Confirmed 12 Nov 2026 · Governance A one-minute read

Anthropic's rules took effect: no Claude for firms China owns, wherever they sit.

On November 12 Anthropic's 2026 usage policy and a rewritten Supported Regions Policy took effect. Caveat: the policy text describes no checks, so enforcement is not published.

Until now

Access to frontier AI is limited by where the user sits.

Now

It is limited by who owns the company, anywhere in the world.

Why it matters

A company has written an export rule and applied it to ownership, not geography. A supported-country firm with a Chinese majority holder is now barred. Governments use entity lists for this. Anthropic now runs one of its own, and no regulator approved it.

Established

The policy took effect November 12 and lists about 190 supported regions.
Entities majority-owned or controlled from unsupported regions are barred, wherever staff sit.
Sustained cruelty toward the models is now a written prohibition.

Not established

How Anthropic checks ownership. The policy describes no verification method.
That any account was closed. No enforcement numbers are published.
That identity checks use photographs. That is reported, not in the policy.

The five-minute readThe record, the timeline, the weight and 3 sources

The record

Anthropic's 2026 usage policy and its Supported Regions Policy both carry an effective date of November 12, 2026, and we read both at anthropic.com. The Supported Regions Policy covers Claude.ai, Claude Code, Cowork, the API and developer platform, and Claude models sold through cloud platforms and authorised resellers. It lists roughly 190 supported countries and regions and states that any region not on the list is unsupported. China, Russia, Iran and North Korea do not appear. Taiwan does. Ukraine appears with an exclusion: "except Crimea, Donetsk, Kherson, Luhansk, and Zaporizhzhia regions." Three separate tests restrict use. First, physical location: "Use by persons while physically located in an unsupported region," including staff of otherwise supported companies. Second, incorporation: entities incorporated or headquartered in an unsupported region, and their users and personnel, even when those people sit in a supported country. Third, ownership: entities majority-owned or controlled, directly or indirectly, by persons or entities in unsupported regions, "regardless of whether such individuals are physically located in a supported region." The announcement post at anthropic.com is dated October 8, 2026 and says the Supported Regions page "was updated to clarify enforcement." The usage policy that took effect the same day prohibits "sustained and needless abusive or cruel behavior toward our models," removed the earlier blanket ban on personalised vote and campaign targeting, narrowed the elections section to voter deception and election disruption, extended the weapons ban to weapons-related software and to arming drones and autonomous vehicles, and bars predictive policing and facial recognition. For 11 high-risk areas, among them legal, medical, finance, credit, insurance, housing, employment, education and public benefits, "a qualified person must meaningfully review a High-risk AI Recommendation" and the affected individual "must be clearly told that AI was used to produce it." For physical actions a qualified operator must be able to stop the equipment, which must "hold a safe state if Claude is disconnected." Not in either document: any verification procedure, any ownership test beyond majority, any account-closure process, or any count of accounts affected. Reported elsewhere and not established by us: that risk signals trigger identity verification with a photograph through a third-party vendor, and that failed checks end in account closure.

How it got here

Oct 8
Anthropic publishes the 2026 usage policy with an effective date 35 days out, bans sustained cruelty toward its own models, and drops the blanket ban on personalised campaign targeting. Edition 045.
Oct 16
China's commerce and science ministries move to control exports of Chinese AI and begin registering them that day. Edition 053.
Nov 3
Nvidia tells cloud providers to sell most capacity to non-Chinese firms, the Financial Times reported. A private rule about nationality, written by a supplier. Edition 071.
Nov 8
Nvidia's board opens an independent investigation into whether its founder was the target of a foreign intelligence effort. Edition 082.
Nov 12
Anthropic's usage policy and Supported Regions Policy take effect. Ownership and control, not only location, now decide who may use Claude.

Who has to react

Chinese-owned subsidiaries in Europe and Asia that use Claude; the cloud platforms and resellers that must apply the list; Amazon Web Services and Google Cloud, which sell Claude; corporate counsel who now screen their own cap tables; the Commerce Department, whose entity list this resembles; OpenAI and Google, whose region policies are now the comparison; Anthropic's enterprise sales team.

What to watch

Would strengthen itAnother frontier laboratory publishes a region or ownership policy barring entities controlled from a named country, or Anthropic publishes a count of accounts closed under the ownership test, by June 30, 2027.
Would weaken itAnthropic adds China or removes the ownership test from the Supported Regions Policy, or confirms the clause applies only where sanctions law already requires it, by December 31, 2027.
Would strengthen itAnthropic's listing documents or a government filing cite the Supported Regions Policy as a national-security control, by September 30, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentHigh

ReversibilityModerate

PersistenceStructural

Cost is medium. The rule closes paying accounts and forces every reseller and cloud partner to apply the list, which is work Anthropic does not get paid for, but no plant and no capital is at stake. Commitment is high: a published policy with an effective date, written into the terms that govern the API and the resale channel, cannot be walked back quietly. Reversibility is moderate because a list is edited with one page change, and 190 entries are a drafting decision rather than a contract. Persistence is structural. Ownership screening, once built into enterprise sales, becomes a permanent step in how frontier AI is bought.

The words

Supported region: A country or territory on a company's published list where its product may be used. Anywhere absent from the list is barred, so the list is the rule.
Entity list: A government register of companies that domestic firms may not supply without a licence. Washington keeps one for Chinese technology buyers.
High-risk AI recommendation: Advice from a model affecting someone's health, legal rights, money, job or access to essential services. A qualified human must now review it.
Majority-owned or controlled: More than half the shares, or effective command of decisions, held directly or through intermediaries. The test looks through holding companies.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.