OpenAI and Google Sold AI to Pentagon-Blacklisted Firms

 We'll say the quiet part first: if the rule is "Chinese companies can't touch frontier American AI," and the actual practice is "Chinese companies can touch frontier American AI as long as the subsidiary address says Singapore," that's not a policy. That's a geography loophole with a national security label stapled to it. 

Key Points

• The Financial Times reported that OpenAI and Google have supplied AI services to Singapore-based units of Alibaba, Baidu, and Tencent, despite those companies' parents sitting on the Pentagon's 1260H blacklist for alleged military ties.
• The arrangement is legal, because current U.S. rules don't broadly stop Chinese-headquartered firms from using American AI models outside mainland China.
• OpenAI suspended API access last month for users linked to Alibaba after detecting "distillation," where a competitor uses a model's outputs to train its own, then reported the activity to the U.S. government.
• Anthropic has taken the opposite approach, banning Chinese companies and their overseas entities outright from its frontier models and pushing for broader export controls.
• Meanwhile, Anthropic's own Fable 5 and Mythos 5 models got suspended in June under U.S. export controls before being restored weeks later, a whiplash that landed on a company already advocating for tighter restrictions.

What's Going On?

The Financial Times reported that both OpenAI and Google confirmed serving Singapore affiliates of Alibaba, Baidu, and Tencent, three companies whose parent organizations the Pentagon has flagged for alleged ties to China's military. Nothing about this violated current law. U.S. export rules restrict access from mainland China specifically, not access by Chinese-owned entities operating from a jurisdiction with better optics. Google said as much, noting its products remain available in Singapore and Hong Kong under existing policy, while admitting that "geographical restrictions alone cannot completely prevent sophisticated users from bypassing controls." That's a company saying, on record, that its own compliance framework is more suggestion than wall.

OpenAI, to its credit, caught something concrete: it suspended API access for Alibaba-linked users after detecting distillation, the practice of training a competing model on another model's outputs, and reported it to the government. That's a real enforcement action against a real technique. It's also happening inside the same company that's still openly serving other Chinese-linked entities through the Singapore side door.

The Anthropic Contrast, and Its Own Bad Timing

Anthropic has drawn a harder line, blocking Chinese companies and their overseas entities entirely and lobbying for export controls with real teeth. That's a defensible, consistent position. It's also the same company whose Fable 5 and Mythos 5 models were suspended in June to comply with Commerce Department export controls, only to have those controls lifted less than three weeks later. We're not accusing anyone of hypocrisy. We're pointing out that the regulatory apparatus governing all of this moved fast enough to disrupt its own strictest advocate before anyone had time to explain the rule to a customer.

That's the actual story here, not which lab is more virtuous. It's that the rules are being applied unevenly, reversed abruptly, and routed around through subsidiary addresses, all under the same umbrella of "national security." We'd note, in fairness, that export control policy genuinely is difficult to draft cleanly when the underlying technology moves this fast and jurisdictional workarounds are trivial to construct. That's a real problem, not manufactured. But a hard problem doesn't excuse a policy that's this easy to route around and this quick to whiplash the companies trying to comply in good faith.

What This Means If You're Building on Any of These Models

If your growth strategy depends on a specific model's availability, this should worry you more than the geopolitics does. Access has already been suspended and restored once this year for reasons entirely outside any customer's control. Build flexibility into your AI marketing stack now, because the rules governing who can use what are being written in real time, inconsistently, by people who clearly haven't agreed on what they're trying to prevent.

The Financial Times broke the original reporting, and it's worth reading directly if you want the full picture before drawing your own conclusions.

If you want an AI strategy that doesn't assume today's access rules will still apply next quarter, that's a conversation worth having with us.