China's AI models narrow gap with US as tensions escalate over IP theft claims

The gap between American and Chinese artificial intelligence has collapsed in a way that would have seemed impossible just a few years ago. Where China once lagged by five years, it now trails by only months—a shift that has alarmed Washington and set off a new round of accusations and threats.

The story of how this happened is not one of matching American spending. OpenAI and Anthropic have poured hundreds of billions into pushing the frontier of what AI can do, building closed systems that they control and monetize through subscriptions and licensing deals. China took a different route. Blocked by US restrictions on the advanced chips needed for cutting-edge AI work, Chinese companies instead built dozens of open-source models—systems that work well enough for most purposes and that anyone can download, run on their own servers, and customize. DeepSeek, Alibaba's Qwen, and Xiaomi's MiMo are now embedded across Chinese businesses, government agencies, and consumer devices. The approach has proven both practical and scalable.

The moment that crystallized the shift came last week when Moonshot AI, a Chinese startup, unveiled Kimi K3. On standard benchmarks that measure AI capability, Kimi K3 matches or even beats Anthropic's Claude Fable 5—one of the world's most advanced models—while using far less computing power and costing a fraction as much. Alvin W. Graylin, a technology specialist, described what this means in plain terms: the model "commoditizes frontier models." For most customers, he explained, the tiny difference in performance is not worth paying ten to twenty times more. The economics have shifted.

Washington's response has been swift and sharp. President Trump has threatened to block American access to Chinese AI models entirely. Treasury Secretary Scott Bessent warned of potential sanctions. The accusations center on a practice called distilling—using queries to extract advanced reasoning from proprietary American models in order to train Chinese systems. Anthropic and OpenAI say this has happened at industrial scale. Beijing has rejected the claims as baseless and called the sanctions threats "technological bullying" designed to slow China's rise.

But the real driver of China's progress may be less about theft and more about strategy. Open-source models allow Chinese labs to enlist global developers in improving their systems. The US chip restrictions forced Chinese companies to be more creative and efficient with the resources they had. Giulia Neaher, a research analyst at the Stimson Center in Washington, argues that distilling played only a limited role in Kimi's success—the real work came from talent, infrastructure, and resources poured into building the model itself. She compares distilling to teaching students to pass tests without deep understanding: you might get good scores, but the actual learning may not be there.

Yet the stakes of this competition extend far beyond market share or pricing. Whoever controls the most advanced AI systems is likely to dominate militarily, economically, and technologically for years to come. Frontier models can accelerate breakthroughs in science and medicine. Washington still holds the edge in the most advanced systems, but the margin is narrowing, and Neaher predicts the US will abandon its hands-off approach to AI policy and escalate its response.

A new rivalry is already taking shape—this one over which nation's AI companies can raise the most capital. American firms have been valued at close to a trillion dollars, and both Anthropic and OpenAI are preparing IPOs later this year. On the Chinese side, DeepSeek and Moonshot AI are also seeking public listings, with Moonshot planning a Hong Kong offering before early next year. US institutional investors have quietly increased their bets on Chinese AI firms, betting they can avoid an American AI bubble while positioning for China's momentum. Graylin sees the bubble clearly: "The Chinese are building comparable models for one-tenth the cost and priced at one-tenth of US frontier." The correction, he warns, could come within three to twelve months.