The Released Emails Showed Why the Talks Collapsed. Anthropic Refused the Pentagon's Demands, While OpenAI Moved Closer to the Administration ── On July 2, 2026, negotiation emails between Anthropic and the Pentagon (the U.S. Department of Defense) were released as court filings. The same day, OpenAI was reported to have offered the U.S. government a 5% stake in the company. The two stories give opposite answers to the same question: how should an AI company deal with the military and the government?

On July 2, 2026, a set of emails exchanged between Anthropic (the AI company that builds Claude) and the Pentagon (the U.S. Department of Defense) was released as part of a court filing. The dispute was not about access to AI. It was about whether a contract could ban two uses: fully autonomous weapons (AI weapons that attack without human judgment) and domestic surveillance. The same day, OpenAI was reported to have offered the U.S. government a 5% stake in the company — worth about $42.6 billion — to ease political pressure. One company is fighting the government in court. The other is offering the government its own shares. The contrast shows, in concrete terms, where AI companies split on business, principles, and philosophy.

01What the Emails Revealed: The Fight Was Not About Access, but About Whether Uses Could Be Banned

On July 2, 2026, the negotiation emails between Anthropic and the Pentagon were released as court filings. The WSJ (the Wall Street Journal) broke the story. The email exchange was between Emil Michael, Under Secretary of Defense for Research and Engineering (and a former Uber executive), and Anthropic CEO Dario Amodei. The two had spent months working out terms. The whole process is now visible, with dates and exact wording.

The point of conflict the emails reveal is surprising. Whether the Pentagon could use Claude was never the issue. Anthropic had already agreed to supply the military. The fight was over the contract itself — whether "prohibited uses" could stay in the written agreement. Anthropic wanted to ban two uses: fully autonomous weapons (AI weapons that attack without human judgment) and domestic surveillance (using AI to monitor a country's own citizens).

Michael demanded a standard of "all lawful uses" — a condition that allows anything the law permits. Amodei replied that he could not accept it. His reason was specific. U.S. law allows certain forms of domestic surveillance. So a contract that says "anything legal is fine" leaves no way to stop the use of Claude for domestic surveillance. The ban would lose its meaning.

Michael called Anthropic's refusal "just not workable." Amodei, for his part, pointed out that the Pentagon's proposed document had "completely removed" the lines Anthropic could not cross. Both men's words show that this was not an emotional clash. It was a head-on dispute over contract terms. The government wanted a blank check on uses. The company wanted limits on uses. Neither side gave in.

The day after the talks collapsed, Defense Secretary Hegseth announced that Anthropic would receive a FASCSA designation — a designation that bars a company from government procurement on the grounds that its supply chain poses a risk. The negotiation ended there. With the emails now public, the timing is on the record: the designation came immediately after the breakdown of talks. From a business standpoint, this was the moment the government switched roles, from customer to regulator.

02From a Terminated $200 Million Contract to a Lawsuit: The Relationship Collapsed in a Year and a Half

The relationship did not start out hostile. In November 2024, Anthropic began supplying Claude to the U.S. military through Palantir (a data analytics company serving governments) and AWS (Amazon's cloud service). In June 2025, it announced a government-only version, "Claude Gov." At that point, Anthropic was one of the few AI companies working with the military.

The turning point came in January 2026. The Pentagon terminated the contract and removed Claude from GenAI.mil, the military's generative AI platform. The contract was reported to be worth about $200 million. The following February, Michael pressed Anthropic to "cross the Rubicon" — meaning cross a line there is no coming back from. The demand was to drop the use restrictions. First came the real damage (the terminated contract) in January, then the ultimatum in February.

Anthropic took the matter to court. On March 9, 2026, it filed suit in the U.S. District Court for the Northern District of California (case number 3:26-cv-01996). The result came quickly. On March 26, Judge Lin found the FASCSA designation to be "classic illegal First Amendment retaliation" — unlawful retaliation against speech protected by the First Amendment, the constitutional rule that protects free expression. The judge granted a temporary block on the designation. At the district court level, the company's argument prevailed.

But on April 8, the appeals court — the D.C. Circuit Court of Appeals — denied Anthropic's emergency motion. The designation stayed in effect. As a result, as of July 2026, work is under way to remove Claude from military systems over a 180-day period. Even a partial win in court does not undo business losses already locked in. Viewed as litigation risk management, this is a textbook case: you can win the lawsuit and still not get the revenue back.

One more fact cannot be ignored. Michael's financial disclosure (the system under which senior government officials publish their assets) included shares of xAI, an Anthropic competitor. Whether that holding actually influenced his decisions remains speculation. But the picture — the official who led the decision to exclude Anthropic held stock in a rival — has been reported as a suspected conflict of interest, a clash between an official's duties and personal financial gain.

03OpenAI Moved the Other Way: It Changed Its Terms, Signed a Contract, and Offered the Government Equity

OpenAI (the AI company that builds ChatGPT) has moved in exactly the opposite direction from Anthropic. Its turning point was January 10, 2024. On that day, OpenAI deleted the language banning use for "military and warfare" from its usage policies. The change, reported by the tech outlet TechCrunch, looked like a quiet edit to fine print at the time. In hindsight, it reads as the first move toward the government and the military.

The effect of the policy change turned into numbers about a year and a half later. On June 16, 2025, OpenAI signed a $200 million contract with the U.S. Department of Defense. The work: developing prototypes of frontier AI, with a term running through July 2026. OpenAI has not hidden the deal. It hosts a page on its own site titled "Our agreement with the Department of War," presenting the government relationship openly. During the same period when Anthropic was fighting to keep prohibition clauses in its military contract, OpenAI was displaying its military agreement as a badge.

OpenAI also closed the distance to politics. In January 2025, President Trump personally announced Stargate — a $500 billion AI infrastructure plan by four companies: OpenAI, SoftBank, Oracle, and MGX. No government money is in the plan. Even so, the format itself — an announcement by the president in person — gave OpenAI strong political backing. OpenAI went further, proposing "Classified Stargate," a government-only data center for classified data, and asking the administration for AI-related tax cuts, according to the industry outlet DCD.

Then, on July 2, 2026 — the same day Anthropic's emails were released — a bigger step surfaced. According to the U.S. business outlet CNBC, OpenAI offered the U.S. government a 5% stake in the company to ease political pressure. At the company's $852 billion valuation, that stake is worth about $42.6 billion. The offer is still only a news report; neither side has formally confirmed it, and no deal has been made. Still, if the offer is real, it means a great deal. It suggests OpenAI treats the government's favor as a business resource that can be bought with equity.

Seen through a business lens, OpenAI appears to treat political backing as a resource to procure, like electricity or chips. Change the terms to clear the obstacle. Build a track record with a contract. Gain credibility through a presidential announcement. Then lock in the relationship with equity. This looks less like a change in principles than a consistent investment decision, made after concluding that the government is both the biggest customer and the biggest risk. What Anthropic is trying to protect in court, OpenAI is putting on the table as a bargaining chip.

04Google and Meta Also Dropped Their Limits: Anthropic Is the Only Company Still Holding the Line

OpenAI was not the only company to drop restrictions on military use. In February 2025, Google deleted from its AI principles the pledge not to use AI for weapons. That pledge had stood for seven years, since 2018. After the deletion, things moved fast. In July 2025, Google won a contract worth up to $200 million from the Pentagon's CDAO (the Department of Defense's AI and data office). In April 2026, it reached an agreement with the Pentagon to supply its AI, Gemini, for "all lawful use" — the very standard Anthropic refused, the refusal that ended its Pentagon talks. Google's agreement has been reported as more permissive than OpenAI's contract.

There was opposition inside Google. About 600 employees sent a protest letter to CEO Pichai. But unlike in 2018, during Project Maven (a program in which Google helped analyze military drone footage and withdrew after employee protests), the company has not changed course. Eight years ago, employee voices stopped a contract. This time they did not. The difference suggests that ethical dissent inside AI companies carries less weight than it used to. That said, this is a comparison within one company, and other explanations remain — the size of the contracts, or the changed relationship with the administration.

Meta moved even earlier. On November 4, 2024, it lifted the ban on military use of Llama (the AI model Meta releases publicly) for the United States. It began supplying defense and national security agencies, plus military and defense contractors such as Lockheed Martin, Anduril, Palantir, and Booz Allen. The timing is telling. The change came right after reports that researchers tied to China's People's Liberation Army had used an older version of Llama to build a military chatbot. If a ban cannot stop foreign militaries from using the model, better to supply your own military officially. That logic justified lifting the restriction.

Line up each company's moves and the pattern is clear.

In just over two years, from 2024 to 2026, three of the major AI companies loosened or scrapped their limits on military use. Anthropic is now the only one still holding a contractual red line — a line that must not be crossed — and clashing with the government over it. Measure the distance between each company's published principles and its actual behavior, and only Anthropic has stayed on the side of its principles, paying for it with pressure from both the market and the government. Whether this is principle carried through, or business isolation, is a question with no answer yet.

05Are Principles a Signboard or a Contract? Facing the Same $200 Million, the Companies Chose Opposites

Every one of these companies posts principles and usage policies (the rules users and customers must follow) on its official site. The difference appeared when the words collided with action — and each company had to choose which one to bend. Anthropic held its line — no use for fully autonomous weapons or domestic surveillance — to the end of the negotiation, and lost a roughly $200 million contract and its relationship with the government. The other three companies did the reverse: they deleted the words and won the contracts.

Put the numbers side by side and the picture sharpens. The contract Anthropic lost by holding its line, and the contracts OpenAI and Google each won after erasing their limits, are all worth roughly the same $200 million. Facing the same price tag, one company chose its principles and three chose the contract. Posting principles costs nothing. The cost arrives only at the moment principles and revenue collide head-on. Between 2024 and 2026, that moment came to every major AI company in turn, and every company answered with its actions.

Another point worth attention is where the change was made. What OpenAI rewrote in January 2024 was not a statement of principles. It was the usage policy — the document that sets contractual terms with customers. Leave the principles document untouched, and quietly rewrite the contract document. That way the signboard stays intact while only the practice changes. Anthropic's ban worked the opposite way: as the emails filed in court show, it was an actual contract term, raised in negotiation with the government and defended at a cost of $200 million. When judging a company's principles, do not look at how fine the words sound. Look at what the company gave up to keep them.

Whether Anthropic's choice makes business sense, though, is still undecided. If a reputation as "the supplier that will not bend its line under pressure" attracts safety-minded customers and talent, and generates more value than the lost government contract, then the $200 million was cheap advertising. If it does not, it was simply a loss. This is an interpretation, not a fact. Which way it goes will show up in Anthropic's future enterprise revenue and hiring.

06If a Company That Sells Safety Bends on Safety, What Is Left?

Anthropic has made AI safety its business signboard. Drawing its own line around safe use is the product itself. Accepting the Pentagon's "all lawful uses" standard would have left no way to stop the use of Claude for fully autonomous weapons or domestic surveillance. That is not a concession on one contract term. It is closer to erasing the company's reason to exist. With OpenAI, Google, and Meta having loosened their limits one after another, Anthropic is the only company left holding this line. Scarcity itself can become the product's value — that was the interpretation in the previous section. Whether that value actually turns into revenue is not yet known.

The problem this case exposed goes beyond one company's business decision. The structure is this: the government punished a company for keeping its principles, using an official designation to do it. As noted above, the district court found the designation to be classic illegal retaliation against free speech. Even so, the appeals court denied the emergency motion, and the designation remains in force. Only the real-world harm of a measure ruled illegal keeps moving forward. This asymmetry works as an unspoken lesson to other AI companies. It functions as a warning: draw a line and this is what happens.

The issue has already become a matter of institutional debate in Congress. The Congressional Research Service (CRS, the research agency that provides neutral analysis to the U.S. Congress) issued a report for Congress on the Pentagon-Anthropic dispute and autonomous weapon systems (IN12669). How far can the government force an AI company to accept certain uses? How is a company's freedom to keep use restrictions protected? This is starting to be treated not as one contract dispute but as the fight that will shape the template for AI procurement rules. The next entity to decide where the line goes may be the law, not a company's usage policy.

There are two things to watch next. One is the 180-day deadline for removing Claude from military systems. The other is the outcome of the appeal. Whether the district court's finding of "illegal retaliation" survives review will decide the answer to the core question: can the government punish a company for keeping its principles? The cost and the payoff of a safety company refusing to bend on safety will show up, in numbers, in those two results.