Every B2B buyer running on one AI model needs a divorce clause this quarter

In July 2025 Anthropic signed a $200 million contract with the Pentagon. On 3 March 2026 Defense Secretary Pete Hegseth decided that buying AI goods or services from Anthropic was a supply-chain risk. On 25 September a panel of the DC Circuit voted 2-1 that he was entitled to do so. The military had been running Claude since 2024. It took less than a year to go from flagship partner to blacklisted supplier.

I expected to write about Anthropic's legal odds, which look reasonable. A federal judge in San Francisco threw out the parallel designation in August as unlawful retaliation, Judge Karen LeCraft Henderson dissented in Washington, and the panel has paused its own ruling so Anthropic can seek a rehearing or go to the Supreme Court. What stuck with me instead was the second half of the ban, the part aimed at contractors.

The designation stops defence contractors using Claude in their Pentagon work. Think about who that reaches. A prime contractor buys analytics from a data firm, a ticketing tool from a SaaS vendor, document review from a services shop. If any of those products has Claude wired into it and touches the Pentagon contract, the prime's compliance team now has a problem, and within a week it becomes the vendor's problem. Most of those vendors never signed anything with the Pentagon. They picked a model because it scored well on their evals and the API price worked.

The Pentagon's own response tells you how the biggest buyer in the room thinks. It is reportedly looking at replacing Claude with Grok, Gemini and OpenAI's GPT models. Once the politics made one supplier impossible, the department treated frontier models as substitutes. Your customers will reach the same conclusion, and they will expect you to have reached it first.

The risk also runs from the vendor's side. The majority opinion says Claude's built-in restrictions more than once stopped it doing tasks government users asked for, and there was a dispute over whether the contract barred use in an ongoing overseas operation. Anthropic's red lines on fully autonomous weapons and domestic mass surveillance strike me as defensible, and the judges wrote that they had "no reason to doubt" its intentions. But a usage policy is a contract term the buyer rarely negotiates and the vendor can tighten. Emil Michael, the Pentagon's research chief, framed the ruling as keeping private companies out of the chain of command. A bank or a hospital group can make the same complaint about its own operations with far less theatre.

The obvious reply is that almost nobody reading this is a defence contractor, the ruling is not final, and commercial buyers can keep using Claude through whatever channel they like. All true today. It was also true for every vendor that sold into the Pentagon in 2024 without imagining a supply-chain label. Anthropic itself says the designation has cost it billions and damaged its reputation ahead of a planned IPO. When a supplier is fighting a sitting president who posts about its CEO by name, and that CEO has just called for an industry-wide slowdown and been left off the guest list for the Xi Jinping state dinner, you are underwriting political risk whether you meant to or not. The same applies to the replacements. Grok belongs to a company whose owner has his own history with the White House, and nobody should assume OpenAI or Google are immune.

So I think buyers should stop treating the model as a single-vendor purchase. The exit clause worth having lets you route workloads to a second model without penalty, keeps your prompts, evaluation sets and fine-tuning data in a portable form you own, and lets committed spend terminate or transfer if a government designation or a unilateral policy change makes the product unusable for a customer. It also needs a fallback that has already handled real traffic, so switching is a config change rather than a project.

Pull your current model contract and search it for any termination right triggered by a regulatory designation of the vendor. If you cannot find one, send the redline before your next renewal, and in the meantime run your twenty highest-volume prompts through a second provider this month, log where the output breaks, and put a price on fixing it. That number is what your defence-adjacent customers will ask you for once their lawyers finish reading the Katsas opinion.

Prompted by U.S. appeals court upholds Pentagon designation of Anthropic as supply chain risk, CNBC.