Anthropic wins its Pentagon case — six months later
A federal judge called the Pentagon's supply-chain-risk label on Anthropic unlawful retaliation. The vendor won. It still took six months. Plan for the gap.
A federal judge ruled Thursday evening that the Pentagon's designation of Anthropic as a supply chain risk was unlawful, ordering it removed. Judge Rita Lin of the Northern District of California found the government had engaged in unlawful retaliation under the First Amendment and denied Anthropic the process the Fifth Amendment requires. Anthropic won cleanly. It also spent roughly six months winning, which is the part that should shape how you plan.
What actually happened
Per TechCrunch and Forbes, the dispute started in February when Anthropic declined Pentagon terms that would have removed its limits on using Claude for fully autonomous weapons and domestic mass surveillance. In March the department designated the company a supply chain risk — a statutory label for entities posing a sabotage or subversion threat — and federal agencies were directed to stop using its technology. Anthropic filed two suits in March, one in California and one in Washington, D.C.
Judge Lin's ruling found the designation arbitrary and capricious, and read the record as showing the government wanted to make a public example of Anthropic for criticizing it rather than acting on any articulable security concern. She wrote that invoking national security is not a blank check to retaliate against critics.
The caveat matters: the D.C. case is still pending. We covered the preliminary injunction and the summary-judgment hearing earlier this year, and the shape hasn't changed — a court can order a label removed faster than an agency will act on it. Anthropic's public response was to say it remains focused on working with the government on national security applications.
Why a vendor ban you're not party to matters for your business
You are never going to be designated anything by the Department of Defense. You are going to be downstream of a company that gets designated something by somebody — a regulator, a payment network, an app store, a cloud provider's acceptable-use committee.
Look at the clock on this one. February: terms rejected. March: designation, agencies told to stop. Late March: injunction. July: summary judgment argued. August 27: ruling. That is six months from "your vendor is unusable" to "your vendor was wrongly made unusable," and the second sentence does not restore anything you lost during the first. Every customer who had built on Claude inside a federal agency spent that window doing something other than shipping.
The operator's version of this is short:
- Measure your switch time, don't estimate it. Pick a Tuesday, route production traffic through your second model provider, and time how long it takes to get back to normal quality. That number is your actual exposure.
- The second provider has to be warm. An account you've authenticated once is not a fallback. A prompt suite you've never run against it is not a fallback.
- Own the artifacts. Prompts, evals, tool definitions, and golden outputs live in your repo. If your accumulated tuning only exists in a vendor console, portability is a slide, not a capability.
- Ask which of your dependencies would disqualify you. If you sell into government, healthcare, or education, that list exists whether or not you've written it down.
Legal remedies run in months. Your business runs in hours. Build for the gap between those two numbers, because that gap is the only thing a court ruling can't shorten.
Key takeaways
- Judge Rita Lin (N.D. Cal.) ruled Thursday evening that the Pentagon's supply-chain-risk designation of Anthropic was unlawful and ordered it removed
- The court found unlawful First Amendment retaliation, a Fifth Amendment due-process failure, and an arbitrary and capricious decision
- Timeline: terms rejected in February, designated in March, injunction in late March, ruling on August 27 — roughly six months of disruption
- Anthropic's separate Washington, D.C. case is still pending
- A favorable ruling restores the vendor, not the quarter you spent without it
- Time your actual provider switch with real traffic; keep the fallback warm and your prompts and evals in your own repo
Could you switch model providers this week? We build AI systems behind one adapter, with prompts, evals, and outputs stored in infrastructure you own — so a designation, a price change, or an acquisition costs you an afternoon instead of a quarter. See how we build vendor-agnostic systems.
Sources: TechCrunch, Forbes.
- #anthropic
- #vendor-risk
- #procurement
- #ai-governance
- #portability
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Tencent Hy4 preview: 770B open weights, Apache 2.0
Tencent open-sourced a 770B-parameter MoE flagship with a 1M context under Apache 2.0. The open-weights fallback in your cost model just got a lot more credible.
Read itModel Hardware Standard: a driver layer for your machines
Anthropic opened a research preview of the Model Hardware Standard, a spec that lets AI agents drive lab and factory equipment. Integration drops from weeks to hours.
Read it