Who Actually Governs AI: Part 3 - Defense

In April 2018, more than 4,000 Google employees signed a letter to their CEO asking the company to cancel a Pentagon contract. The project, code-named Maven, used Google's AI to help analyze drone footage. Google let the contract lapse the following year and published a public pledge: it would not build AI for weapons.

In February 2025, that pledge disappeared from Google's website. No employee letter. No press conference. The page was simply rewritten.

Every major AI lab has now made some version of the same trade. The industry didn't debate whether to take defense money. It found out how much was on the table, and only one company decided the price was too high.

The Year Four Companies Signed the Same $200 Million Deal

In July 2025, the Department of Defense signed contracts worth up to $200 million each with OpenAI, Anthropic, Google, and xAI. The goal, in the Pentagon's own language, was to build AI prototypes that could "improve military advantage, military utility, or enhance military decision making" across the armed forces, from battlefield logistics to intelligence analysis.

The contracts were not one-way sales. Each company agreed to give the Pentagon access to its best models and to brief military officials on how adversaries were using AI. In return, the military briefed the companies on real operational threats. It was less a purchase order than a partnership, and it put four rival labs on the Pentagon's payroll at once.

Nine months later, only one of them walked away from the table.

Did You Know?

The Department of Defense has a second, older name again. In September 2025, an executive order made "Department of War" an official secondary title for the Pentagon, restoring the name the agency carried from 1789 until Congress renamed it in 1947.

The Clause That Broke One Deal and Not the Others

By February 2026, the Pentagon was pushing all four contractors to accept broader terms, including language authorizing use of their AI for "any lawful use." Anthropic read that clause as permission to deploy Claude for mass domestic surveillance and in fully autonomous weapons systems that could select and strike targets without a person approving the decision.

Anthropic said it would adapt many of its policies for the Pentagon, but not those two. The company held that position through weeks of negotiation. When a Friday-afternoon deadline passed without an agreement, the Trump administration ordered federal agencies and contractors to stop doing business with Anthropic.

The Defense Department went further, formally labeling the company a "supply chain risk," a designation normally reserved for security threats, not the loser of a contract negotiation. A federal judge later overturned the ban, but the message to the rest of the industry had already landed: refusing a term the Pentagon wants has a cost, and the government will make you pay it publicly.

OpenAI signed the broader terms instead. Its amended agreement, finalized the same week, called for company engineers to be embedded directly in "warfighting support settings," including combatant commands, so the model could be tuned to military operations in something close to real time.

Did You Know?

Palantir and Anduril, the two companies whose stock and contracts moved fastest during this period, formally teamed up in late 2024, before most of the headline deals were signed. Anduril's Lattice software feeds real-time battlefield sensor data upward, while Palantir's platforms turn that data into the maps and target recommendations that commanders actually use.

The Companies That Never Made the Pledge to Begin With

Palantir and Anduril never had a "no weapons" policy to walk back, because building for the military was always the plan, not a pivot.

That head start is showing up in the numbers. In March 2026, the Army merged 120 separate Anduril contracts into one enterprise agreement worth up to $20 billion over ten years, largely for autonomous counter-drone systems. Around the same time, it consolidated 75 Palantir contracts into a single deal worth up to $10 billion. Palantir's Maven Smart System, a battlefield analysis tool built on the same Project Maven concept Google once walked away from, is on track to become an official Pentagon program of record.

The revenue is following the contracts. Palantir's U.S. government revenue reached $687 million in the first quarter of 2026 alone, up 84% from a year earlier, and the company later raised its full-year 2026 revenue forecast to roughly $8.2 billion. Anduril's annual revenue climbed from about $420 million in 2023 to roughly $1 billion in 2024 to an estimated $2.2 billion in 2025, and its valuation jumped from $30.5 billion in mid-2025 to $61 billion in May 2026, when it raised another $5 billion from investors betting the growth keeps going.

The Policy Language Nobody Noticed Changing First

OpenAI's own reversal predates the Pentagon deal by a year and a half. In January 2024, without an announcement, the company deleted the specific ban on "military and warfare" use from its usage policy, replacing it with softer language about not using its tools to develop weapons. An OpenAI executive later explained that the old blanket ban had been getting in the way of projects the company actually wanted to pursue, like cybersecurity tools for defense agencies.

That pattern, quietly rewriting the rule before the business need arrived, repeated at Google a year later. Where Google's 2018 AI principles had listed specific applications it would never pursue, including weapons and surveillance, the 2025 rewrite dropped that section entirely. The new document describes only what Google will do, with no list of what it won't.

The order of operations matters. In both cases, the policy changed first. The contracts came after, once the public commitment was no longer in the way.

The shift also runs through how these systems get described. Palantir and Anduril's platforms increasingly rely on autonomous software components that gather information and act on it with limited human review at each step, the same AI agent loop that powers far more mundane tools like coding assistants. The difference is what the loop is deciding.

Knowlegic Perspective

None of this happened because the technology suddenly became ready for war. It happened because the money got big enough to make the old policies expensive to keep.

Google's 2018 pledge existed because thousands of its own employees made refusal cheaper than compliance. By 2025, that balance had flipped: the contracts were worth more than the internal cost of quietly rewriting a webpage. Anthropic is the one data point suggesting the old calculation can still hold, but it took a government blacklist to prove the position was sincere rather than convenient.

What actually governs AI's use in war right now isn't a law or a treaty. It's whichever company's terms the Pentagon agreed to sign last, and how much it was willing to pay for the ones that said no.

The industry didn't debate whether to take defense money. It found out how much was on the table, and only one company decided the price was too high.

Read Next: Who Actually Governs AI: Part 4 - Surveillance

Sources & References

• OpenAI Snags $200 Million Contract With US Defense Department, Bloomberg (2025)
• AI Giants Work Hand-in-Hand With the Pentagon, Contracts Reveal, The Intercept (2026)
• Trump administration orders military contractors and federal agencies to cease business with Anthropic, CNN Business (2026)
• Army, Anduril enter into new $20B enterprise agreement, Nextgov/FCW (2026)
• Palantir lands $10 billion Army software and data contract, CNBC (2025)
• Army consolidates dozens of Palantir software contracts into one deal worth up to $10 billion, Breaking Defense (2025)
• Pentagon strikes deals with 8 Big Tech companies after shunning Anthropic, CNN Business (2026)