The Pentagon is no longer treating AI targeting as a side experiment. It is moving the technology into doctrine, classified networks and long-term procurement, and you should read the spending as seriously as the policy language.
For years, the Pentagon could talk about AI targeting as a set of pilots and experiments, useful but still sitting at the edge of official doctrine. That line is getting harder to sustain. Bloomberg reported that the Defense Department has revised its classified targeting doctrine to give artificial intelligence a larger role in combat decisions, formalizing a shift that has already been visible in contracts, exercises and battlefield systems. The department is not saying machines can make lethal decisions on their own. But the direction is clear enough. Palantir is the company to watch first. In a March 9, 2026 letter, Deputy Defense Secretary Steve Feinberg said Project Maven should become an official program of record by September 2026, the close of the fiscal year. That is not a paperwork detail. Program-of-record status means Maven is being moved toward a durable budget line rather than a string of isolated awards. The Pentagon had already raised the Maven Smart System contract ceiling to nearly $1.3 billion through 2029, up from the original $480 million indefinite-delivery, indefinite-quantity contract signed in 2024.
Maven's job is blunt: pull data from military systems and intelligence feeds into a single interface so commanders and analysts can find, sort and act on targets faster. At Palantir's AIPCon 9, Business Insider reported, Pentagon chief digital and AI officer Cameron Stanley showed how the system could move from identifying a vehicle to selecting a weapon for engagement. You don't need to call that autonomy to see the problem. A system built to compress the kill chain changes the work of the human who remains inside it.
Anduril and Shield AI are running into the same procurement current from different angles. Investor's Business Daily reported in March that the Army awarded Anduril an enterprise contract worth up to $20 billion, consolidating more than 120 separate procurement deals into one framework. Shield AI, meanwhile, has pushed its Hivemind autonomy software into Air Force testing, including a February flight on Anduril's Fury drone, also known as YFQ-44A, over the Mojave Desert. These are not research lab curiosities. They are being wired into programs the services expect to buy, test and deploy.
The drone story makes the procurement logic easy to understand. The Low-cost Uncrewed Combat Attack System, or LUCAS, is a one-way attack drone modeled on Iran's Shahed design and reported at roughly $35,000 per unit. The Pentagon confirmed its first combat use in Iran strikes in February 2026, according to Military Times and other defense reporting. Cheap, attritable systems get more useful when software can help route them, coordinate them and connect them to a wider targeting picture. That is why the AI doctrine question cannot be separated from the hardware budget.
Defense Secretary Pete Hegseth has said publicly that AI does not make lethal targeting decisions, and that remains the formal position. DoD Directive 3000.09 requires commanders and operators to exercise appropriate human judgment before force is applied. The latest defense policy language also requires congressional notification when waivers are issued under that directive. The Brennan Center has treated that disclosure requirement as a real guardrail, but a limited one. Disclosure is not a ban. It tells Congress what happened after the machinery has already been allowed to move.
The classified-network fight shows how quickly the machinery is moving. The Verge reported in May that the Defense Department signed deals with eight companies, Google, Microsoft, Amazon Web Services, Nvidia, OpenAI, Reflection, Oracle and SpaceX, to deploy AI tools on classified networks for lawful operational use. Anthropic refused to drop restrictions on the use of Claude for mass domestic surveillance and fully autonomous weapons, a fight Axios and Reuters have both traced to the Pentagon's demand for broader lawful-use language. OpenAI reached its own agreement with the department, while Sam Altman said OpenAI's principles still prohibit domestic mass surveillance and require human responsibility for the use of force.
Frankly, that vendor split is the story in miniature. The Pentagon wants speed, scale and fewer contractual limits. The AI companies want the defense market, but not all of them are willing to sign the same language. If you're a founder watching this market, the lesson is not that defense tech is suddenly open to everyone with a model and a deck. It is that the government is deciding which safeguards it will accept inside classified systems, and those decisions will shape who gets paid.
The money says the doctrine is catching up to procurement, not the other way around. Palantir's Maven ceiling is already in the billions. Anduril's Army framework runs up to $20 billion. Shield AI's Hivemind is being tested as software for future uncrewed aircraft. Once those systems sit inside programs of record, enterprise contracts and classified networks, the oversight debate becomes harder. You can write human judgment into policy, but the operating environment is being built for faster recommendations, faster approvals and more targets moving through the pipeline.
Congress is not irrelevant here. It can demand reports, set waiver rules and force the Pentagon to explain where humans actually sit in the chain. But it is moving at legislative speed while the department buys at wartime speed. On the current path, the next argument will not be whether AI belongs in targeting. It will be whether the human in the loop still has enough time, information and authority to mean anything.
Also read: The White House just put a government checkpoint between OpenAI and the public • Onsemi bets $7 billion in stock on a world where AI runs at the edge, not the cloud • General Intuition raises $320 million on the thesis that video game footage is the most underrated training data in robotics