American Models Open
Chinese Military Researchers Used Western AI Models to Train Defense Systems
A Reuters investigation published this week found that Chinese military researchers used OpenAI and Anthropic models to help train defense-oriented artificial intelligence — apparently accessing the systems through standard API channels or similar means, fine-tuning outputs for military applications, and incorporating results into their own training pipelines. Both companies maintain terms of service prohibiting military use, but the episode illustrates a fundamental technical limit: a powerful general-purpose language model capable of reasoning about logistics, strategy, and planning cannot reliably be prevented from being used for those purposes by determined actors.
The timing coincided with Chinese companies releasing open-weight models that Silicon Valley observers assessed as genuine rivals to frontier American systems. Moonshot AI and Alibaba both released powerful models this week. The significance of 'open-weight' releases is that the model weights themselves are publicly available for download, local operation, and fine-tuning — no API gate, no terms of service to navigate. Silicon Valley companies and researchers are now lobbying Washington for an American open-source response, arguing that if China will release powerful open models regardless, the United States should ensure the best globally available open models are American-origin.
Meta's Mark Zuckerberg offered a nuanced counterpoint, saying this week that Meta will not rely on open-source AI for its own core products — despite Meta being one of the biggest open-source AI contributors through its Llama series. The implication is that for mission-critical applications, even the most prominent open-source advocates want control over their full stack. President Trump, separately, called AI 'bigger than the internet' and warned against overregulation, framing domestic AI policy explicitly as a competition with China where excessive restrictions could cost the U.S. its lead.
The infrastructure ambition underlying this competition came into sharp focus with the announcement that DeepSeek, the Hangzhou-based AI startup, is planning a one-gigawatt data center in Inner Mongolia's Ulanqab region. American data centers typically operate in the tens to hundreds of megawatts range. For comparison, Anthropic is currently attempting to finance a $15 billion data center backed by bank loans supported by Google — a facility that the DeepSeek announcement would reportedly dwarf. U.S. District Judge Rita Lin added another dimension to the Anthropic story, stating that the Pentagon's legal justification for blacklisting the company had gotten 'worse' over the course of legal proceedings — unusually pointed language from a federal judge scrutinizing a government national security designation.