Palantir's Maven Goes Permanent: The AI That Sees Everything

On March 22, 2026, Military.com reported that the Pentagon is formally expanding Project Maven from its current experimental acquisition structure into a permanent, long-term defense program. The announcement was not accompanied by fanfare. There was no press conference, no congressional hearing, no public statement from the Secretary of Defense. The designation arrived as a line item in acquisition documentation — the bureaucratic equivalent of a quiet revolution. Project Maven, the Defense Department's flagship artificial intelligence intelligence-analysis system, is no longer a project. It is infrastructure.

The implications of that transition are more significant than the quiet announcement suggests. Maven's conversion to permanent status means that the AI system now processing and interpreting millions of hours of drone footage, satellite imagery, and signals intelligence every month has moved from the category of experimental capability to the category of essential military function. It means that the $800 million to $1 billion in annual contract value flowing to Palantir Technologies for Maven's operation and expansion is now a baseline defense expenditure rather than a discretionary program subject to annual review. It means that the AI that sees everything the American military watches from the sky has become as permanent a fixture of the national security apparatus as the NSA or DIA.

To understand what that means, you need to understand where Maven came from, what it actually does, why every major technology company except Palantir eventually walked away from it, and what it signals about the trajectory of AI in warfare.

The Origin Story: How Maven Was Born and Google Was Burned

Project Maven began in April 2017 when Robert Work, then Deputy Secretary of Defense, signed a memo establishing the Algorithmic Warfare Cross-Functional Team. The team's mandate was specific and urgent: apply artificial intelligence to the analysis of full-motion video from drone surveillance feeds. The problem it was solving was not classified. The American military, by 2017, was generating more surveillance video than its human analysts could process. Predator and Reaper drones operating over Iraq, Syria, Afghanistan, and a dozen other theaters were streaming hundreds of thousands of hours of footage every month. Analysts could watch only a fraction of it. Targets were being missed. Threats were being missed. The AI was supposed to solve the attention deficit.

The initial contract went to Google. The company's TensorFlow machine learning framework was state of the art, its engineering talent was unmatched, and its leadership saw the Maven contract as an opportunity to demonstrate that commercial AI capabilities could serve defense applications. The contract, worth approximately $15 million in its initial tranche, funded the deployment of Google's AI to help Pentagon analysts automatically detect and identify objects of interest in drone video — vehicles, people, structures, weapons — flagging them for human review.

What Google's leadership did not anticipate was the reaction of its employees. In early 2018, a group of Google engineers published an open letter demanding that the company cancel the Maven contract and commit to never building AI for military surveillance or weapons applications. The letter ultimately gathered approximately 4,000 signatures from Google employees. The internal pressure was sustained and intense, including resignations from senior technical staff who had been recruited specifically because of Google's stated commitment to ethical AI development. In June 2018, Google announced it would not renew the Maven contract when it expired and published a set of AI principles that explicitly excluded applications in which AI would be used to violate human rights or cause harm without appropriate oversight.

Google's departure created a vacuum. The Defense Department's need for drone video analysis AI had not diminished — if anything, the surveillance data volumes were growing faster than ever. The Pentagon needed a vendor that would not be destabilized by internal employee activism, that had the organizational culture and mission alignment to operate in classified environments, and that had the technical capability to develop and operate surveillance AI at the scale the military required. The answer was Palantir Technologies.

The Maven Timeline

What Maven Actually Does: The Technical Architecture of Mass Surveillance AI

Maven's public description — AI-powered analysis of drone footage — is accurate but dramatically understates the scope of what the system does in its current form. The Maven Smart System, as it is now designated, is a multi-intelligence fusion platform that aggregates and analyzes data from several distinct intelligence collection disciplines simultaneously.

Full-Motion Video Analysis

The original Maven capability remains the core of the system. Maven processes video streams from a range of unmanned aerial vehicles, including Predator, Reaper, Global Hawk, and a growing array of smaller tactical drones operated by Army and SOCOM units. The AI performs several functions on this video in real time: object detection and classification, identifying vehicles, persons, structures, and equipment; movement tracking, following identified objects across multiple frames and video streams; pattern-of-life analysis, building behavioral profiles of individuals and groups based on their movements, associations, and activities over time; and anomaly detection, flagging behavior that departs from established patterns as potentially significant.

The scale of Maven's video processing capability is publicly acknowledged only in rough terms. DoD officials have testified that the system processes "millions of hours" of surveillance footage monthly. Independent analysis of publicly disclosed contract values and computing infrastructure suggests that Maven's video analysis alone processes on the order of 5 to 10 million hours of footage per month across all theater deployments — a volume that no human analytical workforce could meaningfully review at any level of detail.

Satellite Imagery Analysis

Maven's expansion beyond drone video into satellite imagery analysis represents a qualitative shift in the system's strategic significance. Satellite imagery provides coverage of areas where drone surveillance is not feasible — deep inside adversary territory, over the open ocean, across regions where overflight rights are contested. Maven's satellite imagery analysis capability, developed primarily between 2021 and 2024, allows the system to monitor fixed installations, track vehicle movements, detect construction activity, and identify military equipment at locations around the globe.

The combination of drone video and satellite imagery gives Maven a temporal coverage capability that neither source provides alone. Satellite passes over a given location occur at most a few times per day. Drone surveillance can be continuous for days at a time, but drones cannot be everywhere simultaneously. Maven's AI fuses data from both sources, maintaining a continuously updated picture of monitored locations that is more complete than either input alone.

Signals Intelligence Fusion

The most sensitive aspect of Maven's expanded capability is its integration with signals intelligence — the interception and analysis of electronic communications, radar emissions, and other electronic signals. The specific nature of Maven's SIGINT integration is classified, but open-source reporting based on congressional testimony and industry sources indicates that the system can correlate signals data with imagery to identify individuals and equipment, track communications networks associated with identified targets, and flag signals activity as a cueing mechanism for imagery and video collection.

The fusion of full-motion video, satellite imagery, and signals intelligence into a single analytical platform is what distinguishes the current Maven Smart System from its 2017 predecessor. The original Maven was a single-INT video analysis tool. The current Maven is a multi-INT fusion engine that creates a more complete operational picture than any single intelligence discipline could provide.

Predictive Analytics and Targeting Recommendations

The most consequential capability added to Maven since Palantir took over the contract is predictive analytics. Maven does not merely describe what it observes. It produces predictions about what observed actors are likely to do, recommendations about which targets merit priority attention, and in its most advanced configurations, targeting recommendations that identify specific individuals or locations as high-probability threats based on pattern-of-life analysis and behavioral modeling.

The targeting recommendation capability is the aspect of Maven that human rights organizations and congressional oversight staff find most alarming, and it is the aspect that is most carefully described in Pentagon communications as "decision support" rather than "autonomous targeting." The distinction matters legally: if Maven recommends a target and a human analyst approves the recommendation and a human operator conducts the strike, the legal chain of accountability for the strike runs through the human decision-makers, not the AI system. Whether that legal architecture accurately reflects where the substantive decision is actually being made is a question that the permanent designation of Maven as long-term infrastructure has made more urgent.

Palantir's Defense Empire: The Full Architecture

Maven is Palantir's most publicly discussed defense program, but it is one piece of a much larger portfolio of military AI systems that the company has assembled since its founding in 2003 by Peter Thiel, Alex Karp, and Joe Lonsdale with seed funding from the CIA's In-Q-Tel venture arm. Understanding Maven requires understanding the ecosystem it operates within.

Gotham: The Intelligence Backbone

Palantir Gotham is the intelligence community's primary data integration and analysis platform, used by CIA, FBI, NSA, DIA, and dozens of other agencies to aggregate, search, and analyze vast quantities of structured and unstructured intelligence data. Gotham was Palantir's original product and remains the foundation of its government business. Maven's multi-INT fusion capability is built on Gotham's data integration architecture, which is why Palantir was the natural successor when Google departed — the company already had the classified infrastructure and security accreditations that the Maven program required.

MetaConstellation: Space-Based Intelligence

MetaConstellation is Palantir's platform for integrating and analyzing data from commercial satellite constellations. As the commercial satellite imagery market has expanded dramatically, with companies like Planet Labs, Maxar, BlackSky, and dozens of smaller operators now providing daily global coverage at resolutions that were classified capabilities five years ago, the challenge has shifted from collecting satellite imagery to analyzing the petabytes of imagery that commercial constellations now generate. MetaConstellation is Palantir's answer to that challenge, applying AI analysis at scale to commercial satellite data streams and integrating the results with classified intelligence products.

AIP for Defense: The LLM Layer

Palantir's Artificial Intelligence Platform for Defense, launched in 2023, brings large language model capabilities to military applications. AIP allows analysts and commanders to interact with military intelligence systems using natural language queries — asking Maven what it observed in a given location over a given time period, requesting pattern-of-life summaries for identified individuals, or querying logistics systems for equipment availability. The natural language interface dramatically lowers the technical barrier to accessing AI-generated intelligence, making the system accessible to operators who lack the technical training to interact with traditional database query tools.

AIP's integration with Maven represents the system's most significant capability expansion since Palantir took over the contract. An analyst can now ask Maven, in plain English, to summarize the recent activity patterns of a specific vehicle, identify any deviations from that pattern, and flag locations where the vehicle has been observed in association with other vehicles of interest. The AI processes the query across millions of hours of archived video and returns a structured summary with supporting visual evidence. The process that might have taken a team of analysts weeks now takes minutes.

The Market Position: How Palantir Won the Wars Its Competitors Refused to Fight

The story of who did not win the Maven contract is as revealing as the story of who did. Google's departure in 2018 was followed by the exits of Microsoft, Amazon, and IBM from various aspects of military AI contracting, each driven by some combination of employee pressure, ethical commitments, and strategic decisions about reputational risk.

Microsoft's Retreat

Microsoft won the Joint Enterprise Defense Infrastructure cloud contract in 2019, a $10 billion award that would have made it the primary cloud computing provider for the entire Department of Defense. The contract was challenged by Amazon, lost in court, relitigated, and ultimately cancelled in 2021 in favor of a multi-vendor approach called the Joint Warfighting Cloud Capability. Throughout the JEDI controversy, Microsoft faced sustained internal criticism from employees who opposed the company's military AI work, particularly after the company defended its augmented reality headset contract with the Army. While Microsoft remained engaged with military cloud computing, it retreated from the most visible AI targeting and intelligence applications that Maven represented.

Amazon's Ambivalence

Amazon Web Services remains the largest cloud computing provider for classified government workloads through its GovCloud and Top Secret Region infrastructure, but AWS has deliberately avoided the AI applications layer of military intelligence — the part of the stack that Maven occupies. Amazon's calculation appears to have been that the cloud infrastructure layer could be provided without the reputational exposure that comes from explicitly developing AI systems that recommend or enable targeting of human beings. The strategy has allowed Amazon to capture significant defense cloud revenue while maintaining a degree of distance from the most ethically contested applications.

IBM's Withdrawal

IBM, once one of the primary technology providers to the U.S. intelligence community, has progressively reduced its presence in military AI applications since 2020. The company's Watson AI platform, which had been positioned for intelligence community applications in the mid-2010s, failed to achieve the performance levels that classified customers required and was eventually discontinued. IBM's departure from competitive military AI left the market to specialized players — primarily Palantir — who had built their businesses around defense applications from the beginning.

The pattern across all three companies is consistent: they engaged with military AI when it appeared to be primarily a cloud computing and data infrastructure play, and retreated when the applications became explicitly about surveillance, targeting, and lethal force support. Palantir never had that ambivalence. The company was founded to serve intelligence agencies. Its founders understood that the ethical questions around surveillance AI were not bugs to be resolved before entering the market but defining features of the market they had chosen to operate in.

Palantir's Financials: What Permanent Warfare Pays

Palantir Technologies trades on the New York Stock Exchange under the ticker PLTR. As of March 2026, the company's market capitalization exceeds $160 billion, up from approximately $32 billion in early 2022 — a gain of more than 400 percent in four years. The primary driver of that appreciation is the market's reassessment of Palantir's long-term revenue trajectory as the depth and permanence of its defense relationships became clear.

The revenue composition tells the strategic story. In fiscal year 2022, Palantir's commercial revenue and government revenue were roughly balanced. By fiscal year 2025, defense and government revenue had grown to represent more than 55 percent of total revenue, driven by Maven expansion, new AIP for Defense contracts across military branches, and international government deployments in allied nations including the United Kingdom and Ukraine. The company that Peter Thiel founded as a data analytics platform has become, functionally, a defense technology company with a commercial analytics business attached.

Maven alone represents an estimated $800 million to $1 billion in annual contract value across its various program lines, a figure that includes the core drone video analysis capability, the satellite imagery integration, the AIP interface layer, and the classified capability expansions that are acknowledged only in aggregate congressional budget data. The permanent designation of Maven as a long-term program effectively converts that revenue from a contract that requires annual renewal and competition to a sustained baseline that will grow as Maven's capabilities and theater deployments expand.

"We did not want to be the company that builds the software that identifies targets. We are the company that builds the software that identifies targets. That is what we do. We are comfortable with it."

-- Alex Karp, Palantir CEO, speaking to shareholders, 2024

The Oversight Gap: Who Watches the AI That Watches Everything

The permanent designation of Maven raises accountability questions that the Pentagon's announcement did not address and that Congress has not yet fully resolved. When an AI system is an experiment, its governance framework can be provisional. When that system is infrastructure — when it is permanently embedded in the intelligence collection and targeting workflow of the most powerful military on earth — the governance framework must be something more durable.

The current oversight architecture for Maven has three layers. The first layer is internal to Palantir: the company applies its own ethical review processes to capability development, maintains audit logs of system outputs, and claims to have designed the system with human-in-the-loop requirements for lethal targeting applications. The second layer is Department of Defense policy: the DoD's AI ethics principles, adopted in 2020 and updated in 2023, require that AI systems used in lethal force applications maintain meaningful human control and be subject to regular review for unintended consequences. The third layer is congressional oversight: the House and Senate Armed Services Committees receive classified briefings on Maven's capabilities and deployments, and relevant subcommittees have conducted hearings on AI in military targeting.

The gap in that architecture is accountability for specific decisions. When a Maven-generated targeting recommendation leads to a strike that kills civilians, the after-action investigation proceeds through the normal military justice and inspector general channels. But those channels were designed for decisions made by human commanders, and they are poorly suited to auditing AI-assisted decisions where the causal chain from data to recommendation to decision to action involves multiple systems, multiple human reviewers, and latency measured in seconds rather than hours.

Congressional testimony by senior DoD officials on Maven's oversight framework has been consistent in its assurances and conspicuously vague in its specifics. Officials confirm that humans remain in the decision loop for all lethal applications. They describe the oversight processes at a level of abstraction that does not allow committee members to assess whether those processes would actually detect and correct systematic errors in Maven's targeting recommendations. The permanent designation of Maven as a long-term program makes the absence of a robust, specific, publicly disclosed oversight framework more problematic, not less.

Maven vs. Lavender: Two Models of AI Targeting

The most instructive comparison for Maven is not American predecessor programs but Israeli ones. The Israeli Defense Forces' Lavender system, described in detail by reporting from +972 Magazine and The Guardian in 2024, is an AI targeting system that generates ranked lists of suspected Hamas and Palestinian Islamic Jihad operatives as targets for potential strikes. Lavender's methodology involves pattern-of-life analysis similar to Maven's, correlating phone data, social media activity, residential associations, and military affiliation signals to generate target recommendations with associated confidence scores.

The surface similarities between Maven and Lavender are real: both use AI to analyze behavioral patterns, both generate target recommendations, and both operate with some form of human approval requirement before lethal action. But the governance contexts differ substantially. Lavender, as described by IDF sources who spoke to journalists, operated during the Gaza conflict with human review times measured in seconds, with targets approved in batches rather than individually, and with a stated policy permitting strikes that would kill multiple civilians per confirmed militant target. The human oversight requirement was formally present but effectively minimal.

Maven's oversight context is different: the American military's targeting doctrine requires more extensive collateral damage assessment, longer review chains, and more explicit human authorization for lethal action. Whether those differences hold in practice under the time pressure of active combat operations, and whether they will be maintained as Maven becomes permanent infrastructure whose outputs are increasingly treated as authoritative, are questions that cannot be answered from publicly available information. The comparison with Lavender is not an accusation that Maven operates the same way. It is a warning that AI targeting systems, once embedded as infrastructure, tend to generate institutional pressure toward faster, more automated decision chains — regardless of the initial design intentions.

The Permanent Warfare State: AI as Infrastructure, Not Project

The most significant aspect of Maven's conversion to permanent program status is not the contract value or the technical capability expansion. It is the institutional signal. When the Pentagon converts a capability from a project to infrastructure, it is declaring that the capability is no longer optional. Infrastructure does not get cancelled. Infrastructure gets expanded, maintained, and improved. Infrastructure becomes the basis for other capabilities that depend on its continued operation. Infrastructure becomes, over time, so deeply embedded in organizational processes that removing it would be operationally catastrophic.

Maven is now infrastructure. The intelligence workflows of multiple combatant commands — CENTCOM, INDOPACOM, AFRICOM, EUCOM — depend on Maven's continuous operation. The targeting processes of forces operating in active theaters incorporate Maven's outputs as a standard input to the analytical process. The institutional knowledge of how to conduct targeting operations without Maven — through manual analysis of drone video, through traditional imagery exploitation, through the laborious human pattern-of-life analysis that preceded AI — is atrophying in the organizations that depend on Maven's capabilities.

This is the dynamic that the phrase "always on" captures. Maven was designed to be always on — continuous surveillance, continuous analysis, continuous pattern-of-life tracking. That continuous operation was initially conceived as a capability advantage. It is increasingly also an organizational dependency. The American military's intelligence and targeting apparatus now relies on Maven's continuous output in a way that was not true five years ago and that will be more true five years from now.

What happens when the AI that sees everything is the system that everything depends on? What happens when Maven has a systematic error in its object classification — misidentifying civilian vehicles as military vehicles in a specific environmental context — and that error propagates through hundreds of analytical products before it is detected? What happens when a sophisticated adversary learns enough about Maven's training data and algorithmic structure to deliberately create patterns of behavior designed to fool Maven's classification systems into identifying legitimate targets as non-threats or non-combatants as high-priority targets?

These are not theoretical questions. They are the predictable failure modes of any large-scale AI system operating at the complexity and scale of Maven, and the permanent designation of Maven as infrastructure means that the organizations that depend on it will have less tolerance for the system's downtime, less flexibility to operate without it, and less institutional capacity to detect and correct its errors over time. The transition from project to infrastructure is, in this context, as much a risk as it is a capability milestone.

What Maven's Expansion Means for the Future

The Pentagon's Maven expansion is not an endpoint. It is a waypoint on a trajectory that leads to something considerably more comprehensive: a continuously operating, globally deployed AI intelligence fusion system that monitors adversary activities, tracks threats, and supports military decisions across every domain of conflict — air, land, sea, space, and cyber — in every theater simultaneously.

The expansion plan disclosed in the permanent designation documentation includes three major development tracks. The first is multi-domain extension: taking the video and imagery analysis capabilities that work in air surveillance and applying them to undersea acoustic data, space domain awareness, and cyber intrusion detection. The second is allied integration: expanding Maven's architecture to allow interoperability with intelligence systems operated by Five Eyes partners — the United Kingdom, Australia, Canada, and New Zealand — and potentially with other close allies. The third is predictive capacity expansion: developing Maven's ability to predict adversary intentions and actions rather than merely describing current observations, moving from pattern-of-life analysis to strategic warning.

The combination of those three tracks, if realized over the next five to seven years, would produce a system that bears only superficial resemblance to the drone video analysis tool that Google was contracted to build in 2017. It would be a persistent global intelligence fusion platform capable of monitoring adversary military activities at strategic scale, integrating seamlessly with allied intelligence systems, and generating actionable predictions about adversary intent rather than merely describing observed behavior. It would be, effectively, the automated intelligence community that DoD planners have discussed in classified forums for years — the AI that watches the watchers, understands what they are doing, and advises on how to respond before they act.

That future raises questions that the current oversight architecture is not equipped to answer. Who approves the parameters of AI systems that make strategic-level predictions about adversary intent? What recourse exists when those predictions are wrong and military decisions based on them prove catastrophic? How does the United States maintain the escalation control and diplomatic signaling that international security requires when the AI systems mediating its adversary awareness are operating continuously, generating outputs at machine speed, and embedded in decision processes that move faster than human deliberation can accommodate?

Project Maven is no longer a project. The questions it raises are permanent too.