US AI companies OpenAI and Google have been supplying advanced AI models to Singapore-based subsidiaries of Alibaba, Baidu, and Tencent — three Chinese tech giants on the Pentagon’s 1260H blacklist for alleged military ties. The sales are completely legal under current US export rules. That’s the problem.
🔍 THE BOTTOM LINE
The US government has spent years building a chip export control regime to slow China’s AI advancement. But it never built the equivalent for AI models. OpenAI and Google confirmed to the Financial Times that they sell API access to Singapore subsidiaries of blacklisted Chinese firms, and nobody in Washington has the authority to stop them. The regulatory gap isn’t an oversight — it’s a structural choice, and one that’s about to come under intense scrutiny.
The Singapore Loophole
According to the Financial Times investigation, both OpenAI and Google confirmed they provide AI services to subsidiaries linked to Alibaba, Baidu, and Tencent in Singapore. The US government’s 1260H list designates these parent companies as having ties to China’s military. But the blacklist restricts specific entities, not their overseas subsidiaries operating in different jurisdictions.
OpenAI said it does not allow its models to be accessed directly in China but permits companies with Chinese ownership to use its tools in countries where “safeguards can be enforced and potential misuse monitored.” The company framed it as a values choice: “We would rather see more of the world using AI shaped by democratic values than AI controlled by autocratic governments,” OpenAI said, while adding that nationality alone should not determine access.
Google said its AI services are available in Hong Kong and Singapore subject to usage policies, including restrictions against distillation — the process of using one model’s outputs to train another. But Google also acknowledged that geographic restrictions alone aren’t enough, because sophisticated actors can bypass them.
The Sri Lanka Guardian’s reporting on the FT investigation notes that the use of Singapore as a neutral hub highlights the complexities and potential loopholes in global tech regulations.
The Distillation Problem
This is where the story moves from a regulatory curiosity to a strategic threat. OpenAI confirmed it suspended access last month for Alibaba-affiliated users of its API over concerns about possible distillation — using GPT outputs to improve a competitor’s model. The company said it reported the suspected activity to the US government.
Distillation is the exact mechanism that OpenAI, Anthropic, and Google previously united against when they called for stronger protections against Chinese competitors harvesting their model outputs. The irony is that the companies themselves are the ones selling the access that enables it. It’s also the same pattern the NSA exploited when it used Anthropic’s Claude despite Pentagon blacklist concerns — the blacklist is a list, not a firewall.
The concern is straightforward: if a blacklisted Chinese firm can access GPT-5.6 or Gemini Ultra through a Singapore subsidiary, they can generate millions of API calls to produce training data for their own models. The API terms say you can’t do this. But enforcement requires detection, and detection across a Singapore corporate structure is non-trivial.
Why Export Controls Cover Chips, Not Models
The US export control apparatus was built for physical goods — semiconductors, manufacturing equipment, dual-use components. The Commerce Department’s Entity List and the Pentagon’s 1260H list both restrict specific entities from receiving restricted items. But AI models accessed via API from a server in Singapore are not physical goods. They’re not even software exports in the traditional sense — they’re remote inference calls.
This is the structural gap. The chip restriction regime that restricts Nvidia H200 sales to China is enforceable because chips are physical objects that pass through customs. API access has no customs checkpoint. The only enforcement mechanism is the provider’s own terms of service and usage monitoring — which is exactly what both companies admitted is insufficient.
Alibaba has separately challenged its inclusion on the 1260H blacklist, asking a US court to remove the designation and describing it as “arbitrary and capricious.” Baidu and Tencent did not provide immediate comments.
NZ Angle
New Zealand sits in an interesting position here. As a Five Eyes partner, NZ’s intelligence apparatus benefits from US export controls slowing Chinese AI development. But as a trade-dependent Pacific nation, NZ has no domestic AI industry to protect and every incentive to access the best available models — from any provider. The Singapore loophole shows that AI model access is already globally fluid. If the US closes this gap with broader restrictions on API access for Chinese-owned entities, it sets a precedent that could extend to other jurisdictions. For NZ businesses building on OpenAI or Google APIs, the question is whether their provider might one day cut off access based on government policy rather than commercial terms.
❓ FAQ
Is what OpenAI and Google are doing illegal? No. Both companies confirmed the sales are legal under current US rules. The 1260H blacklist restricts specific named entities, not their overseas subsidiaries. There is no broad prohibition on selling AI model access to Chinese-owned companies operating outside China.
What is the 1260H list? It’s the Pentagon’s list of Chinese companies accused of having ties to the military. Being on the list doesn’t automatically trigger export restrictions — it triggers scrutiny and can affect contracts with the US government.
What is distillation and why does it matter? Distillation is using one AI model’s outputs to train another model. It’s the cheapest way to improve a weaker model. If a Chinese firm can call GPT-5.6 millions of times via a Singapore subsidiary, they can generate high-quality training data for free — undercutting the billions US firms spent training the original.
🔍 THE BOTTOM LINE
The Singapore loophole is the missing chapter in the US-China AI story. Chip export controls get the headlines, but the models themselves — the actual intelligence — flow freely through API calls that no regulatory framework currently governs. OpenAI and Google are doing nothing illegal. The system is doing exactly what it was designed to do: nothing. Whether that’s a feature or a bug depends on whether you think AI models are more like software (which anyone can buy) or more like enriched uranium (which nobody should sell to the other side).
