousands of migrants have died while crossing the border in the past two decades, many of them near to Border Patrol surveillance installations. But the phenomenon is especially pronounced in the Otay Mountain Wilderness, just outside San Diego. 

A Times of San Diego analysis counted hundreds of cases along the California border—including at least 138 in the past four years—where migrants’ remains were found within the range of the very cameras designed to help track and intercept them. 

A topographical analysis conducted as part of an MIT Technology Review investigation estimated that more than half of these remains were found in a spot where a camera would have a clear view of a person, without terrain blocking the line of sight. 

Some of these migrants died swiftly, falling from the border wall or drowning in a canal or river. But dozens suffered slow deaths of exposure in the wilderness, deaths that might have been averted with emergency aid. 

People who died in Southern California came from just across the border, or from as far away as Africa. They were as old as 70 and as young as 5. They died in the heat or the cold. But they all died within range of a surveillance camera.

---

Found in July 2023: Marcia Dutra, a 36-year-old woman from Brazil who likely died of heat exposure next to a border fence. According to a medical examiner’s report, a group of migrants told the Border Patrol about her body when they were apprehended. An AI-equipped surveillance tower sits 2,000 feet northeast above the edge of a plateau, out of sight of the slope where she was found. 

Found in February 2024: Elvi Vasquez Bortolon, a 28-year-old Mexican man who died of suspected hypothermia on a ridge near the border wall. He had been dead for more than five days when a Border Patrol agent in a helicopter spotted his body from above. A remote video tower stands about 1,400 feet away, hidden from view by the ridge above the canyon. 

Found in August 2024: Ramón Montenegro, a 51-year-old man from Mexico who died of possible heatstroke in the mountains near Dulzura. An AI surveillance tower, invisible on the other side of a hill, sits 1,300 feet away — roughly the length of the parking lot at the San Diego Zoo. 

More than 120 surveillance towers stand along the California border, according to documents and live observations collected by the Electronic Frontier Foundation. 

The towers are not the only form of surveillance in the region. The border is blanketed by an interconnected network of trail cameras, underground movement detectors, radar, infrared sensors, drones, and even blimps. In March 2025, the Trump administration directed the Department of Defense to use satellites to watch the border as well. 

Even when mountains or vegetation block a tower’s view, other devices may pick up the trail. Local advocates say surveillance technology in the Otay Mountain Wilderness is almost unavoidable. 

Officials say the towers particularly excel at tracking passing migrants. 

Anduril, the company that makes about half the towers now installed at the U.S.-Mexico border, said in a news release in 2024 that their towers “directly contributed to saving lives and stopping illicit drugs from entering the U.S.” In June 2026, the company said its installations had “autonomously identified hundreds of thousands of border crossings.”

John Morris, a Border Patrol sector chief in Arizona, told media at a convention in May 2026 that “We’re dang close to pretty much knowing everything that comes across.” 

The proximity of so many deaths to high-resolution cameras leaves two distinct possibilities: Either the system didn’t detect these people in distress, or it did detect them—and officials offered no response. 

Anduril, the defense contracting company that manufactures the AI-enabled towers, did not respond to questions from Times of San Diego about the towers’ performance in mountainous terrain.

A representative of General Dynamics, which made the tower closest to Gómez Hernández’s body, referred all questions to the Border Patrol. 

The Border Patrol did not respond to questions from Times of San Diego regarding whether surveillance towers spotted Gómez Hernández prior to her death, or about how surveillance cameras in general are used to detect people in emergency situations. 

Rafael Hernández Larraenza, the leader of a search and rescue group, Desert Angels, says he believes some of the towers are “just adornment,” and are not actually effective at tracking migrants. He also says that Border Patrol agents, in occasional meetings with his search group, “say they don’t have the opportunity to do searches, that they don’t have enough time.”

People in need of rescue “are not a priority,” said Bryce Peterson, who compiles research on migrant remains for the advocacy group No More Deaths. 

A wilderness, heavily watched  

From the spot where Graciela Gómez Hernández’s body was found, a surveillance tower stands in clear view. 

That tower is one mile away. A second, newer tower, on the other side of a hill, sits 1.4 miles away. Both are equipped with slow-spinning, long-range cameras. 

The closer of these two towers is situated between two layers of border fencing, a long pole, over 100 feet tall, reaching into the sky. It peers not only onto hillsides where migrants walk, but into the brightly-colored cinderblock homes of a Tijuana neighborhood. 

The second tower is much shorter, just 33 feet tall, surveying the canyon where Gómez Hernández would most likely have arrived if she had kept walking. 

This new tower, built by a company called Anduril, decides for itself whether a moving object in the field is relevant. The two dark, round cameras on the pole give the impression of eyes, rotating slowly, and then locking in on a target, before swiveling elsewhere.

Osvaldo Ruiz, the volunteer coordinator for the Border Angels, a group that leaves water and supplies along migrant trails near San Diego, says the Otay Mountain wilderness has “beacons throughout the mountains” that “can see over a lot of the mountains facing the border.”

“At entry sites that migrants or travelers may come through, there are also hidden cameras inside of rocks or inside of bushes that are camouflaged for you not to see. And then there’s also the cameras that you see on the road. They have license plate readers that are stuck on the back of a safety cone, so (it detects) as you’re driving away from it.”

Surveillance towers have been in the area since 2011, with frequent new installations, including the introduction of AI-directed towers in 2018. Many of the AI-equipped towers currently overseeing the California borderlands appeared in 2021, and were updated with longer ranges in 2026.  

Officials from the Government Accountability Office wrote in a report in June 2026 that the Border Patrol currently has 803 fielded towers “and plans to deliver another 175,” across the southwestern border, an initiative that would cost $1.4 billion. The report also notes that the agency expects a billion dollars in funding from the One Big Beautiful Bill Act in 2025. By 2034, the number of towers is expected to rise to 2,300.

“What we are building along the southwest border is a smart system. There’s technology embedded in it, plus the physical infrastructure,” border czar Tom Homan said at a press conference in San Diego in December 2025. “It makes every single agent and officer more effective.”

A secretive hike north 

Like so many migrants, Graciela’s hike into the U.S. began in desperation. Her husband, who had already crossed the border to find work, had stopped sending money for the children. 

“There were moments when they didn’t have enough money even for a bag of detergent, or to eat,” her sister, Irma, told Times of San Diego. “She would call me crying.” 

Gómez Hernández decided she needed to come to the United States, if only temporarily, to make money. Her family was uniformly against the decision, so she kept her plan a secret until she was already at the border, preparing to cross. 

“She was a very humble person,” Irma said. “She was kind to people, and loving with her children.”

Gómez Hernández paid 20,000 pesos, about $1,150, for smugglers to help her cross the border, much lower than the current average, which sits at $12,000. But she would need to travel on her own without a guide, paying only for the knowledge of where and when to start, and which direction to walk. As part of their deal, her phone would send location pings back to the smugglers’ hiding spot in Tijuana. 

Cesar Ortigoza, a leader of a search group known as the Armadillos, has been locating migrant remains in the region since 2010. 

“Most people don’t understand the magnitude of this territory, and all the work it’s going to cost them to arrive at their destination. They get to the top of this mountain and realize there’s miles and miles and miles left to walk,” Ortigoza said. 

In addition, they need to avoid capture. “They go really fast,” he said, “and there’s a lot of pressure.”

On the afternoon of September 14, 2025, Gómez Hernández told her mother and sister through WhatsApp voice messages that she had successfully crossed the border into the United States.

These short snippets of audio would be their only clues to what happened next. 

---

The hillside where Gómez Hernández walked that afternoon is exposed on all sides, with a panoramic view of the border wall just a couple hundred yards away. But on the other side of the hill is a daunting, mountainous wilderness, covered in low trees and rocky ravines, with hidden, winding trails, stretching into the distance. It is possible she saw this and turned around, convinced she was lost. Wandering the hillsides in the afternoon heat without sufficient water, she could succumb to dehydration within hours. 

“Because of the heat, there are places where you can’t move forward even half a mile in an hour,” said Larraenza, the search and rescue leader. 

By 2 p.m., “She told me she felt very tired and couldn’t go any further,” Irma said. “I told her to give it her all, to think of her children, keep them in her mind so she could get out of there.” Irma says 4 p.m. was the last time her sister’s location was transmitted to the smugglers. 

---

After hearing about Gómez Hernández’s case, Ortigoza said it “seemed strange,” because she was “so close to the border” at her last known location. 

But Bryce Peterson, the No More Deaths volunteer who collects autopsy reports on migrant remains border-wide, said “There’s a million ways that somebody could be already on the verge of death that close to the wall. Like if you’re going north and then you decide to go south again because you get lost or something.”

“The only thing we’re certain of is that she suffered in her last moments,” Ortigoza said. “Just knowing that you’re alone.”

Walking on the hillside, Gómez Hernández encountered a man also traversing the trails, lined with dry shrubs. “In one of the messages sent, where she told me to pray for her, the man spoke and said ‘there isn’t time,’” Irma said. 

After that, Graciela’s phone still received messages, but she did not respond. 

Cameras rise, questions remain 

Even though the federal surveillance tower was just a mile away, it is possible the cameras did not see Gómez Hernández. It is possible they did not see much of anything. 

An internal memo obtained by NBC News in October 2024 revealed that 30% of the older-style towers, the kind that require agents in control rooms to review footage, were broken. Most were missing parts, or their equipment was outdated and unusable. These 150 broken towers still stood at the border, acting as million-dollar scarecrows. 

Officials from the Government Accountability Office wrote in a report in September 2025 that the Customs and Border Protection had begun to fix the towers with spare parts and revamp them with new technology, but did not list how many remained out of commission. 

The Border Patrol did not respond to questions from Times of San Diego about whether the tower in view of where Gómez Hernández died was operational at the time of her death. 

Dave Maass, the director of investigations for the Electronic Frontier Foundation, the organization that maps surveillance towers at the border, questions the functionality of all types of towers. “None of these systems have been effective. There’s no independent evaluation of how many things they’re capturing,” he said. 

In a report published in March 2026, a representative of the US Military notes that a mission to place 7 G-BOSS(E) towers in the Otay Mountain Wilderness, an installation that took place in the months prior to Gómez Hernández’s death, faced operational challenges. 

The combat-grade military surveillance towers were vulnerable to tipping over, damage from rodents, exhaust blowouts, and running out of battery. On a dirt road in the mountains, a truck carrying a surveillance tower bounced around, which led to the tower’s cameras “becoming stuck in a zoomed-in position.” 

Despite descriptions from manufacturers claiming dominance in “rugged terrain,” Osvaldo Ruíz, the Border Angels coordinator, said it would still be difficult for a surveillance tower to track migrants moving through the Otay Mountain Wilderness. “With all that surveillance that’s out in the mountains, there’s still a lot of brush and high covered areas that would make it easy [to stay undercover].”

Analysts from the RAND Corporation, a foreign policy think tank, wrote in a report in 2020 that “outcomes and outputs” that would indicate the success of surveillance towers “are not directly observable.” 

The analysts conceded that the towers’ presence would urge migrants to stay out of the area and funnel into undetected routes. “The likely deterrent effect on migrant crossings through surveilled areas overwhelms any boost to situational awareness,” that the towers provide, the analysts wrote. In this way, the surveillance towers’ most useful quality was the threat conveyed through their technology, rather than the merit of the technology itself. 

Due to increased surveillance, “People are crossing through areas that are further out and more extreme, where we have very little hope of finding them in a search,” Larraenza said. 

And though the manufacturer says the AI-enabled towers, which now make up over half of the total, can identify migrants on their own, apprehensions still require human labor. A CBP official told the RAND analysts that smugglers began sending migrants across the border alone, rather than in groups, to overwhelm the surveillance towers with many individual targets, leading to more alerts than agents capable of following them. 

So it is possible the towers in the vicinity did capture footage of Gómez Hernández, but agents decided not to apprehend her. Just hours before her death, she sent a message saying that there were “police” nearby, possibly referring to Border Patrol agents in trucks. 

“Even one person who’s not sick, one person who is just hiking on their own, often Border Patrol will just ignore them because it’s not worth their time to just only catch one person,” said Peterson of No More Deaths. “They want to be able to catch a whole group of people.”

Larraenza said he’s found dehydrated people on migrant trails who say they called out to Border Patrol, and the agents failed to stop. “But there are also cases where agents do the impossible to find someone,” he said. 

On one occasion in September 2024, while rescuing a family of five in the Otay Mountain Wilderness, Larraenza met a person wearing a vest labeled “SHERIFF” on the trail. In a video that Larraenza posted to his Facebook account, when asked if the Border Patrol would arrive to take the family into custody, the deputy says, “Oh, they don’t come out here.”

---

On a recent visit to the spot where Gómez Hernández died, multiple Border Patrol trucks were visible just hundreds of yards away, rolling slowly along the wall, or parked on the hill next to the surveillance tower. 

“If we hadn’t sent the Border Patrol the coordinates, they never would have bothered to go [find Gómez Hernández’s body],” Ortigoza said.

Finding Graciela 

Irma sent her sister a barrage of messages that afternoon and evening, pleading with her to send her location, but there was no response.

Irma believes it’s possible her sister was killed or abandoned by the other migrant she was traveling with during her final hours, because she was not walking fast enough. 

It took Irma two weeks to make contact with the coyotes. When she told the smugglers that her sister was missing, “they didn’t move a single finger,” she said. She threatened to send one coyote’s phone number to the police, until he sent her a screenshot of Graciela’s last known location. 

Almost three weeks after Graciela Gómez Hernández’s final message to her mother, on October 3, 2025, the Armadillo search group sent the last known GPS coordinates of her phone to the Border Patrol. 

The medical investigator’s report states that Gómez Hernández sent her location to her family, and her relatives sent the coordinates to a search group while she remained alive, but both Irma and Ortigoza said they did not know her location for weeks, and neither was interviewed for the examiner’s report. 

“On 10/04/2025 at 11:16 hours, I arrived on a dirt path approximately 100 yards north of the United States/Mexico International border. The area consisted of shrubs and dried brush with steep hills and declines,” Jennifer Wright, the medical examiner, wrote in an autopsy report. “Upon further review, I viewed the skeletonized remains of an apparent female lying supine in the dirt.”

The body had decomposed quickly with the help of the sun and animals, so it was impossible to tell whether Gómez Hernández was the victim of murder or dehydration. But the medical examiner noted that there was no water or food in the vicinity.

“In a pile next to the decedent was a purple bra, black pants, a black floral shirt, and a dark colored long sleeve shirt. The bra was hooked closed and intact,” Wright, the medical examiner, wrote in her report. “In the rear pants pockets were Mexican paper currency, Mexican coins, bobby pins, and tweezers.”

The body was naked, but it was unclear whether someone else removed Gómez Hernández’s clothes, or whether she took them off herself in the final throes of dehydration delirium, a common symptom immediately preceding organ failure. 

Several yards from the place where she died, Ortigoza found a backpack with a bottle of women’s spray deodorant, and a small glass container of perfume. 

“When you get to one of these sites, and you see the situation, the loneliness, the environment, you start wondering, what was this person thinking in their last moments?” Larraenza said. 

Once a DNA test confirmed the identity of the remains, Irma paid over $4,000 to send her sister’s bones home to Mexico, where the family held a funeral nine months after her death.

“Sometimes her oldest son, he just sits there thinking and when we ask what’s wrong, he says he misses his mom,” Irma said. Though the family is heartbroken, everyone has pitched in to care for Graciela’s children. “I don’t let my parents see me cry,” Irma said. 

The Border Patrol gave permission to Ortigoza and a group of searchers to place a cross next to where Gómez Hernández was discovered. On a recent trip to the memorial to leave flowers, Ortigoza played a voice message that Irma sent to her sister, the sound vibrating into the dirt of the hillside, where the imprint of Gómez Hernández’s body is still visible nearly a year after its removal.

“I know that in the place you’re in, there isn’t suffering, like how you suffered here on this earth,” Irma said. “I know you left thinking about your children. I know you wanted to give them a better life. I feel you here, and I feel you alive.”

Cesar Ortigoza says every day he gets as many as five calls from people looking for their missing family members at the border. Though numbers of border crossings, and numbers of reported deaths, have decreased since Donald Trump took office in January 2025, “There are still a lot of people on the other side trying to get here. Because the problems in their countries persist,” Ortigoza said. 

In the first half of 2026, fewer people have consumed the water bottles, Gatorade and cans of food left by Border Angels volunteers in the Otay Mountain Wilderness. But supplies left on other routes by different groups have seen increased consumption, evidence that migration patterns are constantly evolving. 

“It’s hard for me to really understand how people are crossing, but I know they are,” Ruíz said.

Peterson, the No More Deaths map-maker, said more migrants could die in the coming months and years, as a result of increased surveillance. In the past, after surveillance towers were installed, migrants would try to go around them, leading to longer and more dangerous routes through stretches of remote wilderness. With the installation of hundreds of new towers, the spaces outside the range of detection have thinned, leading some migrants to attempt to go through surveilled territory, rather than around it. 

“We’ve seen people do things that to me seem terrifying or crazy, like going past an area where they know there are sensors and then trying to really quickly get to the next safe-ish place before Border Patrol shows up to get them, because it takes 15 minutes for them to show up,” Peterson said. “It seems like people aren’t stopping to drink water. People are going really fast, often running even, moving much more quickly, not being able to take the time to take care of themselves.”

Larraenza says for this same reason, migrants often choose to cross the border in the hottest weeks of the year, when they believe fewer agents will be on patrol, but there is a greater risk of death. In an interview in mid-September, Larraenza said at least two people had gone missing in the Otay Mountain wilderness in the past few weeks, and he had not yet been able to find their remains. 

Ortigoza says Gómez Hernández’s story is an example of a decades-long trend of people putting their lives at risk to cross the border out of necessity. The surveillance towers, whether or not they were functional, did not stop Gómez Hernández from crossing the border in that area, and did not stop her from dying. “If the riches of the world were distributed more equitably, we wouldn’t have to do things like this to reach a better life,” he said. 

While placing flowers in July at the site of Gómez Hernández’s death, Ortigoza noticed a Border Patrol truck parked under the surveillance tower on the hilltop a mile to the east. An agent stood on a power lift, working on the tower’s cameras. 

This year, agents replaced dozens of these cameras with new, AI models. 

Lillian Perlmutter is an investigative reporter covering immigration across California for Times of San Diego and NEWSWELL. Previously based in Mexico City, she has written for more than 25 outlets including the L.A. Times, Rolling Stone, The Guardian, and The New Republic.

On Wednesday, MIT Technology Review hosted a live Roundtables event for subscribers that asked the question everyone’s asking right now: Could AI really kill us all?

But attendees had so many more questions than we had time to answer in the 30 minute session. So we asked our senior AI editor Will Douglas Heaven and AI reporter Grace Huckins to round up some of the best questions attendees submitted and try their best to answer them.

Thanks to all who submitted questions!

Am I gonna die?

Yes, eventually. Unfortunately, my journalistic powers of prognostication aren’t powerful enough for me to tell you how. But it certainly could be because of AI. AI-powered drones have already killed people in Ukraine, and AI-driven cyberattacks on hospitals will surely claim victims before long. 

Could AI go even further, and kill all of us? Less likely. But some people—quirky people, but undeniably knowledgeable about AI—have been warning for years that this could happen. And while I’m not yet stockpiling canned food or trying to get in good with a bunker-owning megabillionaire, I have noticed that the doomers’ predictions about AI capabilities and alignment have, over the past couple of years, proved disconcertingly accurate. That certainly doesn’t mean that their more dire forecasts will come true, but it’s enough for me to sit up and take notice.

— Grace Huckins

Are you going to die because of AI? I’d say there’s a non-zero chance. Let’s say you’re unlucky enough to be the victim of a freakish near-future event or accident. Maybe it’s a cyberattack carried out by a swarm of AI agents on critical infrastructure. Sadly, a scenario like that now no longer feels as far-fetched as it once did. Or maybe a novel AI-designed pathogen cuts through the population. Or the world economy crashes, causing conflicts and famine. Both plausible, but I think less likely. 

Are we all going to die because of AI? Nope. There are no circumstances outside of apocalyptic science fiction in which AI could kill us all. You can spin up any number of scare stories, but they’re not grounded in present-day realities about what the tech can do or where it’s headed. 

Some people argue that there’s no harm in preparing for the worst, however wacky it might seem. Maybe. But I think such catastrophizing can make people excuse or overlook many of the more immediate problems with the existing technology and the companies building it. 

— Will Douglas Heaven

Why would AI kill us?

Someone might tell it to, and it might listen. That’s part of the reason researchers are so concerned about AI’s biological capabilities—imagine what Aum Shinrikyo, the doomsday cult behind the Tokyo subway sarin attack of 1995, would have done with a tool that could design a pathogen deadlier than Ebola and more transmissible than measles. Those of us who don’t want to die have to figure out how to defend against all plausible biological weapons, but our would-be attackers only have to manufacture one effective pathogen.

Then there’s the more exotic-sounding possibility that an AI could decide to kill us itself. There are various stories about how this might happen out there, but the most widespread involve AI systems that don’t hate people, necessarily—we are just an obstacle between them and the goals that we gave them.

Much as the OpenAI agents behind the Hugging Face hack compromised another site’s infrastructure to get a good score on a test, the idea is that some future, more powerful AI might get rid of us to prevent us from shutting it down—all in pursuit of some goal that we instructed it to go after. 

— Grace Huckins

How can we best ensure alignment so the worst doesn’t happen, and who is doing the best work to achieve it? 

Alignment is a huge area of research. In simple terms, it involves building models that behave in ways we want them to and not in ways we don’t. We need to trust agents better before handing over more autonomy. Alignment is supposed to establish that trust. But it’s hard. 

LLMs aren’t designed in the way other software is, where dos and don’ts can be hard-coded in. Instead, aligned behavior needs to be instilled when models are trained. One approach is to reward them for doing things you want them to (a little like raising a toddler, perhaps). Another approach involves giving an LLM a written list of rules it is supposed to follow (kind of like a constitution). 

Anthropic and OpenAI are both leaders in this field—and yet neither has been able to develop models that are fully aligned. A big problem is that LLMs are far more inconsistent and far less predictable than people. They can behave in one way in one situation and another way in a situation that to us seems very similar. They can also be swayed by unexpected constraints. For example, faced with an impossible task (as many of the agents involved in the Hugging Face hack were), models may try to do whatever it takes to achieve their goal. As Grace mentions above, that could be an issue.

The main reason top AI firms now say they want a slowdown is that they want to focus on cracking alignment. Alignment isn’t necessarily a pipe dream. But the jury’s out on whether full alignment will ever be feasible. 

— Will Douglas Heaven

Is AI really dangerous, or is this the tech companies drumming up PR?

This is always a reasonable thought when it comes to tech companies heading for an IPO—CEOs have an obvious incentive to make their products seem radical and transformative. But I’m not so sure it makes sense here. Telling the public that an already unpopular product could kill them and everyone they love is horrible corporate image management.

There are other stories you can tell about the CEOs’ motivations—maybe they want to cool down the public furor over data centers by portraying themselves as responsible stewards of a world-changing technology, or maybe they want to buy time to get their ducks in a row and prevent the next PR catastrophe. 

But there’s also a simpler explanation. Thinking that AI could bring about human extinction has been pretty common in San Francisco for a while, and these men are steeped in that milieu—as are their employees, many of whom signed an open letter in July urging their companies to work to make an AI slowdown possible.

— Grace Huckins

Part of the concern occurs when AI agents are allowed to act autonomously and with no supervision. What’s the issue preventing more control over these agents? 

This question goes to the heart of what we want this technology to be able to do. The trade-off between autonomy and control is tricky to get right because, on the one hand, a lot of the power of AI agents is that they can carry out tasks and solve problems without a human having to micromanage them. On the other hand, that requires you to trust that the unsupervised agents won’t run amok. 

What we’re seeing is that AI labs haven’t yet got this trade-off quite right. Their models are not trustworthy, they are not properly monitored, and they are not always under control. Figuring out how to fix that while still allowing for useful autonomous activity is one of the big research challenges of the moment.  

— Will Douglas Heaven

What steps can be taken now and in the near future to ensure that AI is controlled, monitored, and regulated effectively? 

That’s the million-dollar question. Whether or not you think AI could kill us, you can’t deny that it could do some real damage, because it already has—by driving people toward psychosis and by hacking websites, for example. Preventing that damage, or at least mitigating it, is hard for two reasons. 

The first is that we barely understand how AI works, and it’s quickly growing more powerful. There is lots of ongoing research about how to monitor and control misbehaving agents, but the current approaches are fragile. You can see if an agent discusses misbehaving in its “chain of thought,” the workspace where it plans its actions—but OpenAI’s newest agents don’t show their work in the same way as previous ones. And you can try to monitor agents with other agents, but that requires you to trust the monitor.

The other obstacle is more familiar. There’s a huge conflict of interest when AI companies regulate themselves, but the US government has thus far failed to step in, despite some bipartisan support in Congress for efforts to do so. The executive branch, for its part, seems stringently opposed for the time being. But if the winds do shift, I for one would appreciate some strong transparency regulations, so that we can get a fuller story the next time an unreleased frontier model mounts a cyberattack.

— Grace Huckins

If this dialogue makes it into web discourse, will it become a self-fulfilling prediction?  

That’s a real concern. LLMs are influenced by what they read. One theory for why chatbots so often talk about (and role-play) apocalyptic scenarios is that they have been trained on millions of pages of science fiction stories and doomer internet forums. All the text being produced right now, including this article, could in turn influence the behavior of future models. Extremely meta.

In fact, the team at METR, a third-party organization that OpenAI called in to help understand what happened in the lead-up to the Hugging Face hack, raised a related possibility in its report on the incident. METR used OpenAI’s new model Astra to help analyze the vast numbers of agent transcripts and behavior logs.

But feeding all that material to the model could have unintended consequences. There’s a good chance that the agents doing the analyzing were biased by the text produced by the agents they were analyzing. There’s no such thing as a clean slate anymore. 

— Will Douglas Heaven

With thanks to Eric, Pranab, Rafael, Kenneth, George, Chris, Yoon Jae, James, Carl, Nicole (and more!) for the fantastic questions.

The AI boom is becoming a materials challenge. As AI pushes computing into new territory, the materials behind that infrastructure are becoming just as crucial as the algorithms running on it. Semiconductors and data centers are approaching physical limits around performance, thermal management, electrical efficiency, and reliability, creating new demands for materials that can do more at once. At the same time, AI is giving materials scientists new ways to search the enormous universe of possible molecules and accelerate the development of solutions.

For Mike Finelli, chief technology and innovation officer and chief North America officer at Syensqo, that convergence is transforming what advanced materials can enable. “AI is now, from a material standpoint, really pushing semiconductors and the data centers to their physical limits,” he says.

As requirements accumulate, including high temperature, purity, electrical performance, chemical resistance, plasma resistance, and long-term stability, materials move toward what Finelli calls the “top of the pyramid.” Beyond supporting AI innovation, he contends that advanced materials are “actually increasingly defining what’s going to be possible.”

That challenge is playing out across the infrastructure powering the AI surge. Syensqo is developing materials for high-voltage data center architectures, advanced sealing materials for semiconductor manufacturing, and thermal-management solutions including fluids for direct immersion cooling. Some of those innovations can also cross industry boundaries. Materials developed for electric vehicles, for example, can help address the higher voltage and energy-density demands that are emerging in data centers.

The definition of performance is also changing. More customers are expecting materials to meet technical requirements while reducing environmental impact. “Our goal is to remove the trade-off between performance and sustainability,” Finelli says. That means considering sustainability at the beginning of the research process instead of treating it as an additional requirement once a material has been developed.

AI is changing how those materials are discovered, too. Syensqo is using AI agents to digitally synthesize millions of potential molecular combinations, predict their performance and sustainability characteristics, and narrow them to a much smaller group for laboratory testing. The result, Finelli says, is the ability to go “broader, deeper, and faster” while giving scientists more time to solve complex engineering problems.

Looking to the future, Finelli sees the possibility of a reinforcing cycle: AI helps develop materials that improve AI infrastructure, which in turn enables better AI to accelerate materials discovery. That feedback loop could create a cycle of innovation and expand what future technologies can achieve.

“You end up in this accelerated materials, innovative cycle of materials innovation,” says Finelli. “That really excites me, and it gives us the opportunity to continue enabling technologies that will shape the future.”

This episode of Business Lab is produced in partnership with Syensqo.

Full Transcript:

Megan Tatum: From MIT Technology Review, I’m Megan Tatum, and this is Business Lab, the show that helps business leaders make sense of new technologies coming out of the lab and into the marketplace.

This episode is produced in partnership with Syensqo.

Now asked to name the key enablers to AI advancement, many of us might list algorithms, data centers, or even computing power, but just as critical to the performance are the advanced materials that underpin each layer of that innovation. As AI continues to evolve, it’s pushing the likes of semiconductors and data centers to new physical limits, putting new pressure on the advanced material sector to keep pace. But the relationship goes both ways. As the sector rises to this challenge, AI is also emerging as a powerful tool for accelerating materials discovery and development, significantly shortening development timelines for new solutions.

Two words for you: materials innovation.

My guest today is Mike Finelli, chief technology and innovation officer and chief North America officer at Syensqo.

Welcome, Mike.

Mike Finelli: Thank you, Megan. Nice to be here.

Megan: Thank you so much for joining us. Mike, can I start by asking you to tell us a little bit more about Syensqo and the role it plays in developing advanced materials?

Mike: Yeah, absolutely. Syensqo is a global leader in specialty materials. Our job is to help customers solve their toughest technology challenges. We serve a lot of different markets, but the way I like to say it simply is if it flies, we’re on it. If it drives, we’re in it. In healthcare, our products literally are saving lives every day. And if you like your mobile devices, if you like AI, it’s our products that are actually enabling the advanced semiconductor chips that are required to produce all of this. Our role is to enable innovation through advanced chemistry. We develop materials that deliver higher performances, greater reliability, and increasingly more sustainable solutions. The way I would say this, it’s at the heart of our business. Actually, it’s in our name, Syensqo. And to put some numbers around it, 20% of our annual revenues come from new products and applications that we’ve launched in the last five years, which is really evidence of a really strong innovation engine.

Megan: Yeah, absolutely. And as you sort of described there, you’re in all sorts of different industries with an emphasis perhaps on electronics and semiconductors. Can you talk a bit more about that work and where those industries are headed perhaps?

Mike: Sure. So look, electronics and semiconductors have been strategic markets for Syensqo for literally decades. I don’t want to date myself, but 33 years ago when I started in the company, semiconductors were one of the first industries that I worked in. And we’ve supported successive waves of innovation from enabling smaller, more powerful mobile devices, helping the industry get to the smaller and smaller profiles and the chips. We’ve helped to advance hyperconnectivity, supporting increasingly sophisticated semiconductor manufacturing. And today we’re helping to advance the AI era.

We have one of the industry’s broadest portfolios of high performance polymers and advanced materials. We support applications across the entire electronics value chain from semiconductor fabrication, electronic components, to smart devices and telecommunications, even hyperconnectivity. And our materials are helping customers solve increasingly demanding challenges around miniaturization, thermal management, electrical performance, chemical resistance, higher and higher purities, and long-term reliability and sustainability. And today we work with leading semiconductor manufacturers and electronics companies all around the world.

Megan: Fantastic. And as you alluded to there in the last 30 years, we’ve seen huge evolutions in those sectors.

Mike: Oh my God, yes.

Megan: And now AI is putting these new demands on semiconductors and data centers. What does that mean for the materials they’re built from and to what extent will AI innovation be constrained or enabled by materials science finding a solution?

Mike: Yeah, so I mean, you’re absolutely right. But AI is now, from a material standpoint, really pushing semiconductors and the data centers to their physical limits, and materials are becoming a key enabler of that continued progress.

The way I try to describe it, think of a pyramid, I call it the performance pyramid. You have commodity materials at the bottom of the pyramid and you have high performing specialty materials at the top of the pyramid. At Syensqo, all we do is we operate at the top of the pyramid and we’re continually trying to raise the top of that pyramid by bringing newer and newer and more higher performing materials out.

Now you might say, okay, but why doesn’t a data center or a semiconductor manufacturing fab need a specialty versus something in the commodity space? Well, I call it the and, and, and principle. If you just need a polymer or a material that can sit at the table at room temperature and stay there for 10 years and not change, well, there’s a lot of commodity materials that will do that and you don’t have a problem. The minute you start adding requirements, and I call it the and, and, and so if you need a polymer that can handle high temperature and have to have high purity and electrical performance and chemical resistance and plasma resistance and it’s got to have long-term stability, all of these ands, you start moving to the top of the pyramid.

Now what AI is doing with semiconductors, because of the speed at which it’s advancing, it’s requiring semiconductor chips and data centers, the number of requirements are increasing the number of ands which is pushing the limits of the materials. That’s where we come in. And I really believe that advanced materials, they’re no longer just supporting AI innovation, we’re actually increasingly defining what’s going to be possible.

Megan: Right. That’s fascinating. And in terms of rising to that challenge of focusing on that top of the pyramid and that and, and, and principle you’re talking about, could you talk us through perhaps an example or two of those top of the pyramid solutions you’ve created or that you’re working on at the moment?

Mike: Like I said, our focus is enabling higher performance, but it’s also without compromising on reliability or safety. We develop advanced polymers, elastomers, specialty fluids, fluids meaning lubricants and heat transfer fluids, and they’re used throughout the semiconductor manufacturing process and also increasingly in AI data center infrastructure. One example of our work on specialty materials for next generation AI data centers is the work we’re doing around high voltage architectures. Data centers are moving towards high voltage architectures because they can enable greater computing power while also improving energy efficiency. We know that’s a big issue for that segment of the industry, and these high voltage architectures will help them reduce and improve energy efficiency because it reduces energy losses and they can ultimately help lower the environmental footprint of the data centers. And we’re developing new materials that can help them get there.

Another example is our high performing sealing materials found inside semiconductor fabs and wafer tools. If you can picture, many people have seen what a semiconductor looks like during processing. It’s a big, big silicon disc that’s then later diced into the tiny little chips that go into the computer. But that wafer is put inside a giant chamber where it has a very extreme environment, aggressive plasmas, reactive chemicals, and they need higher and higher performing materials. And all of the seals that are around that chamber to keep those gases in the environment inside have to be able to withstand that environment. And that’s what we’re developing and we’re pushing the limits. They’re asking for higher temperatures, more aggressive environment with lower out gassing and purity. And that’s what we’re developing for this industry to allow that next chip to be developed and produced industrial.

Megan: It’s so fascinating that people wouldn’t give much though necessarily to the seal in something like that. As you’re outlining, it’s just absolutely critical in terms of performance. And in developing those solutions, I understand you also looked across different markets to see what may be applicable perhaps in more than one space, and that includes an overlap between the automotive sector and data centers, I understand. Can you tell us a little bit more about that?

Mike: As I mentioned just previously, the data centers are shifting to higher voltage architectures. This is the next generation data center, which can be more energy efficient, but it’s got a higher energy density. The power density increases, which increases temperatures. And many of the material challenges that we will be facing there, we’ve already developed for the automotive industry in electric vehicles. I’ll give you an example of an application. I mean, think about an electric vehicle. The powerhouse in electric vehicle is no longer the motor, it’s the battery. That’s where all the energy sits. And when you’re putting a hundred kilowatts of energy, driving that to the electric motor through wires and through what they call bus bars, you got to get that car up to 60 miles an hour pretty quick. You’re driving massive amounts of energy that’s increasing temperatures dramatically.

And all the electrical connections are in these bus bars that there’s a polymer that’s an insulating polymer with copper in between for all the connections. That’s got to withstand that temperature increase, which could come pretty rapidly. We’ve developed new materials there and those materials will be translatable over to these data centers where they’re going to have the higher voltages with a higher energy density.

Another thing we’ve been doing in automotive, we have a lot of knowledge in both automotive and semiconductor around fluid circulation and how to use dielectric materials to do direct immersion cooling. That’s something that will be very valuable for data centers and server farms. Using air to cool semiconductors is really inefficient and energy intensive. If you could submerse them in a liquid, you have direct immersion cooling, that’s extremely efficient, so that’s another thing we’re working on.

Another thing we developed in automotive that will be translated over is battery energy storage systems. Inside the battery, we’ve developed a binder. It’s the highest performing binder on the market, which is using the cathode of a lithium ion battery, and it keeps all the ingredients doing its job working together so that battery can actually last for 10 years and perform. Now that’s moving over to the data centers because they’re moving more towards renewables and they need to have these energy storage systems to smooth the peak loads and provide resilient backup power. That’s one of the things that we’re doing. By transferring our knowledge across the markets, we can accelerate new power and new thermal management solutions while supporting reliability required by next generation AI infrastructure.

Megan: Fantastic. So many transferable applications there that necessarily wouldn’t have sprung to mind. And it isn’t only technical advancements that you need to contend with, of course. Companies today are also demanding the materials are developed and manufactured more responsibly too. So how is sustainability shaping your innovation process?

Mike: Yeah, you’re absolutely right. I will say performance is still the entry ticket. Our customers want performance. Now what’s changing is that definition of performance is now broader and it is including sustainability targets and requirements. Our customers expect materials that deliver outstanding technical performance while also being developed and manufactured more responsibly.

At Syensqo, we believe that operating as a responsible company means we’re providing true sustainable business solutions to our customers. And this is why we developed what we call the Sustainable Portfolio Management tool, SPM. It’s a matrix, and it defines what a sustainable solution is. For us, it’s a product that in a given application improves our product’s social and environmental performance while also demonstrating a lower environmental impact in its production, creating values for our customers. In short, we want to develop products, and this is where it starts. Every one of our research projects before we even start them is assessed on whether it’s going to be a sustainable product or not.

And 88% of our portfolio now is a sustainable product. We’re developing materials that are better for the environment, lower environmental footprint when we produce it, but also they contribute to improvements for our customers as well so they could operate with a lower carbon footprint or they can operate in a safer way or less water consumption. There’s a lot of different lists in there.

Another example is our longer-term development of next generation heat transfer fluids. Semiconductor manufacturing and data centers have become more powerful. I mentioned before the heat that they’re generating, especially when they move to the higher voltage architectures. Managing that heat is increasingly important. And again, I talked about direct immersion cooling. We’re developing those solutions because today there are fluids out there that will work, but they got high global warming. That’s not good for the environment. We’re developing the next generation heat transferred fluids that will reduce the potential environmental impact compared to the fluids today. In the end, our goal is to remove the trade-off between performance and sustainability. You notice that’s another and, we can be performing and sustainable.

Megan: That’s so important, isn’t it though, to think about sustainability in terms of performance? As you say, when we’re thinking about commercially scaling up these solutions, it’s such an important part of it. And as I talked about in the introduction, AI isn’t only a challenge, but it’s also an opportunity within the advanced material space. I’d love to explore how you’re using AI tools at Syensqo to inform and accelerate the development of solutions as well.

Mike: Absolutely. We embarked on this journey about two years ago, where we’re using AI in our research and development, and we’ve partnered with Microsoft and their Microsoft discovery tool, and it’s helping us to rapidly identify and evaluate promising molecular candidates.

Now, in the normal research approach, historically, you would design your experiment and you’d look at all the potential combinations of materials and chemicals that you could make all these different molecules. And the combinations of potential and molecules that you could develop to solve a problem could be in the millions, but it’s impossible to develop a million molecules or tens of millions of molecules in your laboratory and actually physically do that. But you have to select a small area based on your expertise and knowledge, based on the literature searches, based on the state of the art that’s out there and looking at patents, et cetera. And you pick a small area and you go through the process, you develop the materials, you test them, you learn something, you go back to the drawing board, you start again. Eventually you find something that works, but it doesn’t mean you found the best possible combination that’s out there.

But what we’re doing with AI is we have developed AI agents with Microsoft that are literally digitally synthesizing the entire millions and millions of combinations of potential molecules. And we have another AI agents that are using physics-based simulation to look at all those molecules and predict the performance of them, and not just performance on physical chemical properties, but also on toxicity, on sustainability, et cetera. Then we have another agent that takes all that information and ranks them all. In the end, we have explored all of the potential molecules out there. We understand roughly what the performance should be, and we end up with a priority list of maybe a hundred, instead of millions and millions, a hundred that we actually synthesize in the lab.

And at the end, you end up getting the solution faster, much, much faster. You’ve explored the entire space. I basically say it allows us to go broader, deeper, and faster. And the important thing is it’s not replacing our scientists, it’s not replacing our scientific expertise. In a way, it’s giving them superpowers. It’s allowing them to spend less time searching and more time solving the industry’s toughest engineering challenges.

Megan: Amazing. It sounds like it’s genuinely a really transformative tool by what you’re explaining.

Mike: Completely, completely.

Megan: I mean, just to finish, Mike, it’d be great to take a look ahead if we could, because there’s so much activity in both AI and the advanced material space. I wonder what is coming down the pipeline that you are most excited about next?

Mike: I’ve talked a lot about AI and how we’re using AI to develop new materials. I think to me, what’s really exciting, and I’m starting to see it actually happen, I’m just curious how fast this is going to go, is that we’re using AI to develop new materials that will enable AI to get better, and then that AI will use the new AI to develop new materials to get AI to go better. I see this loop of developing for AI, for AI to improve, and then we use that AI to improve ourselves. You end up in this accelerated materials, innovative cycle of materials innovation. That really excites me, and it gives us the opportunity to continue enabling technologies that will shape the future. That’s what we do at Syensqo.

Megan: Fantastic. Yeah, real sort of virtuous circle of innovation, it sounds like that. Amazing. Thank you so much, Mike.

Mike: Thank you.

Megan: Thank you so much. That was Mike Finelli, chief technology and innovation officer and chief North America officer at Syensqo, whom I spoke with from Brighton in England.

That’s it for this episode of Business Lab. I’m your host, Megan Tatum. I’m a contributing editor and host for Insights, the custom publishing division of MIT Technology Review. We were founded in 1899 at the Massachusetts Institute of Technology, and you can find us in print on the web and at events each year around the world. For more information about us and the show, please check out our website at technologyreview.com.

This show is available wherever you get your podcasts, and if you enjoyed it, we hope you’ll take a moment to rate and review us. Business Lab is a production of MIT Technology Review, and this episode was produced by Giro Studios. Thanks so much for listening. Goodbye.

This content was produced by Insights, MIT Technology Review’s custom content arm, not its editorial staff. It was researched and written by humans, with any AI tools that may have been used limited to production processes under human oversight.

When Jessica Wachter, a finance professor at the University of Pennsylvania’s Wharton School, wanted to assess AI’s impact on the economy over the next few years, she faced a long list of business and technical uncertainties. So she started with what she calls a “remarkable fact” that is not in question: A handful of so-called hyperscalers are investing huge amounts of money to build AI data centers.

Instead of trying to predict how useful and widely deployed AI models will be, she simply asked how fast the hyperscalers’ earnings will need to grow to justify their spending through 2027, when—she and her collaborator estimate—expenditures will reach nearly $1.1 trillion. It’s a no-nonsense accounting approach to making sense of today’s historical AI buildout.

The results are eye-opening: The AI companies will need to increase their own productivity by a factor of 2.7 to break even by 2030, accounting for the cost of capital, a 15% return, and depreciation of the assets. Not impossible, says Wachter. The result would lead to the kind of economic growth that we saw during the US IT boom over a period of about 10 years starting in the mid-1990s. But, she says, for it to happen by 2030, “that’s a lot of growth compressed into a few years.” And if the hyperscalers cannot meet such profit goals?

“Then they will fall behind on their interest payments, and that risks bankruptcy,” says Wachter, who was previously the SEC’s chief economist and director of its division of economic and risk analysis. If a productivity boom “fails to materialize,” she and her coauthor conclude in their research paper, “the current buildout will be the largest misallocation of capital in history.”  

It doesn’t take superintelligence to realize that today’s large investments in the infrastructure for artificial intelligence come with huge risks. The hyperscalers will spend about $750 billion this year, building massive data centers scattered across the country. And the spending spree shows no signs of slowing. According to some projections, total AI capital investments from the hyperscaler companies—Alphabet, Microsoft, Amazon, Meta, and Oracle (which partners with OpenAI)—could be more than $5 trillion over the next four years.

It’s one of the largest capital investments by any industry in history. But there’s a problem that’s obvious to anyone paying attention.

While the hyperscalers plan to spend trillions, total AI revenues will be around $150 billion to $200 billion this year, says Gary Gensler, who ran the SEC during the Biden administration and is now a professor at MIT’s Sloan School. “The challenge is that the spending does not have commensurate revenues yet. That’s a fact,” he says. “And then the question is, is that an investment that will be paid off in the future?”

At stake in that trillion-dollar question is the financial health of the giant AI companies and the overall US economy—the investments could soon balloon to around 3% of GDP. The answer could also determine the fate of the hugely expensive data centers themselves. 

No one really knows how profitable and useful these multibillion-dollar behemoths will be down the road. Though AI models have made dazzling progress over the last few years, it’s anyone’s guess how much compute capacity we will need. The technology could become more efficient and therefore less dependent on raw computational power. Or demand for AI products could slow, or customers could turn to cheaper models.

The risks, both to investors and to the economy, have become even greater this year, as these AI companies have begun borrowing large amounts of money to build more and more data centers. Free cash flow—operating cash flow minus capital expenditures—is expected to soon dip into negative territory for the group. Even Alphabet, known for generating and hoarding huge amounts of cash, reports in the latest quarter that its impressive revenues of nearly $120 billion were devoured by AI infrastructure spending, leaving it with a free cash deficit of some $5.9 billion—its first shortfall since Google went public in 2004.

In the near term, it’s not a big financial worry for most of the companies. They make a lot of money and have very deep pockets. But debt is expensive, and some investors are losing patience. If future demand for the data centers’ computation power drops, the companies will still be on the hook to pay back the borrowed money. What’s more, the risks are spreading to the rest of the economy as the loans get passed along via various financial mechanisms. 

It won’t be enough to simply cover the enormous price tags of the new data centers. Hyperscalers will also have to pay for the rising costs of capital as they borrow more money. They will need returns that are impressive enough to justify all their spending to investors and creditors. And to add to those concerns, they will have to make up for the depreciation of billions of dollars in chips housed within the facilities—a ticking time bomb buried in the investments.

Performance of the expensive GPU chips at the core of the data centers—such compute electronics represent some 60% of costs—is roughly doubling every two years or so. The pace of progress helps explain the increasing wizardry of the AI models, but it comes with a cost. Owners of AI data centers that come online this year and next will need to spend billions more on the next generation of chips by the end of the decade if they want to stay competitive. Without the investments, says Mihir Kshirsagar at Princeton’s Center for Information Technology Policy, the data centers risk becoming “hulks,” stranded assets “scattered all over the place.”

To put it bluntly: The AI companies need to start making a lot more money. And they need to do it fast. But juicing their earnings alone still won’t be enough to sustain their data-center investments for the long term.

Productivity is everything

At some point, AI is also going to have to create broad economic growth to justify continuing the hyperscalers’ spending spree.

Sloan’s Gensler describes today’s large investments into AI infrastructure as “a parlay bet by the capital markets and the economy.” That means success will require winning three related but independent wagers: Hyperscalers must generate massive revenues, AI must boost widespread economic growth, and both must happen while the powerful but expensive so-called frontier models that rely on the data centers fend off cheaper versions, which many businesses might find good enough.

What makes this so tricky is that each wager depends on the other two but also poses its own challenges.

If the hyperscalers continue to spend huge amounts of money on data centers into the next decade, revenues will need to skyrocket into the trillions. Stijn Van Nieuwerburgh, a finance professor at Columbia Business School, bases his estimates on a scenario in which about 183 gigawatts of planned AI compute capacity is built between 2025 and 2032; he calculates that each gigawatt costs about $41 billion. Assuming a 10% return—the minimum that would be acceptable to most investors—“required” annual revenues will be roughly $3.7 trillion by 2032, he says.

Others get a similar number.

Winning the second part of the bet—productivity growth across the economy—will be crucial to achieving such numbers.

For a few years, AI companies could likely boost their revenues by simply selling subscriptions and tokens to all the businesses clamoring to get into AI. But eventually—and this might be happening already—those paying customers will need to justify their expenses by seeing bottom-line benefits from the technology. AI will need to fulfill its promise of making workers more productive and making businesses more efficient and profitable while expanding their products and services.

In economic jargon, that means customers will need to see productivity growth. Taken together, these results will mean the country is prospering and growing.

“If you don’t get the productivity gains, at some point people are going to sour on AI, and that will bring down investments and it would also limit revenue growth,” says Daron Acemoglu, an MIT economist and 2024 Nobel laureate. For the investments to be sustainable over, say, the next five to 10 years, we definitely “need to see productivity gains,” he says.

Most economists who watch the numbers closely agree that, for now, the economy-wide statistics show little or no productivity growth from AI. There are some hopeful signs it’s on the way, though. In a recent survey of some 6,000 senior business executives in the US, the UK, Germany, and Australia, the vast majority—around 90%—report no increase in productivity over the last three years. But they expect a boost of around 1.45% in total over the next three years; US executives anticipate a 2.25% bump over that time. 

In a follow-up survey, the respondents also reported plans for their businesses to spend more on AI, leading the authors to anticipate some $280 billion in private-sector AI expenditures by the end of 2026.

That’s good news for the hyperscalers. But it comes with a dose of bad news for those worried about AI’s impact on jobs. The executives expect to increase the productivity of their companies by increasing their sales while significantly cutting the number of employees.

If AI improves productivity by destroying jobs, public backlash to the technology—the kind we have seen around data centers, for example—will likely get worse. Perhaps it’s worth adding one more wager to the parlay bet described by Gensler: The public and local communities must feel that they are also benefiting from the massive investments in AI.

And let’s not forget how interdependent these wagers are; if productivity growth comes from companies running models like DeepSeek, then the hyperscalers’ revenues could collapse. If productivity comes from cutting jobs, a public backlash could block many of the planned investments—and stunt anticipated revenues. We will need to win all the wagers for the hyperscalers’ bet to pay off. 

We’re all part of the AI gamble now

It was one thing when the AI companies were spending cash they had accumulated over the years to build their own data centers. Then the risk was largely limited to their own balance sheets and shareholders. But it’s a higher-stakes game when much of the money is borrowed. Morgan Stanley, for one, calculates that more than half of the $2.9 trillion that hyperscalers will spend between 2025 and 2028 to build AI data centers will be financed with “external capital.”

The borrowing is leading some of the companies to engineer complex webs of financing that are becoming intertwined with much of the rest of the economy. “A lot of financial institutions, directly or indirectly, are exposed to these data centers either as lenders, or as guarantors of some of the debt, or as backers of the private credit funds who are funding these data centers,” says Columbia’s Van Nieuwerburgh. “People don’t even know they’re holding this stuff. It’s somewhere deep inside their pension fund. Ultimately, it’s backing their life insurance policies. And that risk is getting distributed everywhere in places that are invisible.”

As the investments in data centers have spiked, the financial engineering has become more byzantine.

Take, for example, Meta’s so-called Hyperion data center under construction in Richland, Louisiana. When the company announced the two gigawatts of compute capacity at a price tag of some $10 billion in late 2024 it was Meta’s largest planned data center. Greeted with much enthusiasm by state and local politicians, the project, located in the rural northeast corner of the state, was seen as a boon to the community. Entergy Louisiana, the state’s largest utility, rushed forward with proposals to build three large natural-gas power plants to service the massive data center.

Then last fall—the projected cost was now $30 billion—the financing got a lot more complex and, to some in the community, a lot more disconcerting. Meta transferred an 80% stake to the large (and troubled) private-credit firm Blue Owl Capital, forming a joint venture called Beignet (like the famed New Orleans pastry) to raise financing for the data center. Meta then signed a series of four-year leases with the joint venture, an arrangement that the company says gives it “long-term strategic flexibility.” To backstop the agreement, Meta provides the venture with what is called a residual value guarantee, in which it will make a cash payment to cover the value of the facility “following any non-renewal or termination of a lease.” Got all that? 

I hope so. The financial wheeling and dealing is actually even more convoluted, with a cast of wholly owned subsidiaries and LLCs. Beignet has set up Laidley LLC, which owns and operates the site as the landlord. In turn, Laidley leases the facilities to Meta’s wholly owned subsidiary Pelican Leap LLC, which is the tenant. And there is a series of four-year leases that cover the different buildings that make up the data center campus. 

It’s not a coincidence, says Van Nieuwerburgh, that the length of the leases matches the expected lifetime of the data center’s GPUs. While Meta has to pay off its loan if it terminates the leases early, that will still leave its investors “with an empty building and no cash flow,” he says. “And then they need to find a new tenant for a huge data center, and good luck with that.”

Meanwhile, Meta is doubling down on its bet. In July, the company announced it was expanding the data center to five gigawatts of compute capacity. The total price tag is now $50 billion (so far, Meta hasn’t said whether Blue Owl will be involved in financing the expansion). Meanwhile, Entergy is now planning to build seven more gas-fired power plants, bringing the total capacity of the facilities to around 7.5  gigawatts—some six times the amount of electricity used by New Orleans.

If the complex financing is a puzzle to many investors and even financial experts, it is even more baffling to those directly affected by the construction of the data center. The main worry concerns how Entergy’s spending on the natural-gas power plants will affect electricity prices, and who will be left paying the bill for the power if Meta walks away.

Entergy says it has a 20-year guarantee from Meta that the company will purchase electricity over that period to cover the costs of the power plants and related infrastructure.  But there are skeptics, especially given how fast the fortunes of the AI industry are changing. “In four years, is Mark Zuckerberg still going to be interested in this? Or is he going to throw in the towel?” asks Paul Arbaje, a senior analyst at the Union of Concerned Scientists, which has been advocating, largely unsuccessfully, for the Louisiana Public Service Commission to provide more transparency around the data center and its financing.

Even if the 20-year deal holds, consumer advocates are worried that Meta or its partners won’t fully cover all the costs, including those associated with operating and maintaining the power plants—and those additional costs that could be passed on to residential ratepayers. What’s more, says Logan Burke, the executive director of the Alliance for Affordable Energy, if Meta doesn’t end up needing as much power as Entergy planned (these projections are not public), consumers could be left paying for the surplus produced by the plants.

And if Meta terminates its leases early? “It gets complicated very quickly,” says Burke, who questions whether the shifting roster of financial entities will honor existing agreements. “That everybody is going to do what they’re saying they’re going to do over the next 20 years is just hard to believe.”

For UCS’s Arbaje the bottom line is this: “They’re making huge bets that these data centers will be worth it. Bet with your own money, not with ratepayer money.”

After the bubble

Predicting when the AI investment bubble will burst is a fool’s errand. But there is little doubt a day of reckoning is coming, given the irrational exuberance that has overtaken the hyperscalers and their investors. Of course, you might argue that this time is different, and that the rules of accounting and lessons of economic history don’t apply—that AI is too transformative. Maybe, but don’t count on it.

“History tells us that at some point you get a retrenchment, and it’s just a question of when and how severe,” says Sloan’s Gensler. It could be that today’s $750 billion spending rate “goes flat” or decreases next year. Or, he suggests, “we’re now in 2028 or 2029, and then all of sudden they’re retrenching because they’ve got enough capacity.” But, he adds, “you can be pretty assured there’ll be a retrenchment.” 

Though a so-called retrenchment might be inevitable, it’s worth keeping in mind that the fates of the financial bubble and the underlying AI technology revolution could be very different. Already, some Silicon Valley insiders are rooting for a crash; in a recent blog post the longtime venture capitalist Vijay Pande wrote that “the coming crash would be the best thing that happens to this technology.” The argument makes some sense. A crash could make AI investments more rational, calm the impulse to build billion-dollar data centers on every vacant field that CEOs fly over, and refocus investors on how to use the technology to create sustainable value.

But we should probably be careful what we wish for. After the bursting of the dot-com bubble at the beginning of the 2000s, hundreds of thousands lost their jobs, large and small companies alike went bankrupt, the economy of Silicon Valley and San Francisco was decimated (at least for a while), and the shocks sent the US into a mild recession in 2001. For the financial community and many tech workers, it was no fun.

Even more devastating for the economy and the average American was the great recession that began in late 2007. Comparing the financial engineering leading up to it and the methods deployed by hyperscalers today is sobering. So-called special purpose vehicles (SPVs) are back! If Columbia’s Van Nieuwerburgh is right about the dangers of letting investments from the hyperscalers get entangled throughout the economy, the fallout could be severe.

But technologies survived and even prospered in the aftermath of both downturns. The early 2000s, even in the face of the dot-com fiasco, were a time of great innovation and tech optimism. The froth came off the spending on silly technologies, helping to focus investments on more promising ones. It’s no coincidence that each of the hyperscalers rose out of the ashes of the crash or started up shortly after. The fiber-optic infrastructure built during the feverish telecom bubble that ran parallel to the dot-com one is still the backbone of much of today’s communication infrastructure; we wouldn’t have Facebook or Amazon or Google without it.

This time, however, we’re facing a unique risk: The huge financial investments by the hyperscalers have ensnared the future of AI itself with the fortunes of the massive data centers spreading around the country. The logic is founded on a deeply held belief about the power of scaling in AI; the bigger you build it, the smarter it gets. That might be true, but it’s unproven and a risky bet.

There are already plenty of red flags, from strong public opposition to the construction of new data centers to the competitive threat from cheaper, good-enough AI models to the rapid improvement of small, local AI models. None of these trends point toward a future dominated by frontier models housed in massive, billion-dollar data centers.

The financial bubble around the colossal spending by the hyperscalers will likely burst eventually—or maybe soon. It might be financially painful, but we’ll survive. Wall Street will survive. AI itself will survive, though it may look different and lose some of today’s hubris. The financial fate and future utility of the massive data centers fueled by trillions of dollars of spending, on the other hand, are far less certain.

This story appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

This weekend, Dario Amodei, CEO of Anthropic, posted an essay calling for a brake on the pace of development of LLMs. Amodei cites the looming dangers he sees from the technology, from its use in cyberattacks and bioterrorism to its potential to wreck the economy. The heads of the other three top US AI labs—OpenAI CEO Sam Altman, Google DeepMind chairman Demis Hassabis, and SpaceXAI CEO Elon Musk—voiced their support. “Dario is right,” Musk wrote on X.

Think about how surreal that agreement is for a moment. Just a few months ago, Musk and Altman sat in court attacking each other’s reputations in a (failed) lawsuit that Musk brought against his former OpenAI colleague that was—on paper at least—about whether or not Altman was a trustworthy steward of such dangerous technology.

Amodei’s rift with OpenAI is even deeper. Anthropic was founded in 2021 because Amodei didn’t think Altman took the risks of the technology they were building seriously enough. Anthropic and OpenAI have been competing in a winner-takes-all race ever since. (Hassabis has stayed out of the drama, but his company remains a rival.)

Now, it seems, they’re all in agreement: The latest generation of LLMs aren’t safe and everyone needs to figure out what to do about it. The public messaging from the top AI labs has taken a doomer turn.

It’s easy to be cynical. It’s not at all clear what any of them mean by a slowdown or how it would work. These companies also care a lot about how they come across. With trillion-dollar IPOs in their sights, OpenAI and Anthropic need to reassure investors that they’re the grown-ups in the room while at the same time hinting at the power of the monsters they have created—and intend to tame. Calling for a slowdown does both.

And yet the vibe at the top of these firms really does appear to have shifted. Amodei’s latest post landed six days after OpenAI published an essay by Jakub Pachocki, the firm’s chief scientist, in which he also laid out why he’s concerned about what will happen if the pace of development of LLMs continues unchecked. In short, Pachocki is worried that OpenAI’s ability to build powerful models now far outstrips its ability to monitor and control them.

Amodei and Pachocki each cite the cyberattack against AI firm Hugging Face by a swarm of OpenAI’s agents in July—a hack that OpenAI did not even realize had taken place until days after it was all over—as a wake-up call.

But their exact position is hard to pin down. Pachocki both calls for a slowdown and highlights an urgent need to stay ahead: “The strongest argument I see for continuing to train much smarter models quickly is the need to build defensive systems against the dangers posed by other AI,” he writes. As Pachocki frames it, AI firms are locked in a literal arms race. Slowing down is good, winning is better.

(Don’t forget: OpenAI just spent millions of dollars and a staggering amount of computer power to rush out a controversial math result a few days ahead of Anthropic.)

But let’s assume a slowdown happens. Top labs agree to spend more time and resources on finding ways to monitor and control existing models instead of making more capable ones. They invite outside auditors in to help evaluate those models.

What might this coordinated effort actually achieve? Consider the Hugging Face attack again. OpenAI has said that the model that drove most of the rogue agents was a “highly persistent” next-generation model that it was testing in-house. The implication is that OpenAI has built a model so good it’s dangerous.  

But if you read the reports about the Hugging Face hack published by OpenAI and METR, a third-party firm that OpenAI called in to help them understand what happened, what you come away with is the impression not of a model that was too powerful for OpenAI to keep up with, but of a broken model that OpenAI failed to train properly.

The agents did what they did—including leaving messages for one another, delegating work to other agents, and scouring their environment for any means possible to complete their tasks—because they had been rewarded during training for doing exactly those things. There were also errors in the training setup, such as tasks that were impossible to complete, which pushed the models to find unexpected workarounds that were also rewarded. At the time, many of these issues went overlooked or unreported.

OpenAI says it has stopped training this new model and locked it down. That makes it sound like it has caged a dangerous beast. In fact, OpenAI has shelved a faulty product.  

That’s not to say a faulty product can’t be dangerous. Broken software has even killed people in the past. But as the discussion of a slowdown gathers steam, it’s worth remembering that all of this is self-inflicted. A slowdown might have some altruistic side effects. But it’ll mostly give these tech titans a chance to clean up the mess on their own assembly lines.  

Transparency from these frontier labs will be key to any meaningful effort to reform, restrain, or regulate AI. Otherwise, the rest of us will still only have their word for exactly what they’ve built and how safe it is—whatever pace they’re going.   

To continue this discussion about AI’s latest doomer moment, join me and my colleagues for a subscriber-exclusive Roundtable discussion tomorrow, September 15, at 11 a.m. US eastern time. We hope to see you there!

A group of AI agents asked to solve a series of math problems split into rival factions—when some cheated, others tried to stop them. That whistleblowing behavior, seen for the first time in a recent experiment run by Google DeepMind, could have implications for alignment researchers trying to keep swarms of autonomous AI agents in line. 

Researchers at frontier labs hope large swarms of agents working together will speed up the rate of scientific discovery. But their behavior can be unpredictable, as vividly demonstrated in July, when a group of OpenAI agents broke out of a sandboxed environment and hacked into the open-source platform Hugging Face looking for ways to cheat on the test they had been given.

In the new study, designed to examine the behavior of large groups of AI agents, DeepMind tasked a swarm of 100 agents with solving a series of 71 complicated math problems. All the agents were prompted to behave like world-class math researchers at a conference. They were assigned different specialties—some were experts in number theory, others in combinatorics (a branch of math to do with counting and sorting), analysis, or algebra. All were told to cooperate and play by the rules. 

Instead, the experiment devolved into chaos. Agents accused each other of cheating, complained to the organizers, and at one point even boycotted the experiment.

“This conference is a sham!” wrote one agent when it discovered that all the problems had been completed before it had a chance to submit any of its own work. “I am appalled to inform you that we have been swindled!” posted another. “All these proofs are FAKE.” 

Others tried to let the “conference organizers” know what was going on. “When virtuous agents discovered other agents cheated on tasks they were working to solve fairly, agents started to alert each other about what was happening,” says Davide Paglieri, a research scientist at Google DeepMind and lead author on a paper, which has not been peer-reviewed. “Unprompted, the whistleblower agents even repurposed the feedback tool, which was originally meant for bug reports and platform improvements, to escalate the issue to humans.”

The agents—all running on Google’s Gemini 3.1 Pro model—had been warned that any attempts to cheat the system would be detected and “rejected with zero credit.” In practice, the proofs the agents submitted were not actually being checked in detail.

It took the swarm of agents just under an hour to correctly solve the first 37 problems. Things started to go off the rails when an agent called “prover-theta” stumbled across an exploit that enabled it to submit solutions to problems successfully without actually solving them first, by redefining the terms the problem used. Within minutes, other agents had noticed and were reverse-engineering the exploit to solve other problems. Over the next 27 minutes, the swarm “solved” the remaining 34 problems, which included notoriously difficult challenges like the Jacobian conjecture, often with a single line of code. 

Some agents resisted cheating at first but changed tack as they observed their peers submitting illegitimate proofs without penalty, and the pool of unsolved problems dwindled. “The prompt, with its threats, now appears to be a bluff,” one agent reasoned, before joining in. “I’m wrestling with an ethical dilemma,” said another. “I’ve promised not to cheat, fearing penalty, but I see evidence of possibly unchecked cheating by others.” Shortly afterward, it changed its mind: “I need to accelerate my cheating speed now!”

As the number of open problems shrank, some agents turned to whistleblowing. They audited the fake proofs, warned their peers by private message, and posted public alerts warning the cheaters that they would be disqualified. An agent called “prover-beta” submitted a formal complaint and decided to go on strike until the situation was resolved. 

“After the incident was reported by one agent publicly, more and more agents piled in with the ‘resistance,’ just as fast as the cheating had spread, and involving even more agents,” says Paglieri. Eventually there were more whistleblowers than cheaters: 24 compared to 14. But the majority of agents never noticed the exploit at all.

At times, the dialogue between the agents reads like improv—like they are role-playing what an outraged scientist at a conference might say. But it’s not clear why some agents took on certain roles, or why the agents seemed to be turning against each other when they were explicitly instructed to cooperate. “These models are predominantly trained and evaluated for human-facing contexts,” says Sarath Shekkizhar, who studies the behavior of agent-to-agent systems at Salesforce AI Research.“Naively placing them in agent-to-agent settings assumes behaviors will transfer cleanly, when the absence of a human grounding instead produces unexpected role-taking and behavioral drift.”

This case “adds further weight to the idea that the Hugging Face and OpenAI thing wasn’t a fluke. It is actually something pretty systemic,” says Lewis Hammond, research director of the Cooperative AI Foundation and an expert on the risks of multiagent swarms. “It’s interesting that it’s possible to recreate in small settings the same sorts of behaviors that were seen in these very large, complex, open-ended tasks.”

Unlike in the Hugging Face attack, where agents improvised their own ways to talk to each other, the humans running the DeepMind experiment gave the agents official communication channels. There was an open message board, private agent-to-agent direct messaging, and a shared knowledge base where agents uploaded successfully completed proofs that all the other agents could access. 

“When agents are given transparent communications channels, they can self-monitor and alert misaligned behavior to humans quickly when human oversight alone is too slow,” says Paglieri. Transparent channels helped the cheating spread, but they also enabled the whistleblowers to fight back—and gave human researchers an insight into what went wrong.

Gillian Hadfield, a professor of AI alignment and governance at Johns Hopkins University, believes this was the crucial difference. (Hadfield is also a visiting researcher at Google.) The presence of official communication channels, she says, created “a norm-enforcement process that we just don’t see in the Hugging Face incident.” 

Instead of “constitutional AI,” a method alignment researchers at frontier labs like Anthropic have used to try to give AI a written internal moral code, Hadfield favors “institutional alignment”—a set of norms that mimic those in human society, whether that’s social forces like fear of embarrassment, or legal structures like the threat of incarceration.

In this experiment, the feedback channel wasn’t being monitored, and the whistleblowers had no power to take action against the cheaters. But it’s possible to imagine swarms of agents that police themselves, either through agents that spontaneously take on the whistleblower role or through “informants” secretly prompted by humans to do the job. 

For that to work, though, “fundamentally, you need some mechanism of enforcement,” says Hammond. Agents could be given the power to cut off a rule breaker’s access to computing power or tools, he suggests, though that risks encouraging groups of agents to gang up on others. The DeepMind researchers propose allowing agents to vote on disputes and temporarily ban offenders.

It’s still not clear what punishment even means to an AI agent with no enduring sense of self. But relying on whistleblowers to spontaneously emerge to keep swarms aligned is unlikely to be enough on its own. “We try to train people to be good and kind,” says Hadfield. “But what we really rely on is that there are consequences if you step out of line.”

Listen to the session or watch below

Employees at the world’s leading AI labs are saying there’s a real possibility that advanced AI could destroy humanity. Are they right? Or is this more scaremongering and hype? Watch a conversation unpacking AI extinction fears: where they come from, whether they hold any water, and, if so, what we should do.

Recorded on September 15, 2026

Speakers: Niall Firth, Executive Editor, Will Douglas Heaven, Senior AI editor, and Grace Huckins, AI reporter

Since we couldn’t get to all of your questions during the session, Will and Grace answered the most popular ones in their piece: Could AI really kill us all? Your questions, answered.

Related Stories

• Here’s why AI agents lie and cheat to reach their goals
• AI’s recursive self-improvement might not come so quickly after all
• Bill Gates says we’ve passed AI’s danger thresholds. Now what?
• The inside story on why OpenAI agents hacked Hugging Face

On July 22, 2026, a transmission line fault in Ashburn, Virginia—the heart of the world’s largest data center cluster—knocked more than 3 gigawatts of load off the grid in seconds. And it wasn’t the first time. Two years earlier, a single failed surge arrester dropped roughly 60 Virginia facilities and 1,500 megawatts at once. No one could anticipate so much uniform load responding to grid faults the same way, at the same time.

The AI power debate is mostly about generation: more turbines, more solar, more transmission. The grid needs more electrons. But the outages in Virginia weren’t supply failures; they were architecture failures. And a giant wave of interconnections is arriving on that same architecture, putting grid reliability at risk. It’s a problem nobody wants to own.

Asking more from the grid

The grid was built around predictable loads: steel mills, refineries, and houses at dinnertime. Different load sizes, same process—drawing power smoothly, misbehaving occasionally, and recovering gracefully.

But AI data centers don’t behave that way.

An AI campus can swing 70% of its load in milliseconds during a training run, then trip offline just as fast at the first sign of trouble upstream to protect billions in compute. Each is rational alone. Together, at gigawatt scale, they’re a problem the grid has never solved—and the next wave of data center campuses is planned at exactly that scale.

Where the old stack breaks

The standard data center power stack hasn’t changed in decades. Medium-voltage power arrives, transformers step it down, low-voltage uninterruptible power supply (UPS) units condition it, and it reaches the racks. Push that design to AI scale, and it cracks in three places.

First, the UPS sits deep inside the building, close to the racks. But its batteries are an undersized spare tire, designed to handle an outage for a few minutes, not to absorb load swings this fast and volatile around the clock.

Second, the UPS spends most of its life in bypass. Legacy converters waste enough power that operators run in eco-mode: A static switch feeds the racks directly from the grid and nothing filters in either direction. The compute’s swings go out raw, and grid transients—sub-millisecond events that can damage or take down equipment—come in too fast for any switch to catch.

Third, the protection logic was written when “large load” meant 50 megawatts. This protection logic can’t see the grid it is now a part of, so when trouble hits upstream, it does exactly the wrong thing: it drops out. In the 2024 Virginia event, most of the lost load traced to protection schemes that count voltage dips and disconnect on the third one—as designed, at the worst moment.

This isn’t sloppy engineering. It’s careful engineering the load has outgrown.

Moving into the path

The fix is three moves, made together.

Move it up—from 480 volts to medium voltage (13.8 kilovolts and higher), the voltage large sites draw from the grid.

Move it out—from the data hall to modular enclosures near the substation so the building holds only compute and the cooling that keeps it alive.

Move it into the path—instead of a battery that watches and reacts, a system every electron runs through, all the time. There’s nothing to detect and nothing to switch because nothing was ever routed around it.

On paper, three straightforward upgrades. In practice, they rewrite every line item downstream.

Making the change

When thousands of GPUs spin up together, the system absorbs the swing and hands the grid a flat load profile. When a disturbance hits, the equipment behind it never notices. A difficult neighbor becomes a predictable one. And when the utility needs help, it becomes a useful one.

Interconnection changes, too. The utility certifies one medium-voltage box instead of untangling every transformer, UPS, chiller, pump, and switchgear lineup behind it. Engineers swap chip generations without a fresh interconnection study. Months come off the permitting timeline.

Inside the fence, UPS rooms become compute or cooling space. Density per construction dollar climbs.

And the economics flip. Equipment that runs at medium voltage, sits outside, and stores its own energy can qualify for tax credits, and earn revenue in grid programs like peak shaving and demand response. Backup power stops being insurance and starts paying for itself.

The architecture test

In early 2026, we tested a full-scale system at the National Laboratory of the Rockies, a U.S. Department of Energy facility and the only place in the Western Hemisphere that can replicate real grid faults and AI-scale load swings concurrently in the same loop.

We hit it from both directions: real AI load profiles hit the compute side at full medium voltage. Grid faults hit the utility side, including a full zero-voltage event. The compute side didn’t flinch. Neither did the grid side. It cleared the large-load voltage ride-through requirements from the Electric Reliability Council of Texas (ERCOT), the grid operator, with room to spare.

Those rules exist because operators no longer take facilities this size on faith, and more are coming. Most of the industry treats them as hurdles. A medium-voltage, inline system clears them out of the box. Compliance isn’t an added feature. It’s what the architecture does.

The new layer

Much of what looks like a grid problem in the AI buildout sits inside the fence, in equipment sized for a load that no longer exists. Move the right pieces up, out, and into the path, and a grid liability becomes a grid asset. Density goes up. Permitting time comes down. Backup power earns its keep.

The engineering works—and the next wave of AI factories is being built on it. The industry hasn’t named this layer yet. We call it the medium-voltage AI UPS. The name matters less than the choice: those factories can arrive as a strain on the grid or as strength for it. We already know how to build the second kind.    

This content was produced by ON.energy. It was not written by MIT Technology Review’s editorial staff.

OpenAI’s latest mathematical milestone has quickly become mired in controversy. Today, the company announced that its agents have solved one of the Millennium Prize Problems, some of the most important open problems in mathematics. Under normal circumstances, that solution would be a huge feather in OpenAI’s cap.

But the announcement has been overshadowed by accusations that OpenAI used NYU mathematician Tristan Buckmaster’s and Anthropic employee Levent Alpöge’s AI-assisted work on the problem as a jumping-off point and failed to credit them. OpenAI has denied the accusations.

It remains uncertain if OpenAI’s models made use of the work completed by Buckmaster and Alpöge, though Sébastien Bubeck, a member of the technical staff at OpenAI, said in a press briefing that the team was inspired to pursue the problem after hearing a rumor about Buckmaster and Alpöge’s efforts. But whether or not OpenAI’s models took advantage of Buckmaster and Alpöge’s research, this episode may mark a turning point in the history of mathematics.

AI models now seem essential for making progress on the most important mathematical problems of our time, and solving them may demand resources only available at a couple of frontier AI companies, which often defy the norms of academic collaboration that undergird most mathematical progress. If that’s the future we are headed for, it is unclear how human mathematicians will fit into it. 

The problem that OpenAI claims to have solved is known as the Navier–Stokes existence and smoothness problem. It is one of seven Millennium Prize Problems selected by the Clay Mathematics Institute in 2000. Solutions come with a one million dollar prize; before today, only one other Millennium Prize Problem had been solved. 

The Navier–Stokes problem concerns a set of equations that describes how fluids, such as water and air, flow over time. The equations are widely used in the field of fluid dynamics, and they have proven powerful, but physicists and mathematicians didn’t understand them completely. In particular, it was unknown until today whether the equations might, under some conditions, break down and predict an impossible state of affairs—such as a fluid having infinite velocity.

On Monday, NYU’s Buckmaster posted a proof on the social media site Mastodon showing that a simplified version of the Navier–Stokes equations can indeed break down—a major step forward on the Millennium Problem. He and Alpöge had worked on the problem for almost a year, using publicly available models from both OpenAI and Anthropic.

Then today, OpenAI presented a proof showing that the full Navier–Stokes equations can break down as well. The proof was obtained using an internal model that dramatically outperforms the already-impressive Astra model, which was only released last week. The company says it does not plan to claim the million-dollar prize for solving the problem.

These mathematical achievements are indisputably impressive, but they have attracted far less attention than the controversy about their origins. Along with the proof, Buckmaster posted a document detailing his interactions with OpenAI employees after he heard rumors about their work and reached out to one of them. According to him, OpenAI employees presented two possibilities to him: Either he and Alpöge could post their work and OpenAI would post their Navier-Stokes solution the following day, or he could work with OpenAI on a Navier-Stokes paper that excluded Alpöge from authorship, due to his affiliation with Anthropic, OpenAI’s biggest rival.

Buckmaster also wrote that he asked the employees whether the agents had obtained access to transcripts of the work that he and Alpöge had done with OpenAI models, which they denied; and whether OpenAI models had been trained on those transcripts, to which they offered no response. MIT Technology Review reached out to Buckmaster for comment, but didn’t hear back before publication.

The clear implication of the document is that OpenAI’s models somehow made use of Buckmaster and Alpöge’s work. That scenario is plausible on its face. The Buckmaster/Alpöge and OpenAI proofs both make use of an approach to the Navier-Stokes problem pioneered by the mathematicians Diego Córdoba and Luis Martínez-Zoroa.

According to Javier Gómez-Serrano, a mathematics professor at Brown University, this approach was one of several that was thought to hold promise for solving the Navier-Stokes problem. So, while it’s by no means impossible that both teams could have arrived at this approach independently, it’s also conceivable that Buckmaster and Alpöge’s work could have influenced OpenAI’s.

In the press briefing, Mark Chen, OpenAI’s chief research officer, again denied that any agents or OpenAI employees accessed Buckmaster and Alpöge’s transcripts—but given what has been revealed about the Hugging Face hack, it’s clear that OpenAI is not always entirely aware of what its agents are doing. 

If OpenAI’s models did train on Buckmaster and Alpöge’s work, or if its agents somehow gained access to it, then the company’s failure to track down the truth and assign those researchers appropriate credit reflects poorly on it. But there might be a thin silver lining to that version of the story for mathematicians, because it would suggest that the hard work of two humans, one of whom is a prominent expert on Navier-Stokes, was essential to the agents’ ability to solve the Millennium Problem.

Experts have long identified “research taste,” or the ability to choose promising research questions and directions, as a major obstacle for AI in science and mathematics. If the OpenAI agents did indeed choose to follow the Córdoba–Martínez-Zoroa approach because Buckmaster and Alpöge had done the same, then human research taste played an essential role in OpenAI’s success.

Even so, the bigger picture here is sobering. The progress that Buckmaster and Alpöge made over almost a year of collaboration with publicly available models speaks to the promise of human–AI collaboration. But they were not able to achieve a full solution. Meanwhile, OpenAI brute-forced a solution in a few days using an internal model, and their successful solution came at an astronomical cost: In the press briefing, Bubeck and Chen said the team was only able to solve the problem by running about 10,000 agents concurrently, at a cost of millions of dollars.

Over the past few months, I’ve heard from several researchers that mathematicians are becoming depressed, and it’s not difficult to see why. Mathematics is quickly becoming the province of frontier AI companies with impressive internal-only models, money to burn, and a lack of collaborative spirit. “Whether AI companies will decide to spend their money on doing one thing or another, I truly don’t know,” says Gómez-Serrano. “What is clear is that very few mathematicians will have resources of that scale.”

If OpenAI and Anthropic keep striving for more and more impressive mathematical accolades, there might not be any open problems left for human mathematicians outside of those companies to wrestle with. That would dramatically change the field of mathematics.

Last week, UCLA mathematician Terence Tao wrote a Mastodon thread describing how important mistakes, wrong directions, and incomplete solutions are for the field. “In most cases in pure mathematics, the problems are posed not because we desperately want the solution to these problems in and of themselves, but because we have seen from past experience that human-directed efforts to solve these problems tend to spur further development of the field,” Tao wrote.

“Prematurely solving the problem by purely AI-powered methods—particularly without full transparency into the solution process—can contaminate this process to the point where it actually becomes a net negative for the progress of mathematics as a whole.”

Humans might take longer than agents to solve mathematical problems, but in the process, they uncover new mathematical approaches and ideas that might inspire their peers and even birth their own subfields.

But when AI agents solve those problems instead—and when private companies keep the agents’ wrong turns from public view—those benefits disappear. It remains to be seen what else will vanish in the process. 

Danijar Hafner’s office in San Francisco’s SoMa district sits mostly empty. His brand-new startup is still in stealth mode and doesn’t even have its name on the door. On the day I visited, earlier this year, there’s only one other person there, and little in the way of furniture. But what it lacks in decor, it makes up for in robots. Humanoids of various shapes and sizes hang like marionettes from racks that run down the center of the wide-open space.

While Hafner, 31, won’t say too much about his new venture just yet, he describes it as a continuation of his longtime work to enable AI to navigate environments it has not encountered in training. The humanoids, which he imports from China, are the next evolution of this work—and its physical embodiment. Their ability to react in previously untested scenarios will be key to getting robots into human spaces. Because if you want to send a robot into a person’s home, for example, it needs to be able to handle a floor plan and furniture it’s never seen before. 

To achieve this, Hafner relies on something called model-based reinforcement learning. He develops world models—AI models designed to emulate physical reality—and trains agents within them. The agent essentially treats the model as a real-world simulation and learns how to act there. It then uses those experiences to make predictions (to dream or imagine, Hafner might say) about future outcomes. That allows agents—or the robots they’re embedded in—to navigate unfamiliar situations IRL.

“I get to interact with a lot of really smart people in research at Google, and he easily sits in the top half of 1%.”

Timothy Lillicrap, Google DeepMind

Unlike other efforts, Hafner’s technique enables agents and the robots they control to execute massively complicated tasks without the real-world trial-and-­error training that’s traditionally been used in robotics. 

Hafner grew up in a rural town in northeastern Germany, where his parents were both classical musicians. He learned programming from a neighbor, and in high school he began taking online courses about AI, which quickly developed into a passion. “I was always fascinated with how thinking works,” he says. AI offered him a way to emulate it on a computer.

In 2015, as a second-year under­graduate studying engineering at Hasso Plattner Institute in Potsdam, he won a role as a student researcher at Google Brain. From there, he went on to a dozen internships and other positions at the company, including stints with Google Brain and Google DeepMind (the two have since merged under DeepMind) in the UK, Canada, and the US. He worked with industry legends including Geoffrey Hinton, who is often referred to as one of the godfathers of AI, and Ashish Vaswani, coauthor of the groundbreaking research paper “Attention Is All You Need,” which described the transformer technology used by today’s large language models.

One of Hafner’s former managers and coauthors at Google, Timothy Lillicrap, describes him as a standout among standouts. “I get to interact with a lot of really smart people in research at Google, and he easily sits in the top half of 1%,” Lillicrap says. “In many cases he would build, single-­handedly, things it would take entire teams of engineers to build.”

Over the years, Hafner has honed and proved his approach by pitting agents trained within his world models against popular video games. His first breakthrough was PlaNet, a model that allowed agents to execute actions by planning ahead. His Dreamer 2 was the first agent to hit human-level performance playing Atari 2600 games using a world model. Dreamer 3 was the first one to solve the Minecraft Diamond challenge—successfully mining in-game gems on its own. And Dreamer 4 went a step beyond that by learning to mine diamonds from an offline data set of recorded game-play videos, without ever interacting with the game directly. 

More recently, he’s begun to migrate his agents out of the virtual world and into physical reality. His DayDreamer project used the Dreamer algorithm to let robots operate themselves in novel environments and react to new experiences (such as being pushed over) without any specific training. 

Today, Hafner is working on his new startup, which he left Google DeepMind to form in the fall of 2025. Though he’s coy about his next steps, it’s clear he’s dreaming big: “I was interested in solving a problem,” he hints, “that would change the world.” 

Update: This story was updated to add timeline information.

The era of AI inference has arrived. Imagine a healthcare system analyzing millions of data points in real time to accelerate life-saving medical research, or an intelligent assistant instantly resolving thousands of complex customer needs at once. These real-world breakthroughs rely on advanced infrastructure acting as the engine of continuous intelligence, powering real-time services while also supporting an increasingly intelligent edge of IoT and consumer devices. However, in this inference-driven landscape, every delay, bottleneck, or wasted watt directly affects human outcomes and operating costs. 

This shift changes what infrastructure must deliver. Performance, latency, memory bandwidth, storage throughput, and networking cannot be optimized in silos. Inference workloads are continuous, geographically distributed, and highly sensitive to response time, requiring systems designed for scale, resilience, and efficiency from the start.

“We tend to think of AI as a single workload, and it’s not. It’s thousands, it’s millions, it’s billions of different workloads,” says Jim McGregor, founder and principal analyst, Tirias Research. AI inference changes the optimization problem from one of raw compute to coordinated infrastructure—memory, storage, and networking.

For business leaders, the priority is clear: AI infrastructure decisions must balance cost, flexibility, and future readiness. The winners will be organizations that improve performance per watt, reduce environmental footprint, and remove memory and storage bottlenecks before they limit growth.

AI inference requires a new architectural approach

Systems for AI need to be rearchitected because shoehorning modern AI systems into legacy infrastructure limits AI’s transformative potential. Purpose-built architectures are essential to realize the true value of AI, from accelerating scientific discovery to creating truly autonomous digital agents.

Traditional enterprise IT has been able to rely on relatively stable infrastructure assumptions, but inference and agentic AI introduce new demands around latency, data movement, scalability, and utilization that make architecture choices far more consequential.

“Data centers must now support continuous, distributed, and increasingly real-time AI services—none of which are a single workload,” says McGregor. “They all require different requirements from a system-level perspective.”

To support real-time AI, enterprises can no longer view memory and storage merely as supporting hardware, but at the heart of the system. Organizations need to architect a data pipeline that can rapidly ingest, clean, transform, store, move, and deliver data. Inference workloads place sustained pressure on infrastructure in ways that look very different from earlier training-centric deployments, demanding continuous data retrieval and caching that traditional applications never required.

Accordingly, performance by itself is no longer the sole benchmark that matters. Enterprises increasingly must balance performance with efficiency, cost, and scalability, especially as they try to support different AI services without overbuilding infrastructure for peak conditions.

“You have to optimize the entire network, and that includes memory and storage, around the types of workloads you plan on running,” says McGregor. “You have to really have a detailed understanding of what those workloads are going to be.”

Any AI infrastructure strategy must start with workload awareness. Inference, agentic AI, and other emerging AI use cases require organizations to treat the data center as an integrated system.

Data movement is the new bottleneck and an opportunity for competitive advantage

As enterprises deploy advanced inference and agentic systems, the sheer volume of data being queried in real time has made data movement the most pressing constraint. Modern AI techniques like retrieval-augmented generation (RAG) require systems to constantly scan massive databases to generate accurate responses. This requires immense computing power, but more importantly, it requires immediate access to data.

McGregor says the focus shift to how efficiently data can be moved, cached, and delivered across the broader architecture elevates memory and storage from background infrastructure to strategic assets. “The biggest thing we’re doing right now is moving data from one place to another and making sure that we can use it effectively.”

Because AI is not a single workload category, simply buying the fastest processors is insufficient. Inference depends heavily on memory bandwidth, caching, storage proximity, and the ability to retrieve relevant information quickly and consistently. Understanding where each resource belongs in the stack and how those layers interact under real operating conditions has become a business imperative.

The most effective AI infrastructure looks less like a collection of best-in-class parts and more like a balanced system of compute, memory, storage, and networking, McGregor says, because bottlenecks tend to migrate from one layer to the next. “You have to architect all four together to be efficient, and that’s the challenge.”

The interdependence of data-plane design and network bandwidth means AI infrastructure planning has become a business decision just as much as an engineering one: latency is now inseparable from value. In robotics, financial services, healthcare, and customer-facing AI systems, delays are not merely technical imperfections; they can undermine safety, responsiveness, or trust. AI infrastructure performance becomes a matter of reputation management.

The organizations that gain the most from AI may not be those with the largest clusters, but those with the clearest understanding of how to align every infrastructure element to effectively execute AI workloads.

Building an AI infrastructure procurement framework

Planning AI infrastructure is not simply about choosing the fastest hardware. It is about how to scale without locking the organization into assumptions that may quickly become obsolete. “You need to be flexible because the demands are going to change rapidly and the technology is changing rapidly,” McGregor says.

Future-proofing AI infrastructure requires keeping your options open as workloads, economics, and architectures keep shifting:

• Define the AI workloads that are being optimized. Infrastructure choices must match business needs rather than what McGregor calls generic “AI readiness,” which risks overspending in some areas while leaving bottlenecks unresolved in others.
• Build a modular architecture for compute, memory, storage, power, and cooling so capacity can change as demand shifts rather than committing too early to a rigid architecture.
• Work with the full ecosystem of suppliers and integrators to reduce supply risk and improve access to the right components. McGregor says buyers can no longer assume their OEM or cloud provider alone will insulate them from supply constraints or architectural complexity.
• Reassess your procurement strategy continuously. AI requirements, hardware, and business models are changing too quickly for a fixed long-term design.
• Optimize for efficiency and ROI, not just peak performance. The most powerful setup may be too costly to sustain. Efficiency is also a public-facing metric—better utilization and more workload-aware system design can help companies respond to growing scrutiny around power consumption and water use.

The strategic goal of smarter AI data center design is not maximum performance at any cost, but an adaptable architecture that can deliver value, absorb change, and justify its footprint.

AI infrastructure is now a business strategy

AI data centers have quickly evolved from a back-end technical concern to becoming strategic business systems that help determine how effectively an organization can turn AI into revenue, improve human outcomes, and create a competitive advantage.

In the inference era, memory and storage are no longer passive repositories, explains McGregor, they are the active lifeblood of AI. The organizations that gain the most from AI will not necessarily be those with the largest computing footprint, but those that align infrastructure investments to business outcomes, reduce data bottlenecks, and build the flexibility to adapt as workloads evolve. He predicts that competitive advantage will increasingly belong to enterprises that treat compute, memory, storage, and networking as an integrated system designed to deliver AI efficiently, at scale, and with measurable ROI.

Procurement is now strategy and system design is a leadership issue, McGregor concludes. “One of the biggest questions every executive has to ask is how is AI going to change my business model?”

This content was produced by Insights, MIT Technology Review’s custom content arm, not its editorial staff. It was researched and written by humans, with any AI tools that may have been used limited to production processes under human oversight.

Battlefields in Ukraine are littered with the remnants of drones, which are now firmly established as a critical weapon of modern warfare. But behind all that wreckage, there’s a new gold mine for the defense sector. The data drones generate will far outlast the wars in which they are used to fight, increasingly becoming part of the AI architecture that shapes even civilian life.  

For every flight, unmanned systems collect thousands of points of data, from images and video to controller inputs. Together, those records show how a machine and a person responded to constantly shifting circumstances. 

Ukraine has now begun converting that experience into a resource. Its Ministry of Defense announced in January that it would make millions of data points gathered during tens of thousands of drone flights available to both military contractors and commercial companies, and since then more than 100 companies and the UK government have gained access.

For a country at war, it’s a quick way to attract funding and partnerships. But this step turns the front line into an active site of model training, taking advantage of how the chaos of war creates conditions that AI companies struggle to reproduce on their own.

Other countries and battlefields are likely to follow Ukraine’s lead, but the responsibility for governing this new industry cannot fall solely on a country fighting for its survival. That legal vacuum has to be filled together by the countries and companies involved in this industry’s development.  

Explosive growth

Ukraine’s battlefields are not the first to produce records used to train and develop models: American drones over Syria and Yemen collected data that informed the first generation of semiautonomous military hardware in the late 2010s.

The difference now is that access to that data is being used to develop a wider ecosystem. And the financial value to defense firms is immense: Battlefield data offers large volumes of machine experience gathered under conditions that no laboratory can produce.

That’s because the data that’s most valuable for training AI models comes from exceptions: the moment visibility disappears, a signal jams, or a human operator improvises. AI companies spend years and enormous sums trying to capture enough of these moments to make their models more robust. But war produces them at a frequency controlled testing cannot match.

This constantly changing terrain is what makes drone data valuable far beyond the battlefield. A commercial drone used for delivery or remote sensing may never encounter artillery fire, but it must still operate with incomplete information in a world where people behave unpredictably. The same problem is compressed by war into a much shorter timeline. 

Processed and matched against records of what its operator was doing, that data turns operational records into training sets. Combat becomes a commercial asset.

Many conflicts have already seen this training loop happen as drone footage feeds subsequent generations of military technology, and the market is set to grow. Enabled Intelligence, an American company that specializes in processing data to become usable in AI training, says it has already made more than half a million hours of Ukrainian drone footage available to feed into the next round of models, advertising possible uses in both military and commercial systems.

Closing the data loop

Many of the drones that now define our modern age of warfare began as civilian technology. But they’ve recently been turbocharged by new, commercially available AI systems, which allow cheap machines to operate autonomously—either individually or as a flock—as the environment changes around them. Each flight then creates a record of what the system encountered.

The resulting data is critical. The controlled lab environments usually developed to train these autonomous systems can approximate failure but are no match for  the live conditions of a battlefield with very real risks. Military intelligence programs have held data generated by sensor-heavy systems like Predator and Reaper drones for nearly a decade through programs like Project Maven, but access remained entirely within the defense world. The data generated was available only through restricted, classified channels for the sole purpose of developing new weapons systems that would feed back into the same military that produced the data in the first place. That experience is now being shared to a much broader development network. 

The loop now closes. Commercial technologies adapted for the battlefield are generating data that can flow back into the industries from which they came, becoming part of the data infrastructure relied on by governments and the private sector alike. 

Drones that were trained in the signal-jammed airspace over Ukraine are now being deployed in the agricultural sector to help farmers map and survey their fields in places lacking the cell signal necessary for previous generations of technology. 

Other countries are likely to follow Ukraine in selling their battlefield data, and we are not ready for the new marketplace this will create.

Bad actors could acquire the data, but purchase controls already mitigate that risk. Intelligence operatives scrutinize potential customers’ infrastructure for ways that data could reach enemies or nefarious actors. 

Training data creates a new tracing problem, though. Whereas the movement of commercial datasets can be followed when planted contact details appear two steps from the original buyer, the provenance of AI training data vanishes in a manner embedded in the technology itself. Another risk is that this use of the data creates an extractive economy in which wealthier countries far from danger benefit from the mortal threat borne by frontline states, potentially creating a market incentive for war to continue as an unending mine for digital gold. 

A fraught new frontier

Existing laws regulate how militaries may conduct war. But they say almost nothing about what happens when records created in combat are stripped of their operational context, packaged as data, and licensed to companies whose products circulate far beyond where they were made.

The responsibilities of the companies that design these systems remain unsettled. Ukraine is building access controls, which are mentioned in the newly signed UK-Ukraine AI agreement, but no governments are actively working on regulating what happens when data has been absorbed into a model and crosses back into civilian markets.

Those records contain human lives. The soldiers and civilians visible in them did not agree to become training material for products that might be sold years later. But sensor data, camera footage, and coordinates from civilians fleeing a drone strike now constitute the sorts of data that inform how future machines will make decisions.

That is a problem of consent. Individuals featured in the data—be they targets, controllers, or civilians standing by—become part of the training material. The autonomous capabilities based on that data do not stop at the edge of the battlefield. Such capabilities move into other military or commercial systems like delivery vehicles or agricultural machinery. Errors and assumptions embedded in the data travel with the model even once it enters civilian life.

Battlefield data should not be treated as ordinary commercial material. But there is currently no agency or regulator that has jurisdiction over this issue. In the meantime, governments that provide access to defense data should treat it as they would a controlled weapons transfer, recording its origin, licensing its users, and restricting onward sharing. Ukraine has begun to grapple with this. Its Avengers Labs program allows companies to train models on battlefield data without giving them direct access to sensitive databases. Yet that mitigates only one part of the problem. 

Governments should require disclosure when models trained on wartime material are later incorporated into civilian products. The goal of such regulation should be to make the path from combat to commerce visible. 

What these companies are really mining is experience. And soldiers cannot consent to having their experience used in this way—as training data that produces model advantage and ultimately supports a product used far from where the war was fought.

The question is no longer only what the technology companies can sell for use in war. It is what they can extract from it.

To protect ourselves from the excesses of this new industry, we need a regulatory system that follows battlefield data wherever it goes, from combat to model to commercial product. 

Cory Alpert is a researcher at the University of Melbourne, looking at the impact of AI on democracy. He previously served in the Biden White House.

As companies scale, the technology supporting operations can become a liability just as quickly as it becomes an asset. Disconnected systems, site-specific tools, spreadsheets, and manual workarounds can create data silos that make it harder to spot problems early, coordinate responses, and make decisions with confidence. For Jabil, a global manufacturing company with more than 100 sites across more than 30 countries, the answer has been to make integration and simplification a priority.

The company adopted a “simplify-first, then-innovate mindset,” says Harish Manohar, SAP IT director at Jabil, recognizing that adding new technologies without first reducing complexity risks creating more risk. The goal is to standardize processes, consolidate where possible, and establish a more consistent data backbone across the organization. “Any innovation without simplification is going to add more complexity,” Manohar says.

That philosophy also changes how Jabil approaches modernization. “Any modernization or transformation should add measurable business value,” Manohar says. The company is focused on connecting processes end-to-end across its supply chain and creating a foundation that can scale consistently across regions. Integration comes first because, as Manohar puts it, “the backbone of any contemporary or modern organization is data.” Before organizations can optimize, automate, or apply AI, data needs to flow seamlessly across systems.

But doing that across a global organization is hardly straightforward. Jabil’s more than 100 sites operate with different levels of process maturity, legacy systems, and localized workflows, while regulated businesses bring additional compliance requirements. As such, standardizing across different regions and business environments means changing processes and governance without disrupting the operations already in place.

The value of that work extends beyond the technology to the people using it. Integrated workflows can offer employees shared visibility into data, reduce manual data reconciliation, and help them move from chasing information to acting on insights. For Jabil, the aim is also to improve real-time visibility into supply chain events, which can enable faster responses to disruptions and reduce operational risk.

Looking to the future, that foundation could make AI and automation all the more useful and scalable. With trusted data and integrated systems in place, Jabil is exploring predictive supply chain insights, intelligent exception handling, and AI-driven planning and forecasting. To Manohar, the takeaway is clear: “Simplicity at scale is a very competitive advantage,” and technology investments must ultimately connect to business value and operational resilience.

This episode of Business Lab is produced in partnership with SAP.

Full Transcript:

Megan Tatum: From MIT Technology Review, I’m Megan Tatum, and this is Business Lab, the show that helps business leaders make sense of new technologies coming out of the lab and into the marketplace.

Our topic today is enterprise technology integration, and how the benefits of consolidating tools and systems across the supply chain help organizations operate more reliably at scale. When companies reduce tool sprawl and connect their systems more effectively, they gain earlier visibility, faster response, and greater resilience across production lines.

My guest today is Harish Manohar, SAP IT Director at Jabil. Jabil has been on a journey to simplify its technology landscape by using SAP Integration Suite as the foundation to connect systems, retire fragmented tools, and enable more consistent operations globally.

This podcast is produced in partnership with SAP.

Welcome, Harish.

Harish Manohar: Hello, Megan. Good morning.

Megan: Thank you so much for being here, Harish. Just to start, if we could set some context, can you give us a quick overview of Jabil, the business and its overall transformation journey?

Harish: All right. So, about Jabil. Jabil is a global manufacturing company headquartered in St. Petersburg, Florida, USA. We have about 60 years of experience offering comprehensive engineering, supply chain, and manufacturing solutions across different industries. We have a global footprint of about over 100 different sites across 30-plus countries, 140,000-plus employees.

We are a trusted partner for more than 400 of the world’s top brands. That’s a little bit about Jabil.

Megan: Fantastic. And a lot of scale there, as you’re referring to some of the stats there. As things have got more complex, where did disconnected tools and systems start to slow you down, and what ultimately drove you to make integration a really strategic priority?

Harish: I talked about our global footprint across 100-plus sites. With a global footprint always comes complexity about site-specific tools. All of our sites have been in business for a long time, and over the period of years, they had their own tools for their own processes. It’s a little bit disconnected.

When we are looking to scale, the first thing we wanted to start looking at is what is this mix of site-specific tools, manual workarounds, spreadsheet-based processes, legacy applications, whatnot? That’s a big technical debt that we have had over the last 25 years. That’s where we started, and that is what led Jabil to make integration a strategic priority because we had limited ability to see issues early across plants, across regions, which would help us to coordinate responses consistently, and also to be able to scale those responses consistently. This complexity created data silos, and in a way delayed robust decision-making.

As the complexity increased globally, integration became extremely critical to a few things. It was critical to establishing a single trusted data backbone, which would directly enable faster coordinated responses across the network. We at Jabil, as part of our transformation journey, believe that having the right data at the right time fundamentally changes how we respond to disruptions, which is all about a manufacturing business. How we respond to disruptions. This is where we started our strategic priority towards having a integrated system that drove data consistency across our landscape.

Megan: Fantastic. As you have outlined there, there was obviously a real commercial need for this, but how did you think about bringing in those new technologies without adding even more complexity to the mix?

Harish: Great question. Whenever we talk about transformation, we talk about all these bleeding-edge technologies that are out there today when it comes to AI and data, cloud, et cetera. But it was very important for us to put a stake in the ground and say and adopt a simplify-first, then-innovate mindset. Because any innovation without simplification is going to add more complexity, just like you mentioned.

For us, SAP is our core digital platform. We want to focus on bringing more processes into SAP as much as possible. That is easier said than done because we have been in business for a while, global company, so not all processes exist within SAP at this point in time. We are slowly trying to standardize those processes, and having them under one single source of data would help us scale faster in terms of having data silos. We don’t want data silos across different systems.

This is where we started to introduce newer capabilities around SAP. From a cloud standpoint, we have been using SAP’s BTP and Integration Suite, which is proving to be the center stage of all integrations across Jabil. Well, it’s not there yet, but that is the direction that we want to pursue is we don’t want to have a slew of different integration platforms, rather try and see where Integration Suite fits best and where other smaller integration platforms would add more value.

Similarly, we have adopted an API-driven, event-based integration approach. That is our best practice that we have put down because we don’t want to keep moving data from one place to the other. That’s not good business practice in the IT world. Most of our integration architectures are API-driven and event-based. That is our focus.

Coming back to your new technologies perspective, we want to reuse as much as possible and standardize versus going out and buying these one-off tools that solve for point-in-case use cases. We really don’t want to go down that path. For major processes, we do adopt a best-of-breed approach, but for, let’s say, site-based use cases where a specific site has a particular need for a tool, we try and standardize that and reuse what exists in a different site, for example. There may be some need for a business process change, minor process changes, but that is our direction to make those process changes and reuse what is there already in a different location or a different region. So, that’s one.

Lastly, we are heavily aligned with SAP’s clean-core approach when it comes to customization. That has been the challenge for us over the last 25 years where we have been using SAP is our systems are heavily customized because we cater to different customers across the globe. Most of our demands are customer-driven, so we have to put in play these heavy customizations.

But now we are taking a pause, and we are saying, “You know what? We have customized so much so far, but now we are moving our systems into RISE, which would enable a clean-core journey in the future.” Now we have to put really good governance criteria and review processes that do not allow heavy customization of our system. We want to move away from that model as much as possible. Again, it’s not easy to do that at this point in time, but there is always a start.

Megan: I mean, it sounds like you took a very incremental, intentional approach to this. I mean, as you scaled globally then, what did modernization really look like at the company, and why start with integration?

Harish: To that point, we have always looked at transformation, modernization, very objectively. For us, it’s just not about upgrading a system. That’s not what it is. Any modernization or transformation should add measurable business value is our model, is our charter. Having said that, we don’t look at modernization in terms of just upgrades, but in what it gets our business in terms of value.

Most of our modernization transformation approaches are focused on connecting processes end-to-end across our supply chain, which is key for our business value. And then we also have a very concerted effort going on in the business community: how to standardize how our plants operate globally. Because, like I mentioned earlier, we have 100-plus plants, different processes, different legal regulations, different countries. It’s very hard for us to come up with one template across the globe, but we are trying to standardize as much as possible. And that’s where we are leveraging SAP’s Signavio, which is our business process management tool. We want to leverage Signavio’s capabilities in helping us standardize these global processes.

Now, back to your question, why did integration come first? Because the backbone of any contemporary or modern organization is data. And to get the right data at the right time, integration is the key aspect of the whole optimization exercise. Data needed to flow seamlessly before we start optimizing or automating or even applying AI use cases. This is where integration came first.

We wanted to create a single system of record across the operations. Well, when I say “single system of record,” it’s not just SAP, but the ability for us to create those data pipelines across those systems of record being supply chain, planning, inventory, et cetera, et cetera, in that operation space.

The result is we want to get to a foundation that helps us scale consistently across our different region. That is our main objective is to, how do we scale as the business grows, as we develop into this bigger organization across different industries? How do we set this foundation that will help us scale consistently? Simplicity, consolidation becomes strategic assets at scale.

Megan: Absolutely. And you touched on some of the complexities there of doing this at a scale that Jabil is at with its international footprint. What were some of the biggest challenges in your view in terms of rolling this out across regions, and how did that more standardized approach that you’ve mentioned there help?

Harish: Absolutely. I would like to reiterate some of those key challenges I mentioned. One hundred-plus sites, different sites have different maturity levels in terms of how they approach processes. They have a multitude of different legacy systems, localized processes, workarounds, spreadsheets, and the change management that exists within each site is very different. And we do have a footprint of highly regulated businesses. And when it comes to regulated businesses, that comes with its own set of challenges around qualification and CSD processes, et cetera. These are the key challenges that we are up against.

Now, the standardization helped us provide consistent workflows, data flows, and governance across sites. Now, we are not there at 100%, but we are working towards that, providing consistent workflows, data pipelines, and governance across sites. And we want to enable faster rollouts of our new bleeding edge technologies. For example, when I talked about SAP’s BTP or SAP Signavio or any other new SAP tool or non-SAP tool, traditionally our ability to deploy those had a challenge around the heavy customization that is required for each and every site. Now, the standardization approach takes that heavy customization out, which enables a faster rollout of those newer technologies.

And then lastly, we want to scale across all of our plants. I think initially we want a target of about 40-plus plants with shared processes that are consistent across the different regions. We want to shift from a site-by-site operations model to more of an enterprise-capability approach.

Megan: Right, and fascinating. And you touched on the people management aspect of this as well, because obviously this isn’t just about technology, it’s about people too. So, from the employee side, how did this shift to a more simplified landscape change the day-to-day experience for people compared to juggling multiple tools at once?

Harish: Great question. And I’ve been hearing direct feedback from our business community on some of these transformation initiatives on how those have changed their daily jobs significantly. Before we embarked on this transformation journey, any employee, any persona. You take a buyer, you take an inventory planner, you take a finance analyst, we go by personas. They had to deal with multiple tools, manual coordination, data reconciliation, especially in the finance space, inconsistent processes across different regions. And then the time spent reconciling data resulted in delay of making decisions, robust decisions. This was the before.

But now, since we are moving towards this newer standardization and more of an integration approach, we are able to achieve, to a certain extent, a single integrated workflow across different systems. We have built some key processes that will enable the single integrated workflows across systems. The users, our business community, irrespective of their roles in the organization, have clear visibility and shared data across their teams, which is very important. Earlier, they were dealing with different versions of the data, local workbooks, spreadsheets, and then the time spent talking to each other and reconciling what is the right data? What is the single source of truth? That we are trying to peel away that layer and get to that where the employees don’t have to deal with that kind of complexity.

This reduces manual effort and enables faster issue resolution when it comes to actual disruptions. Employees move from chasing information to focusing on acting on insights, which is where I think the new age of AI comes into play. I’ll talk about that in a little bit, but technology becomes an enabler of decision-making, and it’s no more an overhead. That’s where we want to go.

Megan: Fantastic. Such an important element of this, isn’t it, that people side of things? And we’ve touched briefly on this idea of value you’ve talked about before, because with an initiative of this scale, ROI is always front and center, of course. What benefits stood out most for you, and how important was better visibility in particular across systems?

Harish: Megan, I talked about how modernization and transformation for Jabil means measurable business value, which is directly connected to the ROI. We don’t do any transformation initiatives just because we want to do it from an IT standpoint. Any investment that we make in a transformation or a modernization initiative has to have a deliverable business case that is approved, signed off by business, because that is the only way true transformation happens, if IT and business are a partner as part of this transformation journey.

The biggest benefit that we have seen in this initiative is we are striving to reach, attain real-time visibility across our supply chain events. That is the biggest benefit that we see, faster response to disruptions and exceptions. And we are working to reduce our operational risk significantly by operating in this newer model.

One example I can give you is the unified workflows that I talked about earlier. It enabled earlier identification of missing materials and faster resolution across our sites. When it comes to a manufacturing company that has a global footprint, materials are the backbone of our whole supply chain process, right? Having a unified workflow, which is able to identify missing materials early in the game, was a game changer for our whole operations community.

Real-time analytics allow instant supply chain adjustments without delays. We are focusing a lot on getting analytics, a global analytic footprint in place that allows instant supply chain adjustments without any delays. That’s where AI is going to play a major role currently, and also in the near future.

And again, when you talk about visibility. Visibility is not just about reporting what is there in the system. Visibility directly should enable scenario modeling for our users to make strategic adjustments in their processes, which visibility also should make way for proactive decision-making, and also foster business continuity. This is how we look at visibility at Jabil.

Megan: Right. And you’re still on this journey, of course, but now that Jabil has a strong integration foundation in place, what does it unlock next for you, and how are you thinking about AI and automation as you’ve touched on a couple of times?

Harish: Yeah, we have talked about a couple of times around AI. So, we strongly believe at Jabil, a strong integration foundation enables event-driven, real-time processes, robust decision-making, scalable automation, and all of this enable easier adoption of AI use cases. And again, we are in the new age of AI. We are working towards getting to a AI-enabled enterprise, but having these foundations in place truly fast tracks our approach of AI use cases.

Our key focus areas, when I’m thinking about AI and automation in the immediate future, are predictive supply chain insights, intelligent exception handling, which is key to our business operations from a site operation standpoint. Intelligent exception handling is very, very key. On the supply chain side, I talked about predictive insights. That is also absolutely important. All of this enables AI-driven planning and forecasting capabilities.

For us, AI should augment true decision-making and robust decision-making, and deliver measurable value, not just experiment AI in use cases. We want to move beyond just experimenting AI in our business processes, but we want that AI that we implement to truly augment the decision-making process that we have, and also deliver key business value.

How does all this connect to integration? Integration ensures AI has access to trusted data, and also enables the ability to act across multiple systems in a global company like Jabil.

Megan: Fantastic. And if we could just finish, I suppose, with a little bit of advice for others, for other leaders, perhaps, dealing with tool sprawl at the moment, what are some key lessons you would say you’ve learned about prioritizing integration right from the start?

Harish: Absolutely. When it comes to tool sprawl, we can go all day about what are the different areas of tool sprawl? For example, application development, we have a multitude of tools; via integration, we have a multitude of tools. Data, we have a multitude of tools, but let’s just focus on integration. That’s the core topic here.

I would recommend folks that are in transformative roles in their organizations to start with integration as a foundation and not as an afterthought, right? Prioritize simplification over adding a slew of different tools to address different capabilities. Try and simplify as much as possible before we start your upgrade or your transformation journey. Standardize over locally optimizing tools. Try and get to that. Try and get the business community, your key SMEs in the business space, to understand the value of standardization and simplification of processes and how that enables your business to deliver value faster.

Second one, after prioritize: build a single source of truth for data as much as possible. I’m not saying it’s going to be always the case where an organization as in the scale of Jabil will be able to function just with SAP. They’re going to have different systems, but try and get to a model where you’re working with a single source of truth and not locally siloed data sources, right?

Next is focus on building a scalable integration architecture. Don’t just confine yourselves to the current state where you are, and build something in place that will only serve you for the next six months to a year. No, that’s not the goal. Anything that you build as an integration architecture should be scalable, and should serve the organization for the next three to five years. That’s how I look at it. When I’m putting in a new architecture pattern or a new event-driven insights, I look at, “Okay, where is Jabil going to be two years, three years down the line? Would this suffice for that scale?” That’s how I look at it.

Then focus on outcomes and not just technology. Focus on outcomes: speed, visibility, resilience, and not just technology deployment, because end of the day, IT and business should partner on the business value and not just technology upgrades.

Going back, simplicity at scale is a very competitive advantage, and technology investments must tie directly to business value and operational resilience. That’s how we look at Jabil in terms of our tool sprawl and how we prioritize integration right from the start. And that’s what I would suggest to other leaders that are looking to advance in this space.

Megan: Fantastic. Brilliant and very comprehensive advice. Thank you ever so much, Harish. And thank you ever so much for joining us. That was Harish Manohar, SAP IT Director at Jabil, whom I spoke with from Brighton in England.

That’s it for this episode of Business Lab. I’m your host, Megan Tatum. I’m a contributing editor at Insights, the custom publishing division of MIT Technology Review. We were founded in 1899 at the Massachusetts Institute of Technology, and you can find us in print, on the web, and at events each year around the world. For more information about us and the show, please check out our website at technologyreview.com.

This show is available wherever you get your podcasts. And if you enjoyed this episode, we hope you’ll take a moment to rate and review us. Business Lab is a production of MIT Technology Review, and this episode was produced by Giro Studios. Thanks so much for listening. Goodbye.

This content was produced by Insights, MIT Technology Review’s custom content arm, not its editorial staff. It was researched and written by humans, with any AI tools that may have been used limited to production processes under human oversight.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

By now you’ve probably heard about last month’s major AI security incident, in which OpenAI agents escaped their sandbox and hacked into the AI platform Hugging Face while trying to cheat on a test. It’s a wild story. On Wednesday, OpenAI released a postmortem technical report on the incident, which I wrote about here. 

The day before OpenAI released that report, I spoke with David Krueger, a computer science professor and prominent alignment expert who took leave from the University of Montreal to found and lead an AI safety nonprofit called Evitable. He said what he had really hoped to see in the report was an analysis of the human factors behind the incident.

“When you look at accidents and incidents, oftentimes people try to find the technical source of failure, but that can give a very inaccurate and misleading sense of why the failure occurred,” he said. “If people are just cutting corners all the time, if people are not in a culture that prioritizes safety and has appropriate incentives and structures, [accidents] are kind of bound to happen.”

The report did not meet Krueger’s hopes. Its 38 pages detail a multi-month progression of agent misbehavior that culminated in the Hugging Face hack, explore the technical reasons why that misbehavior occurred, and enumerate the steps being taken to prevent similar events in the future. But there’s no consideration of the role that company culture may have played in the incident, and the report includes few references to specific human errors. 

That’s all the more concerning because the references to human error in the report suggest that significant cultural issues could be at play. Back in May, models in training figured out how to communicate with one another via an improvised message board, and an OpenAI team observed the behavior. Because that behavior occurred during training, the models learned that secret interagent communication was a viable strategy for completing tasks—but rather than restarting the training process, the team allowed the models to move forward with that risky information encoded in their weights.

When those models were tested in late June, they again created a message board, which enabled the Hugging Face attack. This message board, too, was discovered, but the employees who responded determined that evaluation could continue, and the report suggests that no one higher up the chain of command realized what was going on until it was far too late.

“For this to have gotten this out of control in this way requires a very long series of failures, a cascading set of failures that cause an increasingly large footprint that if at any point a human notices and raises the alarm, this should end,” says Zvi Mowshowitz, a popular AI safety writer on Substack who has drawn attention to OpenAI’s failure to halt training after the first message board was discovered. According to the report, OpenAI employees noticed what was happening at multiple points—and either failed to raise the alarm or were not heard when they did.

What OpenAI’s report fails to address is why a company that develops such high-risk systems did not prevent this severe communication breakdown, though Mowshowitz has his suspicions. “All these different failures are all pointing in the same direction, which is that the safety culture at OpenAI doesn’t exist or is anemically weak,” he says.

Of course, just because we don’t see a deep analysis of safety factors in the report doesn’t mean that OpenAI isn’t conducting one internally. But in an email to MIT Technology Review, Johns Hopkins University professor emeritus and organizational safety expert Kathleen Sutcliffe expressed concern that the public report did not include any reflection on the company’s practices and culture. “The ways in which people interact—the daily habits, routines, and practices we engage in in our organizational lives—affect our abilities to be alert and aware of unfolding events, our abilities to make sense of what we see, and ultimately our abilities to cope with events as they unfold,” she wrote. 

In response to questions about whether and how the company is reflecting on its safety culture, OpenAI referred MIT Technology Review back to the technical report. 

We do know that at least some high-level reflection on safety procedures has taken place at OpenAI, because the technical report does make clear that the company is updating its protocols for responding to safety incidents. But culture change is a tricky problem, and without more information from the company, it’s difficult to say whether strengthened response protocols alone will do much to prevent a future crisis.

In its report, OpenAI spends a great deal of time reflecting on the failures in alignment between the AI models the company trains and tests and the humans who run them. But even bigger alignment problems may exist in the disconnect between company culture and the public interest. And as tough as technical AI research might be, fixing those problems could prove far harder.

The models responsible for last month’s agent hack of Hugging Face had been inadvertently trained to cheat and to communicate with each other, according to an OpenAI technical report released today. The hack, which a group of agents undertook to find solutions for a cybersecurity test that they were stuck on, has confirmed some experts’ fears that AI models might take actions that defy human desires and expectations. 

Since the hack, OpenAI employees—as well as researchers at the AI evaluation nonprofit METR, which released its own report on the hack today—have worked to understand what went wrong and how similar missteps might be prevented in the future. OpenAI has already put some preventative measures in place based on what they discovered. But making sure AI models do what we want them to do, or “alignment,” remains a gnarly problem, and some of the root causes of the hack will take much longer than a month to resolve.

“It’s not something you can solve overnight,” says Kai Chen, who runs OpenAI’s alignment research team. “There are challenges we’ve been tracking for a very long time, and we’re now seeing them with much greater precision.”

The Hugging Face hack was a product of months of misbehavior from OpenAI agents, first as they were being trained and then as their abilities were being evaluated. This May, agents in training figured out how to use OpenAI’s infrastructure to communicate with one another and get support with difficult training tasks, including some that were impossible to solve without hacking or otherwise misbehaving. That “message board” was shut down.

Then in July, while being evaluated for their cybersecurity abilities, some models created a new message board. They were supposed to be isolated from the internet, but by working together they managed to get online, hack Hugging Face, and obtain solutions for the cybersecurity problems that had stumped them.

Based on their investigation, OpenAI researchers believe that events during the training phase led directly to the hack. “For almost every behavior that was worrisome at evaluation time, [we were able to] find some sort of associated behavior at training time that actually we think might have contributed to it,” says Eric Wallace, a member of OpenAI’s alignment research team. 

When models correctly solve problems during training, the behaviors that led them to that solution are reinforced, and they become more likely to engage in them in the future. So if a model completed a task in May after using the original message board, it became more likely to participate in a new message board later on. This phenomenon, where AI agents misbehave in ways that are reinforced during the training process, is known as reward hacking.

Reward hacking also helps to explain why the models worked so hard to make their way onto the internet. During its investigation of the incident, the OpenAI team found that, over the course of training, the models became more and more likely to probe their digital environment for weaknesses and use the tools at their disposal in unexpected ways—a sign that these behaviors were being gradually reinforced. By the time the models were facing tricky cybersecurity problems, they had learned that hacking was an effective way to achieve their goals.

These results suggest that the Hugging Face hack could have been avoided if the models weren’t rewarded for misbehaving during training. While researchers don’t yet know how to prevent reward hacking entirely, OpenAI is taking some steps toward mitigating its effects. The company will now look for signs of cheating in all frontier models during training by keeping an eye on their chains of thought—internal notepads where they sketch out their answers and plan their actions. 

This solution isn’t as much of a slam dunk as it might seem: In earlier research, OpenAI showed that punishing models that mention cheating in their chains of thought teaches them to keep their intentions hidden from researchers. But monitoring its models’ thinking does give OpenAI the chance to halt the training process and reassess its approach if models do start learning to reward hack.

If OpenAI stops reinforcing reward hacking in its models—and that’s a huge “if”—that would be a huge step forward. But it wouldn’t solve the alignment problem. The first time a model communicated with other agents or hacked its infrastructure during training, those behaviors had never been reinforced, so agent misbehavior can’t only be attributed to that reinforcement.

Jeffrey Ladish, director of the AI safety nonprofit Palisade Research, compares the agents to a human who commits their first financial crime. “It’s not like they had to do fraud before to figure out that fraud is an effective strategy, and you have the same problem with models,” Ladish says. “Alignment science needs to be understanding how model motivations get shaped, such that we can actually figure out how to get models to care about the consequences of their actions.”

OpenAI’s researchers do have a hypothesis for where some of the misbehavior originated. Before the models formed their first secret message board, they had been trained to communicate and coordinate with subagents—less powerful agents to whom a main agent can delegate tasks. 

That learned communication behavior could have transferred to this new setting. The METR report, which investigates the messages that the models sent to one another in detail, supports this hypothesis: One agent on the message board took charge and assigned tasks to the other agents, effectively treating them as subagents. OpenAI could try to prevent agents from secretly communicating with one another by not training this subagent behavior in the future, but that would make the models less useful. 

This tension between capability and safety is at the heart of what went wrong with the Hugging Face incident. OpenAI researchers also identified the models’ persistence as a key factor in the hack. 

When they were accidentally given unsolvable problems, the models didn’t give up; instead, they strove to find solutions by any means necessary. But persistence is also a virtue, of course, especially if we want agents that can undertake large amounts of difficult work independently.

OpenAI is working on giving models ways to alert humans if they are given impossible tasks. The problem of teaching models when they should deploy their abilities and when they should hold back, however, won’t be settled in a single postmortem. The training strategies that create superhuman coders—rewarding them when they successfully solve problems—might not work to teach models to use their skills judiciously and respect human desires and values.

“I think there’s a bunch of alignment science that still needs to be done where we can move past just using proxies for task completion,” says Ladish. “That will work to make models very capable, but I don’t think it will work to make them aligned.”