America needs to stop getting shocked by Chinese AI

Last week, two Chinese AI companies—Moonshot AI and Alibaba—released new models, Kimi K3 and Qwen3.8 preview, claiming to compete with the best systems from OpenAI and Anthropic. Markets reacted immediately: stocks fluctuated, commentators said Silicon Valley was 'shaken', and policymakers again used terms like 'arms race' and 'wake-up call'. The Associated Press said the Chinese models 'surprised US tech circles', while Bloomberg described them as 'unexpected breakthroughs' causing market turmoil. Xprize founder Peter Diamandis even called it America's 'AI Sputnik moment'.

However, this 'surprise' itself is what's truly surprising. For years, warnings about China catching up in AI have been loud, yet when the catch-up seems to become reality, the world shows shock.

Moonshot AI claims Kimi K3 is second only to OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 in performance, priced at $15 per million output tokens, compared to about $30 for GPT-5.6 Sol and $50 for Fable 5. Alibaba's Qwen3.8 claims to be 'second only to Fable 5'. Both companies plan to open-source their models, allowing developers to download, use, and modify them, in stark contrast to the closed proprietary approaches of leading US AI labs like OpenAI, Anthropic, and Google.

Notably, token prices are not directly comparable; more expensive models may generate better responses with fewer tokens, and companies often subsidize inference costs to attract customers. Thus, cheaper does not automatically mean better or lower overall cost. But with the performance gap narrowing, Chinese models have a clear price advantage, which has already prompted some US startups to switch to cheaper Chinese models.

OpenAI and Anthropic are preparing for IPOs that could be worth up to a trillion dollars, with valuations partly based on expectations of global AI market dominance. Capable Chinese models challenge this assumption, potentially attracting customers, squeezing profits, and undermining the growth expectations that support these valuations. Tech stocks account for a large portion of the US market, and recent growth has been tied to expectations of sustained AI demand. If Chinese labs capture some demand or demonstrate lower-cost production and operation models, investors will question the rationale for massive infrastructure investments.

On security, highly capable open-source Chinese models, even if behind US frontiers, could make advanced AI systems more widely available, especially when US companies restrict access or strengthen safety measures. When the US government asked Anthropic to restrict access to its latest model, cybersecurity leaders warned that this would make it harder for defenders to respond.

Beijing has been actively supporting domestic AI efforts, including incentives and funding for innovation, and cracking down on companies trying to cut ties with China. Meanwhile, Washington's AI strategy swings between over-intervention (making allies question US reliability) and laissez-faire (assuming the market will self-regulate). Such a strategy is hard to sustain against a competitor ready to use state power to support a single technological goal.

The controversy over whether Chinese companies use US models to train their own remains unresolved. US companies accuse Chinese companies of using methods like distillation to boost performance at very low cost. If true, this could explain how Chinese models achieve near-parity performance at lower cost.