ft.com web signal

OpenAI, Google sold AI to Singapore arms of Alibaba, Baidu, Tencent

4 sources tracking this story

TL;DR

  • OpenAI suspended Alibaba-linked API access after detecting suspected model distillation involving what reports describe as 25,000 fake accounts.
  • Alibaba has filed a legal challenge to its own Pentagon 1260H designation, complicating any enforcement action built on that list.
  • Export control experts argue for shifting from geography-based to capability-based controls, calling Singapore residency an ineffective security threshold.

The interesting thing about the Financial Times' reporting that OpenAI and Google have been selling AI services to Singapore-based subsidiaries of Alibaba, Baidu and Tencent is not that it happened. It is that both companies confirmed it on the record, and that under current US law the sales are legal.

The three parents all appear on the Pentagon's so-called 1260H list, which the FT describes as companies accused by Washington of having links to the People's Liberation Army. But US export controls target specific entities and locations rather than the technology itself, so a Singapore-incorporated subsidiary is on paper a Singaporean company and can enter contracts the Beijing parent cannot. OpenAI told the FT it blocks direct access to its models from China but allows some Chinese-owned companies to use its services in jurisdictions where safeguards can be enforced. Google said its AI services remain available in Hong Kong and Singapore subject to usage policies, including restrictions against distillation, while conceding that geographic restrictions alone are not enough because sophisticated users can circumvent location-based controls.

The reason this matters is that the arrangement is the kind of thing Washington could collapse very quickly. OpenAI already suspended API access for Alibaba-affiliated users last month over suspected distillation, the practice of using the outputs of advanced AI models to improve competing systems. Anthropic has gone further and now prohibits Chinese companies and foreign entities owned by them from using its advanced models, and it separately warned Congress that Alibaba allegedly used thousands of accounts to generate millions of exchanges with its Claude AI model in violation of its terms of service. If that pattern hardens into evidence rather than allegation, the political case for treating frontier model access the way chip exports are treated becomes hard to argue against.

The honest caveat is that the reporting is largely about what is legal, not what is proven. The distillation concerns against Alibaba users are described as suspected, the 1260H designations are Pentagon judgments rather than court findings, and neither OpenAI nor Google disclosed how much revenue their Chinese-owned Singapore customers actually represent. What the piece does not give you is a Commerce Department timeline. The forward-looking read is who benefits if that timeline arrives sooner rather than later: Anthropic, which has already priced in the security-hawk position, and whoever ends up building the customer-ownership audit tooling the frontier labs will suddenly need.

What others are reporting

Coverage cluster as of 24h after publish

  1. The Next Web Read →

    Connects to Nvidia chip export gap as precedent; names Alibaba's 25,000 fake accounts; cites CFR expert McGuire on capability-based vs geography-based controls.

    We don't think nationality alone should decide access.
  2. The National Read →

    Leads with company countermeasures defense; includes Mark MacCarty's counterargument framing export restrictions as protectionist measures rather than security policy.

    strong terms of use and protections, including disabling accounts and blocking access to AI services where necessary
  3. Il Sole 24 Ore Read →

    Reports Alibaba's active legal challenge to its 1260H designation; frames story as U.S. strategic policy debate with Anthropic's enforcement stance as the contrasting case.

    The matter has also been reported to the US government.

Shared on Bluesky by 1 AI expert

  • Stefanie Hane @stefihane.bsky.social amplified

    @lemmk.bsky.social

    #OpenAI mastering the art of the squeal: • confirms it sells access to some blacklisted Chinese customers; • claims it would “rather see more of the world using AI shaped by democratic values”; • adds that “we don’t t…

    View on Bluesky →