FT scoop — OpenAI and Google sell AI to Pentagon-blacklisted Chinese firms via Singapore; only Anthropic refuses
TL;DR
FT: OpenAI and Google supply AI models to Singapore subs of Alibaba, Baidu and Tencent — all three on the Pentagon's 1260H list — legally. Anthropic is the only US frontier lab that blocks China.
The Financial Times reported on July 10 that OpenAI and Google have confirmed selling AI model access to the Singapore-based subsidiaries of Alibaba, Baidu, and Tencent — three Chinese tech giants whose parents all sit on the US Defense Department's 1260H list for alleged ties to the People's Liberation Army. The supply chain is fully legal under current US rules.
US export controls on chips have gone airtight — Nvidia H200 and GB200 exports need individual licenses, tracked shipment by shipment. But the "lighter" AI model layer flows freely through the compliance wrappers of Singapore and Hong Kong subsidiaries. OpenAI told FT it "prohibits direct China access but permits certain overseas affiliates." Google Cloud said its terms of service ban distillation while allowing Singapore users to call the API.
OpenAI has already been bitten once. In June it detected Alibaba-linked API users attempting distillation — training competing models on GPT outputs — and cut off those accounts. The incident is itself a footnote on "current rules can't reach the Alibaba parent": accounts can be banned, the legal layer can't be.
Anthropic is the only one of the three US frontier labs with a hard policy line: neither Chinese firms nor their overseas subsidiaries can use Mythos/Fable. Anthropic concedes enforcement is hard, but the policy is drawn. That stance also explains why Cloudflare's June report pinned Anthropic as the greediest crawler on the web — with China-tier enterprise revenue voluntarily off the table, training crawls have to make up the compute-and-data bill.
Council on Foreign Relations analyst Chris McGuire told FT: "The most advanced AI models should not be made available to companies based in China, regardless of the country from which they access them." The regulated frontier list currently covers only GPT-5.6 and Anthropic Mythos/Fable, but has no "no-selling-to-Chinese-parented-entities" clause. The entity-list logic that governs chips has simply not been ported to the model layer.
Win the bet, and the FT expose pushes the White House to add a model-layer entity list before September, OpenAI voluntarily tightens up, and China faces a two-tier lockout (chips + models). Lose it, and Commerce holds the status quo, 1260H-listed firms keep calling GPT-5.6 Sol via Singapore subs, and Anthropic's lone-gatekeeper stance turns into a pure market-share sacrifice.
via FT / Il Sole 24 ORE / Crypto Briefing / Benzinga / UDN
US export controls on chips have gone airtight — Nvidia H200 and GB200 exports need individual licenses, tracked shipment by shipment. But the "lighter" AI model layer flows freely through the compliance wrappers of Singapore and Hong Kong subsidiaries. OpenAI told FT it "prohibits direct China access but permits certain overseas affiliates." Google Cloud said its terms of service ban distillation while allowing Singapore users to call the API.
OpenAI has already been bitten once. In June it detected Alibaba-linked API users attempting distillation — training competing models on GPT outputs — and cut off those accounts. The incident is itself a footnote on "current rules can't reach the Alibaba parent": accounts can be banned, the legal layer can't be.
Anthropic is the only one of the three US frontier labs with a hard policy line: neither Chinese firms nor their overseas subsidiaries can use Mythos/Fable. Anthropic concedes enforcement is hard, but the policy is drawn. That stance also explains why Cloudflare's June report pinned Anthropic as the greediest crawler on the web — with China-tier enterprise revenue voluntarily off the table, training crawls have to make up the compute-and-data bill.
Council on Foreign Relations analyst Chris McGuire told FT: "The most advanced AI models should not be made available to companies based in China, regardless of the country from which they access them." The regulated frontier list currently covers only GPT-5.6 and Anthropic Mythos/Fable, but has no "no-selling-to-Chinese-parented-entities" clause. The entity-list logic that governs chips has simply not been ported to the model layer.
Win the bet, and the FT expose pushes the White House to add a model-layer entity list before September, OpenAI voluntarily tightens up, and China faces a two-tier lockout (chips + models). Lose it, and Commerce holds the status quo, 1260H-listed firms keep calling GPT-5.6 Sol via Singapore subs, and Anthropic's lone-gatekeeper stance turns into a pure market-share sacrifice.
via FT / Il Sole 24 ORE / Crypto Briefing / Benzinga / UDN
