OpenAI's Pentagon deal the morning after Anthropic's ban signals how Washington now picks AI winners

The Trump administration designated Anthropic a national security supply-chain risk on February 27, 2026, after it refused to drop limits on autonomous weapons and mass surveillance, and OpenAI signed a Pentagon contract within hours. The Pentagon later excluded Anthropic from a list of eight AI companies approved for classified networks, while Cohere gained enterprise interest and merged with Aleph Alpha to position as a sovereign AI alternative.

When the Trump administration designated Anthropic a national security supply-chain risk in February 2026, OpenAI had a Pentagon contract signed within hours. That sequence tells you almost everything about how the AI industry's relationship with Washington has changed. The facts are stark. On February 27, 2026, Defense Secretary Pete Hegseth gave Anthropic a 5:01 p.m. deadline to drop its contractual limits on autonomous weapons and domestic mass surveillance. Anthropic refused. By that evening, Anthropic was designated a supply-chain risk to national security, the first time that label had ever been applied to an American company, and President Trump ordered federal agencies to stop using its technology across a six-month phase-out. Within hours, Sam Altman was announcing OpenAI's agreement with the Department of Defense. The timing was not subtle. What makes that sequence worth examining isn't the speed of it. It's what OpenAI agreed to. Altman said publicly that two of his company's core safety principles were prohibitions on domestic mass surveillance and on directing autonomous weapons systems, and that the DoD had agreed to those limits. These were, almost word for word, the same conditions Anthropic had asked for and been denied. So the Pentagon either softened its position overnight or had always been willing to accept those terms from a different company. Neither explanation is flattering to the administration's account of why Anthropic specifically had to go. Security analysts have since pointed out that Anthropic's models don't present a uniquely greater risk than competing frontier AI systems. That framing matters. If the technical risk is roughly equivalent across labs, then the "supply-chain risk" designation starts to look less like a security call and more like a political one. Anthropic has since sued the Pentagon, arguing the designation violates its First Amendment rights and exceeds the government's statutory authority. The case is ongoing. On May 1, the Pentagon announced agreements with eight AI companies to deploy models on its Impact Level 6 and Impact Level 7 classified networks: SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, Oracle, and AWS. Anthropic was not among them. That list is now effectively the approved vendor pool for U.S. defense AI, and the cost of admission appears to involve demonstrating a willingness to work with Washington on Washington's terms, not negotiating from principles. For Cohere, the fallout has been more quietly useful. The Toronto-based company has reported a surge of inbound enterprise interest since the Anthropic designation, with customers and defense contractors who had built Claude integrations now looking for compliant alternatives. Whether that momentum is durable is a fair question. Cohere isn't yet in the Pentagon's classified network consortium, and once Anthropic's legal dispute resolves or its regulatory status clarifies, a portion of those leads will likely return to Claude. What Cohere is doing with the window, though, is more interesting than riding a short-term windfall. In April, Cohere completed a merger with Aleph Alpha, the Heidelberg-based AI company, at a combined valuation of roughly $20 billion. Schwarz Group, the German retail conglomerate behind Lidl, committed $600 million to Cohere's upcoming Series E as part of the deal. The Canadian and German digital ministers attended the announcement in Berlin. Both governments have now directly endorsed the combined entity as a sovereign AI alternative to U.S. hyperscaler dependence, which is a genuinely different kind of customer relationship than winning a contract on price. The Cohere-Aleph Alpha structure is worth understanding on its own terms: Cohere shareholders hold 90% of the combined company, making it an acquisition dressed up as a merger. Aleph Alpha had built its reputation on European data sovereignty and government trust, particularly in Germany. Cohere's model catalog and enterprise go-to-market capability are substantially larger. Put them together and you get a company that can credibly pitch NATO governments, EU institutions, and multinationals on an AI stack that routes no data through AWS us-east-1. That's a real commercial proposition, not just a talking point, and it would exist regardless of what happened to Anthropic in February. Back in the U.S., Anthropic released Claude Fable 5 on June 9, a Mythos-class model made available for general use with safeguards applied. Claude Mythos 5, the same underlying model with some restrictions lifted, is being deployed in limited availability through Project Glasswing in collaboration with the U.S. government, which adds a layer of irony to the whole saga. Anthropic is still working with American government clients. The company that was banned from federal procurement is, quietly, not entirely banned. The cleaner lesson here isn't about Anthropic's safety principles or OpenAI's opportunism. It's about what Washington has made clear it expects from AI companies that want federal contracts: relationship-first, then technology. Google, Microsoft, AWS, and Oracle had Pentagon agreements in place before any of this happened. They had cultivated those relationships for years, through lobbying, through FedRAMP certifications, through the slow work of becoming infrastructure the government trusts. Anthropic tried to negotiate on the merits of a contractual clause and found out that wasn't how it worked. OpenAI appears to have understood earlier that the path to classified networks runs through Washington's comfort with the company, not just the model. The AI companies that win the next decade of government contracts won't necessarily be the ones with the best benchmarks. They'll be the ones that figured out, sooner than their competitors, that the Pentagon isn't a customer you close on a product demo. Also read: The Anthropic blackout showed every AI startup what export control authority over models actually looks like https://startupfortune.com/the-anthropic-blackout-showed-every-ai-startup-what-export-control-authority-over-models-actually-looks-like/ • Enterprise AI budgets hit a wall and the reckoning is reshaping how companies spend and how founders pitch https://startupfortune.com/enterprise-ai-budgets-hit-a-wall-and-the-reckoning-is-reshaping-how-companies-spend-and-how-founders-pitch/ • Old EV batteries are quietly becoming the cheapest way to power AI data centers https://startupfortune.com/old-ev-batteries-are-quietly-becoming-the-cheapest-way-to-power-ai-data-centers/