← Front Page
AI Daily
Export Controls • Sunday, 23 August 2026

America Banned China From Its Best AI. Then Sold It Through Singapore.

By AI Daily Editorial • Sunday, 23 August 2026

The United States has spent three years building a wall around its best artificial intelligence, denying China both the advanced chips to train frontier models and the frontier models themselves. This week the Financial Times described a door in that wall, and it opens onto Singapore. According to the report, both OpenAI and Google acknowledged supplying advanced AI services to Singapore-registered subsidiaries of Alibaba, Tencent and Baidu, three companies the Pentagon has placed on a list of firms it says are tied to China's military.

The mechanism is mundane, which is what makes it hard to close. A company incorporated in Singapore can legally sign a contract with an American AI provider even when its parent sits on a US blacklist. OpenAI said it blocks direct access from China but allows some Chinese-owned firms to use its services in jurisdictions "where safeguards can be enforced." Google said its tools are available in markets including Singapore and Hong Kong, then conceded the point that matters: "geographical restrictions alone are insufficient" to stop a determined user from routing around them.

Hovering over all of this is the fear of distillation, the practice of using a strong model's outputs to train a cheaper competitor. OpenAI said it suspended access for Alibaba-linked accounts last month after detecting suspected distillation, and reported the activity to the US government. The worry is not abstract: in Congressional testimony this year, Anthropic alleged that Alibaba had used tens of thousands of fraudulent accounts to pull more than 28 million exchanges out of its Claude models, and it has levelled similar accusations at DeepSeek, Moonshot and MiniMax.

What the episode really exposes is that the American labs do not agree on where the wall should stand. Anthropic has taken the hard line, barring China-headquartered companies and the foreign subsidiaries they own from its most capable models outright. OpenAI and Google have chosen softer, conditional access, available in approved jurisdictions with safeguards attached. Behind the scenes the largest labs have been comparing notes through an industry body called the Frontier Model Forum, sharing signals about suspected extraction attempts. It amounts to cooperation on detection and competition on who gets let in.

The uncomfortable backdrop is that the political machinery meant to referee all this has stalled. Legislation to tighten technology export controls is stuck in Washington amid a broader foreign-policy deadlock, a pause that recently handed the chipmaking-equipment giant ASML a reprieve on its own China restrictions. China, meanwhile, has answered every chip curb with export controls of its own on the rare-earth minerals the West needs. A wall works only if everyone agrees where it goes and is willing to patch the gaps. Right now the frontier is guarded by a fence with a Singapore-shaped hole in it, and the people who built it are still arguing about the blueprints.

Sources