The Blacklist Stops at Changi

OpenAI and Google are legally selling advanced AI to Singapore subsidiaries of three Pentagon-blacklisted Chinese tech giants — Alibaba, Baidu, and Tencent — because Washington wrote export controls for silicon crossing borders, not software logging in from Changi.

What's new: A Financial Times investigation published July 10 found OpenAI and Google have supplied AI models and services to Singapore-based affiliates of Alibaba, Baidu, and Tencent. Both companies confirmed the sales to the newspaper. All three parent groups sit on the Pentagon's Section 1260H list of "Chinese military companies," expanded by 65 entities on June 8, 2026. The transactions are legal under current U.S. rules.

Why it matters: Washington spent years ring-fencing Nvidia GPUs and lithography tools. The July 10 disclosure shows the software layer never got the same fence. A Hangzhou headquarters cannot touch GPT-5.6 directly. A Singapore subsidiary can sign the contract, route the API calls, and — per reporting on suspected distillation — feed outputs back into rival Chinese models. The chip war had a side door. Now the model war has one too.

The numbers:

• 3 — Chinese tech giants on the 1260H list with confirmed Singapore AI access: Alibaba, Baidu, Tencent
• 65 — new entities added to the 1260H list in the June 8, 2026 Pentagon update, per WilmerHale analysis
• 188 — total entities on the updated 1260H list
• S$300 million+ — OpenAI's May 2026 commitment to its first Applied AI Lab outside the United States, in Singapore
• 28.8 million — exchanges Anthropic alleges Alibaba-linked operators ran against Claude between April 22 and June 5, 2026, through nearly 25,000 fraudulent accounts
• June 30, 2026 — date Pentagon procurement bans on 1260H-listed firms take effect for direct contracts, per the FY2024 NDAA

The quotes:

• OpenAI, to the Financial Times, July 10: "We don't think nationality alone should decide access."
• OpenAI, to the FT: It blocks direct access from mainland China but permits some Chinese-owned companies in jurisdictions where it can enforce safeguards and monitor misuse.
• Google, to the FT: AI services remain available in Singapore and Hong Kong under usage policies that forbid distillation — while acknowledging geographic restrictions alone cannot stop sophisticated circumvention.
• Chris McGuire, former Biden NSC official and CFR senior fellow, to the FT, July 10: Export controls remain the sharpest tool to slow China's AI progress; the most advanced models should stay out of Chinese firms' hands "wherever those firms log in."
• Alibaba, in its U.S. court challenge: The Pentagon's 1260H designation is "arbitrary and capricious."
• Anthropic, June 10 Senate Banking Committee letter: Operators affiliated with Alibaba and Alibaba Qwen generated "more than 28.8 million exchanges with Claude through almost 25,000 fraudulent accounts" in violation of its terms of service.

How the side door works: U.S. export controls rest on named places and named entities. Mainland China is restricted. Singapore is not. A Chinese firm blacklisted in Washington registers a subsidiary in Singapore — on paper a Singaporean business — and buys API credits its Shenzhen parent cannot. Current rules block direct access to specific frontier models, including OpenAI's GPT-5.6 and Anthropic's Mythos and Fable. They do not impose a blanket ban on Chinese-headquartered firms using advanced AI software through overseas units. The gap is wide enough to drive a cloud contract through.

What each lab did:

• OpenAI suspended API access for Alibaba-affiliated users last month after detecting suspected distillation — using a frontier model's outputs to train a rival system — and reported the activity to the U.S. government, per the FT.
• Google has not disclosed enforcement action against a specific Chinese-linked account, though it confirmed Singapore availability under anti-distillation policies.
• Anthropic bars Chinese companies and foreign entities they own from its frontier models entirely. It pressed Congress for broader AI software export controls and named DeepSeek, Moonshot, and MiniMax in earlier distillation allegations.

The investment irony: In May 2026, OpenAI signed a government MOU committing more than S$300 million to "OpenAI for Singapore" — its first Applied AI Lab outside the U.S., with 200+ planned technical roles. Google DeepMind opened its first Southeast Asian research lab in Singapore in late 2025, per CNA and Google. American AI infrastructure is concentrating in the one jurisdiction where Pentagon-listed Chinese subsidiaries can legally rent the models those chips were built to train.

What we cannot verify: The FT's primary reporting is paywalled; secondary accounts relay company confirmations but not contract values or specific model tiers sold to each subsidiary. Anthropic's 28.8 million exchange figure is an allegation in a Senate letter — Alibaba denies wrongdoing, and no court or regulator has ruled on the claim. Whether OpenAI's Alibaba suspension is permanent or partial is undisclosed. Baidu declined to comment; Tencent and Alibaba did not respond to the FT's questions. We follow wire language: the sales are legal today; whether Congress rewrites the rules is live policy, not settled law.

Convina's view: Washington built an export regime for atoms and left APIs to geography. That was a policy choice, not a loophole discovered by accident — and July 10 made it impossible to pretend otherwise. OpenAI's Singapore lab and Alibaba's Singapore API receipts are the same story: frontier AI follows the path of least regulatory resistance. Anthropic drew the hard line because it got robbed at industrial scale; OpenAI and Google kept selling because nationality "alone" shouldn't decide access — right up until distillation forced a suspension. Congress will now face the same question it faced on chips: close the subsidiary gap, or accept that the 1260H list is a procurement memo, not a technology embargo. The blacklist stops at Changi until Washington decides it doesn't.