The Pentagon Made Palantir Maven Its Permanent War OS

Deputy Defense Secretary Steve Feinberg ordered Palantir’s Maven Smart System onto the Pentagon’s lasting books in a March 9 memo. The four-page directive told senior leaders to make the targeting software a program of record by the close of fiscal 2026 on September 30, with a dedicated funding line and a seat in the five-year defense plan.

That same winter, Defense Secretary Pete Hegseth labeled Anthropic a supply-chain risk and moved to pull Claude off classified networks. In Iran, Maven helped the force hit 13,000 targets in 38 days. It also sat in the kill chain that struck a school in Minab.

Feinberg’s Memo Puts Maven on the Five-Year Budget

Feinberg sent the memo to senior Pentagon leaders and military commanders. He wrote that putting Maven on lasting books would give the force “the latest tools necessary to detect, deter, and dominate our adversaries in all domains.”

Transitioning MSS to a POR will provide the stable funding and resourcing necessary for deliberate, sustained development and enterprise-wide integration to give our commanders the C2 necessary to fight and win wars

Steve Feinberg, Deputy Secretary of Defense, March 9 memo

He also wrote that the move would mature “a promising initiative into an enduring, joint, and fully supported capability.” Oversight of system administration was to shift from the National Geospatial-Intelligence Agency to a Chief Digital and Artificial Intelligence Office Maven program office within 30 days. Research and Engineering was told to take authorizing-official duties for Maven and its commercial cloud. Future buys were to ride the Army’s existing enterprise agreement with Palantir.

Virginia Burger, a senior defense policy analyst at the Project on Government Oversight and a former Marine officer, said program-of-record status puts Palantir’s software “in as permanent a position as a program can get within the DOD.” Emelia Probasco, a senior fellow at Georgetown’s Center for Security and Emerging Technology, framed the paperwork as a signal. Services, she said, do not fully fund training, maintenance, and the rest of the force-design checklist until they see that stamp.

As of September 22, Pentagon officials still described the March memo as a deadline for the end of the fiscal year. Fiscal 2027 budget materials already treat the shift as the reason for a large joint-command request. In August, Army Col. Molly Solsbury was named Maven program director inside CDAO.

Project Maven Began as Computer Vision in 2017

The screen now being treated as joint command software started as an intelligence experiment. Pentagon leadership launched Project Maven in 2017 to detect, tag, and track objects in aircraft, drone, and satellite footage. Google was an early partner and left in 2018 after staff protests. Palantir stepped in as the prime software integrator and has stayed there.

THE MAVEN PATH TO A LASTING PROGRAM

1. 2017: Project Maven starts under the undersecretary of defense for intelligence and security as the Algorithmic Warfare Cross-Functional Team.
2. 2018: Google declines to renew. Palantir begins building the interface that becomes Maven Smart System.
3. 2022: Responsibilities split between NGA, which handled geospatial models, and CDAO, which paid licenses and ran text functions.
4. May 2024: CDAO awards Palantir a $480 million, five-year production contract, first covering five combatant commands.
5. May 2025: The department raises that ceiling to nearly $1.3 billion through 2029.
6. March 9, 2026: Feinberg orders a full shift of Maven oversight to CDAO and lasting-program status by September 30.

Matt Mande and Gregory C. Allen, writing for the Center for Strategic and International Studies, separate the names. “Maven” is the wider Pentagon program, including automated target recognition. “MSS” is the flagship software layer. Palantir documentation they reviewed says “MSS is powered by the Palantir Platform,” and how Maven Smart System actually works is now public in unusual detail, including live demos.

Drew Cukor, the retired Marine colonel who was Maven’s founding director, said that before these tools, staff spent “like 3 percent of our time actually analyzing it because 97 percent of the time was just to get the damn data together.” CENTCOM’s Maven instance drew from 179 data sources in 2024, a figure industry sources told CSIS has grown since.

What a Program of Record Changes for Palantir

A program of record is not a new algorithm. It is the Pentagon’s way of saying a tool is no longer a prototype that must justify itself every budget cycle. Feinberg told undersecretaries that if any transition task would “negatively impact delivery of MSS capability to the warfighter,” they should adjust his timelines rather than stall the fielded system.

WHAT THE MARCH DIRECTIVE LOCKS IN

• A budget line: Maven gets dedicated funding and a place in the Future Years Defense Program, the five-year map of Pentagon resources.
• A new owner: Administration moves from NGA to a CDAO Maven program office, with Research and Engineering as authorizing official.
• A buying path: All Maven contracts shift onto the Army enterprise agreement that already covers Palantir’s joint work.
• A force-design trigger: Doctrine, training, people, and facilities work that units delay for pilots is supposed to start once the stamp is visible.
• A possible next home: Feinberg ordered a CJADC2 acceleration plan that could later park Maven under a joint command-and-control program office.

A former senior defense official, speaking after the memo, said the test is whether Palantir turns Maven from a shared picture into “the platform of platforms.” That, the official said, “would lock them in for a decade.” Probasco called the questions around that shift “gnarly” and said Congress would want more oversight as the tool moves from intelligence fusion into command.

Palantir’s unpublished baseline list, summarized by CSIS and spelled out on Palantir’s own Maven capability sheet, includes battlespace management, target management, computer-vision detections, and generative AI. Large language models, an anonymous NGA official told author Katrina Manson, added about a fivefold jump in targeting speed on top of the gain from computer vision. Palantir’s AIP layer is the pipe those models sit in. The models can change. The screen they feed is what the memo makes hard to unwind.

Users Doubled to 100,000 During Epic Fury

James Mazol, deputy undersecretary of defense for research and engineering, put a headcount on the Iran campaign at DefenseTalks on September 22. “In January of this year, we had about 50,000 people using Maven. Then [Operation] Epic Fury kicks off, and now we’re up to over 100,000.” He said every combatant command plus the National Guard Bureau was in the queue. Cameron Stanley, the Pentagon’s chief digital and AI officer, tied the same weeks to strike volume.

That’s data-centric warfare. That’s using AI at scale

Cameron Stanley, Chief Digital and AI Officer, DefenseTalks, September 22, 2026

Stanley said Maven helped the United States strike 13,000 targets in 38 days during Epic Fury. CSIS, citing earlier campaign reporting, said the first 24 hours of the Iran war used Maven to help hit more than 1,000 targets, a tenfold rise over the pre-Maven era. In a March presentation, Stanley walked through a targeting loop on one screen: an AI detection, a ranked list of nearby weapons, an order, then battle-damage video.

MAVEN AT COMBAT SCALE, 2026

• January users: About 50,000 Defense Department people on Maven, per Mazol.
• September 22 users: Over 100,000, more than double, after Epic Fury.
• First-day strikes: More than 1,000 targets in the first 24 hours of the Iran war, per CSIS.
• Campaign strikes: 13,000 targets in 38 days, per Stanley.

He had already described the old workflow as eight or nine systems, with people “literally moving detections left and right.” A CSET assessment of the 18th Airborne Corps’ Scarlet Dragon exercises found the unit could match Operation Iraqi Freedom’s time-critical targeting cell with 20 people matching a 2,000-person cell. That comparison is an exercise result, not an Iran-war audit. Stanley later said Maven had already replaced “six, eight, ten” analysis tools and that the department wanted the same pane for logistics, readiness, and supply data, not only intel.

Mazol’s other warning was compute. Classified models cannot train in ordinary commercial halls, he said, which is why the department is pushing an “AI Arsenal” of government-owned, contractor-run data centers and containerized compute at the edge. Speed at the terminal still depends on iron the services do not yet own.

The Minab School Strike Maven Did Not Catch

On February 28, the opening day of the Iran war, two Tomahawk missiles hit Shajarah Tayyebeh Elementary School in Minab. Officials involved in an unreleased internal Pentagon review said the blasts killed more than 150 people, including at least 123 children. Outdated intelligence had listed the site as an Islamic Revolutionary Guard Corps facility. Officials said some Central Command staff leaned too hard on the AI inside Maven, which fused more than 150 inputs and helped compress target-list work from hours toward minutes.

WHAT WE KNOW

• The strike: Two Tomahawks hit the Minab school on February 28, 2026, killing more than 150 people, including at least 123 children, per officials familiar with the review.
• The data error: The site had been cataloged as an IRGC facility on stale records; the physical use of the building had changed.
• The software role: Officials said Maven surfaced the site among recommended day-one targets and that some staff expected the AI to flag stale or inconsistent intel.
• The vendor line: Palantir said it is not responsible for underlying data or for spotting intelligence gaps, and that there is no evidence its software was at fault.

WHAT IS UNCONFIRMED

• The full report: The Pentagon has not released the investigation, and the department declined to comment on the account of it.
• Exact model blame: Officials described overreliance on Maven’s AI as one failure in a chain that also included compressed planning and thin civilian-harm staffing; they have not published a finding that the software chose the school on its own.

Stanley later said Central Command added open-source feeds and that Palantir shipped dozens of Maven upgrades after the strike, including agents that re-review intel for disqualifying facts humans missed. Palantir personnel familiar with the contracts described those agents as a post-Minab build. The company’s line is that Maven does not fire. Humans still approve targets. The review, as officials described it, is about what happens when that approval is rushed through a screen that makes a bad record look complete.

United Nations bodies have spent years arguing that humans must keep judgment over force. In December 2025 the General Assembly adopted a UN vote on lethal autonomous weapons that again flagged humanitarian, legal, and ethical risk. Diplomats later stripped human-review language from a separate draft treaty in Geneva. Palantir’s public claim is that its software does not independently make lethal decisions. Minab is the case that tests what “human in the loop” means when the loop is measured in minutes.

Claude Left Classified Maven Months Ago

Anthropic was the only frontier lab that had, through Palantir, put models on classified Pentagon systems. That status ended in a fight over use limits. Hegseth told Anthropic it had to allow “all lawful uses.” The company refused to drop bans on fully autonomous lethal weapons and on mass surveillance of people in the United States. On February 27 he designated the firm a supply-chain risk, a label usually aimed at vendors tied to rival states. The designation was announced March 3. Feinberg’s Maven memo followed on March 9.

An Anthropic executive had questioned Palantir’s use of Claude during the January 3 operation that captured Venezuelan President Nicolás Maduro. Under Secretary of Defense Emil Michael later said that objection “led to alarm” about whether the software might stop working in a fight. On September 25 a divided D.C. Circuit panel left one of the designations in place. Judge Gregory Katsas wrote that the department had “ample support” for treating continued integration of Claude, by the department or its contractors, as a covered national-security risk. The risk, as the Pentagon framed it, was a vendor’s power to withhold lawful uses, not a foreign back door.

I’d say about 90% has transitioned. All of the Maven Smart Systems or Palantir work has been transitioned months ago, and we’re on track to get it done by the end of the month

Emil Michael, Under Secretary of Defense for Research and Engineering, September 2026

Michael also said relying on one model had been a mistake. Warfighters on Maven, he said, preferred newer ChatGPT builds to Anthropic and found Grok “way faster than both of them.” In March, Palantir chief executive Alex Karp said the company was still using Claude and would add other models. By September the department was describing OpenAI, Google, and xAI as the classified replacements, with Maven already off Claude. The interface did not move. The brain behind the chat box did.

That split is the part the March memo does not advertise. Program-of-record status freezes Palantir’s screen, the workflows, and the Army buying path. The language models that sit under generative features are treated as parts. People who watched the Iran campaign and concluded that “Claude ran the war” were looking at a component. Maven was the system. Claude was one model inside it, and the Pentagon spent the spring and summer proving it could be swapped.

CDAO Now Runs Maven as a Joint Program

The Army is the cashier. On July 31, 2025, it awarded Palantir an enterprise agreement that folds 75 contracts, 15 of them primes and 60 related, into one vehicle with a $10 billion cap over 10 years. Army public affairs was explicit: the figure is a ceiling, not a promise to spend. Other Defense agencies can buy Palantir commercial products off the same vehicle. Feinberg’s memo sends Maven orders down that pipe. Leo Garciga, the Army’s chief information officer, said the point was faster delivery and volume discounts, not a new product.

MAVEN MONEY ON THE BOOKS

Fiscal 2027 budget language says the jump “reflects transition of Joint C2 from rapidly deployed product to program of record.” Officials wrote that Maven “aggregates thousands of discrete data streams into AI-enabled workflows.” A smaller $60 million line funds a Virtual Joint Operations Center for low-bandwidth users. Solsbury’s CDAO office is the near-term home until Feinberg’s wider command-and-control plan is approved. NATO allies already hold Maven licenses. The joint force is no longer sampling the tool. It is writing it into the way headquarters are supposed to see, decide, and shoot.

Stanley, after citing 13,000 strikes in 38 days, asked whether logistics could support that pace for 90, 120, or 180 days. The software problem the memo solves is permanence for the screen. The problems it does not solve are stale intel, thin human review, a model vendor the department was willing to brand a supply-chain risk, and magazines that empty faster than the map refreshes.