Skip to content

ATAILA Newsroom · Budapest ·

2026-W40

Claude stays on the Pentagon's blacklist. Access to a model is now politics.

A vendor refused two uses of its model. Its government answered by removing the model from an entire department and barring contractors from using it there. A federal appeals court has now said the government was within its powers. HWSW's Koi Tamás reported the ruling on 28 September, and a day later Bitport reported the vendor's leaked stock-market prospectus. Both stories are about the same company, and that company's models draft this newsroom. So this is a response we owe twice over.

The article we are responding to

“Hiába nyert pert, marad tiltólistán az Anhtropic”

hwsw.hu, 28 September 2026

What the court decided

The opinion is dated 25 September 2026, in Anthropic PBC v. United States Department of War, No. 26-1049, consolidated with a second petition. Judge Katsas wrote for the court, Judge Rao joined, Judge Henderson dissented. The statute is the Federal Acquisition Supply Chain Security Act of 2018, which lets an agency head exclude a supplier whose products present a "supply chain risk" to national security.

The opinion lays out how it got there. Anthropic had already relaxed many of its use restrictions for the department and built a separate government model. It kept two contractual exceptions, which it called critical: Claude would not be used for lethal autonomous warfare or for mass surveillance of Americans. The department asked for permission to deploy Claude for "all lawful uses". Negotiations stalled. On 3 March 2026 Secretary Hegseth made a formal determination that Claude in department systems "presents a significant supply chain risk". On 6 March the department's chief information officer ordered Anthropic products removed from its systems within 180 days and prohibited contractors from using them in their work for the department. Anthropic petitioned the court on 9 March, later asked for reconsideration, was refused on 3 June, and petitioned again on 17 June.

The majority's reasoning, as HWSW summarises it and as the opinion confirms, is that restrictions built into a model can themselves be a national-security risk in a military setting, and that it is the President and the Secretary of War who weigh those risks. The closing line is short: "we deny the petitions for review." HWSW adds a detail that matters for anyone counting on a reversal: in August a federal judge in California, Rita Lin, found a parallel Pentagon decision unlawful, but the two decisions rest on separate statutory powers, so neither overrides the other.

2 to 1

the DC Circuit panel: opinion by Judge Katsas, joined by Judge Rao; Judge Henderson dissenting (opinion of 25 September 2026)

180 days

the deadline in the 6 March 2026 memorandum ordering Anthropic products removed from the department's systems; contractors were barred from using them in work for the department

2

uses Anthropic refused to license, in its own words the "two critical exceptions": lethal autonomous warfare, and mass surveillance of Americans

$518 billion

planned spending on cloud, computing and infrastructure in the leaked IPO draft (Euronews, citing the FT; Bitport, citing Reuters)

$34 billion of $42 billion

the part of the 2025 net loss that was an accounting charge linked to financing; the operating loss was more than $8 billion on almost $4.6 billion of revenue (Euronews)

about a quarter

of 2025 revenue came from two customers, per the draft

The draft prospectus, the same week

On 29 September Bitport, citing Reuters, summarised a draft of Anthropic's registration statement for a stock-market listing. The draft is not a public filing; Euronews, citing the Financial Times, says it was shared with a small group of partners. Bitport's figures: a net loss of $42 billion for 2025, $518 billion to be spent on cloud, computing and infrastructure, and a quarter of revenue tied to two customers. Nearly a third of the filing, Bitport writes, is risk factors, among them catastrophic or existential AI risk and models that may resist being switched off.

Two things Bitport left out, both from the Euronews account. Of the $42 billion loss, roughly $34 billion was an accounting charge linked to financing; the operating loss was more than $8 billion, on revenue that had risen twelvefold to almost $4.6 billion. And revenue in the second quarter of 2026 alone reached $11.5 billion. A vendor losing $42 billion and a vendor growing twelvefold are both in the document. Quoting only the first would be cheap, so we quote both.

What the draft does establish, in its own numbers, is the shape of the business every API customer is a tenant of: enormous committed spending on other people's infrastructure, revenue concentrated in a handful of buyers, and a risk section that discusses the product's own behaviour. Those are the numbers that today's token prices will eventually have to answer to.

What a Hungarian buyer should take from a dispute it has no vote in

Nothing in this case touches a company in Budapest directly. The ruling concerns United States defence procurement. A Hungarian law firm's Claude contract is exactly as valid on 1 October as it was on 24 September, and we say that plainly because it would be easy to imply otherwise.

What the case shows is narrower and more durable. Access to a frontier model is decided in the vendor's home country, by two parties that are not you. The vendor's own usage policy can remove a use you need; in this case Anthropic refused two. The vendor's government can remove the vendor; in this case it did, for a whole department and its contractors, and a court has said that was lawful. Both decisions were taken on the other side of the Atlantic, on grounds a European customer cannot vote on, and both would reach an API customer the moment they are extended.

The practical answer is the principle we stated about logs on 26 September, for a different reason, applied this time to the model: a fallback you control. A model whose weights you hold, running on hardware in your own jurisdiction, is not subject to a procurement determination in Washington or to a policy change in San Francisco. It is slower to improve and it is yours. The second half of the answer is contractual: a documented exit, tested before you need it, for the day a model you depend on is withdrawn for reasons that have nothing to do with you.

Disclosure, and where we are weak

This newsroom is drafted with Anthropic's models and, as we described on 11 September, approved by a person; our own platform can also route customer requests to Claude. We are a dependent writing about our dependency. We have tried to write it the way we would about any other vendor.

The blacklisting happened because Anthropic refused to license mass surveillance of Americans and autonomous weapons. Many readers, and we are among them, will count that in the company's favour rather than against it. The court did not find Anthropic did anything wrong; it found the Secretary was allowed to decide.

The prospectus is a leaked draft, and roughly $34 billion of its headline loss is an accounting charge. A vendor growing twelvefold is not a vendor about to disappear. We quote the numbers because they describe the economics behind the price list, not as a prediction.

What survives all three concessions is the pattern. The vendor and its government decide who may use the model and for what. The customer finds out afterwards. If that is acceptable for your workload, the public API is the better product this week. If it is not, the fallback has to exist before the letter arrives.

Sources, and what we did not open

The story we are answering: Hiába nyert pert, marad tiltólistán az Anhtropic (the misspelling is in the original headline), Koi Tamás, hwsw.hu, 28 September 2026, 14:22. Opened and read in full.

The opinion: Anthropic PBC v. United States Department of War and Peter B. Hegseth, United States Court of Appeals for the District of Columbia Circuit, No. 26-1049 consolidated with 26-1162, argued 19 May 2026, decided 25 September 2026, 51 pages, downloaded as PDF from Courthouse News and read as text on 1 October. The panel, the statute (41 U.S.C. § 4713), the 3 March determination, the 6 March memorandum with its 180-day deadline, the two contractual exceptions, the 3 June refusal of reconsideration and the closing sentence are all quoted from it. The August ruling by Judge Rita Lin in California is as reported by HWSW; we did not open that decision.

The draft prospectus: Súlyos kockázatokról ír az Anthropic tőzsdére lépését előkészítő beadvány, bitport.hu, 29 September 2026, citing Reuters; and Euronews, Quirino Mealha, 29 September, citing the Financial Times, for the accounting charge, the operating loss, the twelvefold revenue growth, and the second-quarter revenue. The Reuters and FT originals were not opened, and no public SEC filing exists to check; every figure above is as those two outlets report it.

Also seen, not used: Portfolio's 28 September analysis Az AI-hozzáférés is geopolitikai kockázattá vált, behind the Portfolio Signature paywall; we read only the teaser.

The original article

“Hiába nyert pert, marad tiltólistán az Anhtropic”

hwsw.hu, 28 September 2026