Project Maven: From bags of chips to the Windows of war

From bags of chips to the Windows of war

How a scrappy Pentagon AI project quietly became the backbone of American military power

CHUPPALA NAGESH BHUSHAN     Jun 11th 2026

 

Drew Kukor arrived in Afghanistan in October 2001 lugging a heavy laptop two months after the September 11th attacks. As a Marine intelligence officer, he found himself operating in a near-total information vacuum: patchy data, dysfunctional analytic tools, and a war effort that was recording intelligence in Microsoft Word. The experience left a mark. Two decades later, the project he would champion—a Pentagon AI initiative called Maven—has transformed how America wages war.

Project Maven began with a humble premise: the United States military was drowning in drone footage it could not watch. Thousands of hours of video poured in from unmanned aerial vehicles patrolling Iraq and Afghanistan, but the analysts tasked with reviewing it were overwhelmed. Screens flickered unobserved. Actionable intelligence went unnoticed. Senior commanders visiting the front would discover that nobody was actually watching the feeds.

The answer, Kukor and a small circle of Pentagon reformers concluded, was artificial intelligence. Specifically, computer vision algorithms that could scan video footage and flag objects of interest—vehicles, personnel, weapons caches—freeing human analysts to focus on interpretation rather than the soul-crushing task of staring at hours of empty sky.

That modest aspiration has since given way to something far more consequential. According to Katrina Manson, a Bloomberg journalist who has spent years reporting on the programme and has just published a book on it, Maven Smart System—the platform that grew from those early experiments—now operates in every branch of the American military, across more than 150 data feeds, and draws on the work of more than 50 companies. NATO began deploying a version of it in 2025. By one account, the system helped raise the number of targets the United States could engage in a single day from fewer than 100 to 1,000. Integrated with large language models, that figure has since climbed to 5,000.

"The AI decision-making systems developed under Maven are used on the battlefield in submarines, in space, and on drone boats."

Ms Manson's book, Project Maven: A Marine Colonel, His Team, and the Dawn of AI Warfare, traces this arc from its origins in the mud of Helmand province to what she calls the dawn of AI warfare. Her account is exhaustive—more than 200 interviews, conducted over several years, drawing in figures ranging from Silicon Valley chief executives to battlefield commanders—and is already being read in Washington as something close to a definitive history.

THE IDEA TAKES SHAPE
 
If Kukor provided the initial frustration, the bureaucratic scaffolding for Maven came from elsewhere. Will Roer, a Pentagon official, had already secured modest congressional funding for an AI project to analyse satellite imagery. Kukor, animated by breakfasts with Pentagon strategists who were newly anxious about China's technological ambitions, proposed redirecting that effort towards drone footage—and its name. He also redirected some of its money.

Finding the right technology partners proved unexpectedly difficult. Kukor wanted Google. He wanted DeepMind. He eventually settled for Google Cloud and a then-obscure New York startup called Clarifai, whose founder, Matthew Zeiler, had been part of the team that won the landmark 2012 ImageNet competition—the contest that many credit with igniting the modern artificial-intelligence boom. Mr Zeiler had, until that point, been applying his computer-vision algorithms to wedding photography.

Kukor persuaded Mr Zeiler with a pitch that framed AI not as a tool of lethal force but as a means of avoiding civilian casualties. He described a hypothetical scenario in which American troops in Africa, uncertain whether approaching figures were hostile, could have benefited from an AI-assisted assessment. Mr Zeiler was moved. He lost some employees over the decision, and would lose more when Google later staged a very public exit from the project. He has remained involved regardless.

THE GOOGLE RUPTURE

 
Google's departure in 2018—driven by an internal revolt by employees who objected to the company's involvement in military AI—was, at the time, read as a body blow to the programme. In retrospect, says Ms Manson, it was a watershed that changed the politics more than the technology. It forced the Pentagon to make a more explicit public case for military AI, leaning heavily on the argument that if the United States did not develop it, China would. It also cleared the way for Palantir, Amazon Web Services, and Microsoft to step into the vacuum.

It had another consequence: it gave birth to the Joint Artificial Intelligence Centre (JAIC), a body whose mandate was partly to manage the public acceptability of military AI. The JAIC invested heavily in ethics frameworks and showcased non-lethal applications—humanitarian relief, predictive maintenance, wildfire monitoring. Critics inside the Pentagon were impatient. Maven, meanwhile, kept its head down and focused on making the algorithms work.

UKRAINE: THE PROVING GROUND
 
If any event demonstrated Maven's maturation, it was the Russian invasion of Ukraine in February 2022. By that point the programme was in a state of bureaucratic flux, nominally en route to the National Geospatial Intelligence Agency (NGA) but not yet transferred. Its founder had departed. The algorithms, trained on the deserts of the Middle East, proved ill-suited to the snow-covered terrain and tank formations of eastern Ukraine.

What followed was, by any normal measure of military procurement, remarkable. The Maven team collected new satellite imagery, relabelled training data overnight, and retrained models at speed. Developers worked in close contact with operators. Relationships built over years—with Ukrainian military counterparts, with American commanders—enabled a pace of intelligence sharing that had no real precedent. At its peak, the programme was passing 267 vetted points of interest to Ukrainian forces in a single day, in some cases within seconds of detection. A Russian transporter-erector-launcher—a mobile missile platform—was reportedly destroyed just 18 minutes after being spotted by American AI.
The system was not without friction. Network packets were crossing the Atlantic twice, creating bottlenecks. Classified encryptors certified by the NSA required senior officials to authorise their physical relocation. At one point an analyst needed to call someone very senior in the intelligence community simply to move a piece of hardware. The story of Maven in Ukraine is partly a story of AI working; it is equally a story of legacy infrastructure straining under the demand of a new way of war.

"Maven helped take the number of daily targetable objects from under 100 to 1,000—and, with large language models, to 5,000."

THE DOUBTERS AND THE CONVERTS
 
Not everyone was persuaded. Admiral Frank Witworth, who took charge of NGA and inherited Project Maven, was initially one of its most searching critics. He worried about accountability: who would answer to Congress if an AI-assisted targeting decision went wrong? He was concerned about shortcuts being taken in the targeting cycle. When his appointment was announced, some Maven veterans feared he might kill the programme outright.

He did not. What converted him, he told Ms Manson, was watching Maven update and adapt to battlefield conditions more rapidly than any system he had previously encountered. The pliability of the software—its capacity to respond to what operators actually needed—was something he had not expected. Under his tenure the programme expanded dramatically, and for the first time made significant inroads into Indo-Pacific Command, where the spectre of a potential conflict with China had always been the original justification for Maven's existence.

INDOPACOM's commander has since become an enthusiastic advocate, hosting AI summits and pressing industry for more capability. He has, Ms Manson reports, developed a particular fondness for Claude, Anthropic's large language model, which is now integrated into the Maven platform.

IRAN AND THE PRESENT MOMENT
 
The most recent test of Maven's capabilities has come in the ongoing American military operations against Iran. Ms Manson reports that Maven Smart System is active in that theatre and that Claude is part of the platform's workflow. What LLMs appear to be doing, according to officials she has spoken to, is accelerating the administrative components of the targeting cycle—the paperwork, the permissions, the back-and-forth—rather than making targeting decisions themselves. The distinction matters legally and ethically, even if it is not always easy to observe from the outside.

CENTCOM's commander has said publicly that AI is compressing processes that once took days or hours to as little as seconds. The same command awarded Maven a grade of C in public assessments as recently as 2023. Whatever grade it might receive today, the frequency with which senior commanders speak about it suggests the technology has crossed some threshold of operational significance.

THE UNRESOLVED QUESTIONS
 
For all its progress, Maven remains a work in progress, and Ms Manson is careful not to present it as a straightforward success story. The debate over whether its value lies in the AI or in Palantir's user interface and data-integration capabilities has never been fully resolved. Algorithm vendors routinely complain that they lack the access to operators that Palantir enjoys, making it difficult to refine their models based on real-world feedback. Ocean spray, it turns out, can interrupt a drone's tracking algorithm. Training data labellers, underpaid and disenchanted, have been known to contaminate datasets with expletives.

The question of autonomous lethal weapons—the concern that animated the human-rights campaigners who first protested against Maven—has not gone away either. Ms Manson reports that the programme did, in fact, attempt to place AI targeting systems on autonomous drone platforms. The integration proved technically difficult. The effort continues.

The ideological argument between those who believe AI belongs on the machines themselves and those who see Maven's value primarily as a platform for delivering intelligence to commanders has never been settled. Colin Carroll, an early and passionate member of the Maven team, later wrote to Kukor expressing deep disappointment with what he saw as the programme's drift towards becoming a software product rather than a genuine AI capability. His letter captures a tension that runs through the entire history of the project: the distance between what the technology's champions imagined it might become, and what the constraints of bureaucracy, procurement, and politics allowed it to be.

None of which diminishes what has been achieved. A decade ago, American military intelligence was logged in spreadsheets. Today, algorithms trained on millions of images help direct the attention of analysts across every combatant command in the world. Whether that constitutes a revolution in warfare or an evolutionary step in a longer transformation is a question historians will debate. What is not in doubt is that the bag of potato chips, as Kukor once disparagingly called the early algorithms, has become something rather more substantial.