A new wave of powerful, low-cost Chinese artificial intelligence models is reshaping the global AI race, challenging long-held assumptions about American technological supremacy and exposing deep contradictions in Washington's efforts to contain Beijing's AI ambitions. Models from Chinese developers including Moonshot AI, Alibaba, DeepSeek, and Zhipu are now benchmarked regularly alongside leading American systems from Anthropic and OpenAI — and in many cases perform at comparable levels at a fraction of the cost.

The competitive threat is compounded by a technique known as distillation, which has rapidly become one of the most politically charged concepts in AI and geopolitics. Distillation involves using the outputs of a large, capable "teacher" model to train a smaller "student" model, allowing developers to replicate much of the original model's performance at significantly lower computational cost. US officials and AI executives have accused Chinese firms of using distillation to close the capability gap by training off American frontier models — a practice some in Silicon Valley are calling little short of intellectual property theft. Yet the accusation sits awkwardly alongside the reality that distillation is a decade-old technique used by virtually every major laboratory worldwide, including OpenAI, Anthropic, Google DeepMind, and Meta.

The contradiction was thrown into sharp relief when Thinking Machines, a high-profile AI startup founded by Mira Murati, formerly OpenAI's chief technology officer, revealed that its debut foundation model was built in part using Chinese models — specifically DeepSeek-V3 and synthetic data generated by Moonshot AI's Kimi K2.5. The disclosure came just hours after Anthropic's chief national security officer, Tarun Chhabra, had publicly flagged Chinese distillation as a national security concern. The juxtaposition illustrates a broader dynamic: while American firms seek to prevent unauthorised extraction of their frontier capabilities, they are simultaneously incorporating advances from China's rapidly improving open-weight ecosystem. American companies including Airbnb are also reported to be relying on Chinese models in their operations.

The entanglement runs deeper still. Apple has received approval from China's Cyberspace Administration to launch its Apple Intelligence service in China using models from Alibaba and Baidu as core components — both of which have been designated by the US Department of Defense as Chinese military-affiliated companies. The arrangement underscores how thoroughly Chinese frontier models have become embedded in global commercial infrastructure, moving well beyond domestic cloud services to power hundreds of millions of devices in the world's largest smartphone market.

The policy implications are significant. The Trump administration is actively debating responses, including placing Chinese AI companies on trade blacklists, even as it simultaneously announces plans to redirect around $200 billion in federal research funding toward individual scientists with AI at the centre of the effort. Voices within Silicon Valley are divided: some executives argue for restricting Chinese models on security and fairness grounds, while others contend that open competition and American-developed open models are the better answer. Analysts caution that export controls, while effective at restricting advanced semiconductors, are far less suited to limiting the spread of published research, open-weight models, and widely shared engineering techniques. As innovation in frontier AI becomes increasingly reciprocal, the deeper question may be less about building walls and more about remaining the most attractive place to generate the next generation of ideas.