US AI access to China groups draws scrutiny

OpenAI and Google have supplied advanced artificial intelligence services to overseas subsidiaries of major Chinese technology companies identified by the Pentagon as having links to China’s military, exposing a gap in Washington’s technology controls.

The services were provided through entities in Singapore connected to Alibaba, Baidu and Tencent. Transactions involving the subsidiaries were permitted under existing US rules, even though their parent groups appear on a Department of Defense list of Chinese military companies.

The arrangements have intensified debate over whether restrictions designed around semiconductors and computing infrastructure can adequately control access to powerful AI systems delivered through cloud platforms. Unlike advanced chips, software models can be accessed remotely and distributed through corporate structures spanning several jurisdictions.

OpenAI and Google said their services are governed by safeguards intended to prevent prohibited activity and detect misuse. Both companies maintain restrictions on direct access from China, where many leading US AI products are unavailable. However, multinational groups can operate subsidiaries in supported markets, creating uncertainty over when the identity and ownership of a customer should determine eligibility.

OpenAI has suspended access for some users linked to Alibaba after detecting activity that raised concerns about model distillation. The technique involves generating large volumes of answers from an advanced model and using those outputs to train another system. It can enable developers to reproduce aspects of a rival model’s capabilities without bearing the full cost of its development.

The company has previously accused China-based developers of attempting to extract knowledge from its systems through distillation. US technology executives and officials regard the practice as a growing threat because frontier models require vast investments in computing power, data, engineering and electricity.

Google has also acknowledged that restrictions can be circumvented, although it says it employs monitoring systems, usage policies and technical controls to identify abusive behaviour. Its AI services are delivered through a range of consumer and enterprise products, including the Gemini model family and the Vertex AI cloud platform.

The Pentagon’s Section 1260H list identifies companies that Washington believes contribute to China’s military-civil fusion strategy. The designation is not the same as placement on the Commerce Department’s Entity List and does not automatically prohibit all commercial dealings.

That distinction has become central to the controversy. Alibaba, Baidu and Tencent are among China’s largest technology groups, with extensive interests in cloud computing, digital payments, online platforms, autonomous systems and AI research. Their global operations include separately incorporated companies outside mainland China.

The Defense Department’s June 2026 assessment said Baidu contributed to China’s defence industrial base through affiliations with government technology bodies. Beijing and affected companies have repeatedly rejected US allegations that commercial enterprises are improperly linked to military activity.

Washington has imposed increasingly strict limits on exports of high-performance processors and semiconductor manufacturing equipment to China. Nvidia and other chipmakers must obtain licences before selling certain advanced products, while restrictions also cover equipment used to manufacture cutting-edge chips.

Controls on access to trained AI models remain less comprehensive. Policymakers face difficulties defining which systems should be restricted, measuring their capabilities and preventing customers from reaching them through cloud accounts registered in third countries.

Security advocates argue that access to leading US models can accelerate research by China-based companies, improve military planning tools and support offensive cyber operations. They have called for ownership-based rules that would apply to subsidiaries controlled by organisations from restricted jurisdictions.

Industry representatives warn that broad controls could damage US cloud providers, encourage customers to adopt competing models and speed the development of independent AI ecosystems outside the United States. They also argue that centrally hosted models offer greater visibility than downloadable software because providers can monitor prompts, block accounts and update safeguards.

Anthropic has adopted a stricter position than several competitors. The Claude developer bars companies controlled from China and other unsupported regions from using its services, including through overseas subsidiaries. It says corporate ownership can create security risks regardless of where an entity is registered.

The dispute is unfolding as China considers tighter controls on foreign access to its own advanced models. Chinese systems from DeepSeek, Alibaba, Moonshot AI and Zhipu AI have gained users abroad because of their competitive performance and lower operating costs.