跪拜 Guibai
← All articles
Artificial Intelligence

Anthropic Bans Claude Access for Any Company Majority-Owned by Chinese Capital

By IT乐手 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Any team building on Claude whose cap table includes Chinese investors above 50% — even through a US or Singapore entity — lost API access overnight with no grace period. The ban turns model-provider lock-in from a cost concern into an existential supply-chain risk, and it accelerates the fragmentation of the global LLM market into separate US-aligned and China-aligned ecosystems.

Summary

The policy, effective immediately on September 5, 2025, covers all mainland China-registered firms, overseas subsidiaries of Chinese parents, and startups with majority Chinese investment. ByteDance's Singapore-registered Trae code editor is among the products caught in the ban. Anthropic cites US export controls, national security, and legal compliance risks from authoritarian states as justification.

Chinese model providers responded within hours. Zhipu launched a Claude-compatible migration plan requiring only three lines of code changes at one-seventh the price. Alibaba, Baidu, and Tencent all position their own large models as drop-in replacements.

The cutoff reflects a broader shift: AI infrastructure is now treated as a geopolitical asset subject to the same export-control logic as advanced semiconductors. DeepSeek's efficiency breakthrough intensified US wariness about Chinese teams reverse-engineering or competing from Western-origin models.

Takeaways
Anthropic's restriction applies to any company where Chinese entities hold more than 50% equity, directly or indirectly, no matter the country of incorporation.
API access was cut immediately on September 5, 2025, with no transition window; accounts were disabled the same day.
ByteDance's Singapore-registered AI code editor Trae is affected because of its ultimate Chinese ownership.
Anthropic groups China with Russia, Iran, and North Korea as "adversary nations" under US export-control rules.
Zhipu released a drop-in Claude API replacement the same day, requiring three lines of code changes and costing 1/7th of Claude's price.
Alibaba, Baidu, and Tencent all offer domestic large-model APIs positioned as alternatives for displaced Claude users.
Individual user accounts are not targeted by the policy, though China's GFW already blocks access to claude.ai without circumvention tools.
Conclusions

Anthropic's equity-based test — rather than a geographic or end-user one — makes the restriction harder to evade through offshore incorporation and signals that US AI firms now treat ownership structure as a compliance vector.

The speed of Zhipu's same-day migration plan suggests Chinese model providers anticipated this cutoff and had API-compatible wrappers ready to capture displaced revenue.

DeepSeek's resource efficiency is treated here not just as a technical achievement but as a threat multiplier that changed how US labs assess the risk of any Chinese access to frontier models.

The ban completes a de facto US industry consensus: OpenAI restricted China earlier, and Anthropic's move leaves Google's Gemini as the last major Western frontier model without an explicit China ownership prohibition.

Concepts & terms
Export controls on AI models
US regulations that treat advanced AI systems as dual-use technology subject to export restrictions, similar to semiconductors. Companies must prevent access from designated adversary nations or risk penalties.
API-protocol compatibility migration
A strategy where a competing model provider mimics the request/response format of an incumbent API so that users can switch by changing only the base URL and authentication, without rewriting integration code.
From the discussion

The discussion splits between pragmatic acceptance and geopolitical grievance. Some see the ban as an inevitable business and compliance reality, noting domestic alternatives like DeepSeek and GLM already soften the blow. A sharper current treats the move as political theater dressed in legal language, arguing the US unilaterally designates threats and that the ban would vanish if China were weakened or subordinated. A smaller thread mocks the grouping with Russia and North Korea, while others dismiss the topic as non-technical or crack jokes about banning Chinese-language input.

Domestic models like DeepSeek and GLM are already strong enough on coding tasks to make the Claude ban manageable.
The ban is a political decision, not a technical or compliance one; the legal wording is a pretext to occupy the moral high ground.
China is grouped with Russia and North Korea because it is perceived as not fitting in with the US-led order.
Blocking Chinese-language input would be functionally self-defeating, turning Claude into a backward model.
Anthropic's ties to the US Department of Defense make its stance predictable; a founder with experience at a Chinese company faces a binary choice of being seen as anti-China or pro-China.
The ban would not apply if China were fragmented and non-threatening, or if it accepted a subordinate relationship like Japan or South Korea.
Unifying global language and currency is floated as a sarcastic long-term fix to such fragmentation.
Some view the ban as a positive development, though without elaboration.
Featured comments
子非鱼不想说话55640 3 likes

I've noticed a lot of AI self-media or tech people like to chat about this topic, actually tech people aren't suited to discuss it, because this isn't a tech issue at all. All that compliance, rules, and clause jargon is just to have a righteous pretext and seize the moral high ground. So Sudan, Palestine, Haiti, Syria and other refugee nations are fine? Well, that figures, these countries are struggling to survive, so they definitely have no democracy issues, pose no threat, and can't possibly be used to develop competitive AI products. The day we fall apart and become a refugee nation losing our threat value, we can also be removed from the list. Or learn from Japan and South Korea, invite a military presence + sign a Sino-US Plaza Accord and this problem can also be easily solved. After all, hostile nations are unilaterally determined by the US, they say you are, and you are.

ToCodex_AI

This was bound to happen sooner or later, the pressure from business and compliance is right there. Fortunately, there are already quite a few alternatives domestically, DeepSeek, GLM and these are already very capable on programming tasks, plus API relay channels are also improving, the impact isn't as big as imagined, just adapt the workflow a bit.

用户56249810366 2 likes

Should look more at our own reasons, why we're grouped with Russia and North Korea.

老青蛙  · 4 likes

Tell me why then, I've got my little stool ready.

俊之以茉

No matter how you think about it, it's just that way, because you don't fit in with others.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗