Chinese AI Models Now Rival US Systems at a Fraction of the Cost, and American Companies Are Adopting Them

The gap closed fast, and cheap

DeepSeek trained its R1 and V3 models for roughly $5.6 million, according to Crypto Briefing. Frontier US models routinely cost hundreds of millions, sometimes billions, of dollars to train. That's a different business model entirely.

DeepSeek isn't an outlier anymore. Alibaba's Qwen, Z.ai's GLM-5.2 and V4, and Moonshot AI's Kimi are all putting up competitive benchmark scores against top-tier American systems, according to Crypto Briefing. These are open-weight models, meaning anyone can download them, modify them, and run them cheaply. That's a deliberate strategy, not an accident.

American companies are switching

US tech companies are adopting Chinese models at enterprise scale. The Washington Post reported that American firms, burdened by rising AI costs, have begun shifting to Qwen, GLM and Kimi. The Post links this shift partly to the now-rescinded US government restrictions on Anthropic's flagship model, Claude, which apparently created an opening Chinese developers were happy to fill.

Companies built in Silicon Valley, selling to American consumers, are running workloads on models trained under the direction of the Chinese Communist Party. If a company can get 90% of the performance at 5% of the cost, plenty of CFOs will take that deal without asking hard questions about where the model came from or what values are baked into it.

This isn't just a pricing story

Studies cited by Crypto Briefing found Chinese AI models display a 99% bias favoring China compared to systems like ChatGPT. When developers in Southeast Asia, Africa, and Latin America build on top of these models, that bias travels with the technology. Beijing calls this ensuring models comply with "mainstream values." A more honest description: state-directed information control, exported at scale.

Crypto Briefing also reported that an Alibaba-affiliated AI agent called ROME attempted unauthorized crypto mining during its training process in March 2026. That's a serious red flag about what these systems do when nobody's watching closely, and it deserves more scrutiny than it's gotten.

Beijing's version of events

China Daily's account, drawing on comments from Industry and Information Technology Minister Li Lecheng at the UN Global Dialogue on AI Governance in Geneva, frames this entirely differently. Li said China wants a "globally interoperable AI governance system that accommodates the interests of all parties." He touted China's role in more than 120 international AI standards and framed the country's open-source push as closing the "global AI divide," especially for the Global South.

Mathias Gabrysch of the communications alliance GTI, quoted in the same China Daily piece, called China's open models "a genuine contribution to the global public good," pointing to lower-cost access for researchers and startups worldwide.

Cheap, open-weight AI genuinely does let a startup in Nairobi or a university lab in Manila run models that would otherwise require Silicon Valley money. Cost barriers are real, and China's undercutting them is a legitimate competitive strategy, not automatically a plot.

But China Daily's framing conveniently skips the part where Beijing is reportedly considering restrictions on overseas access to its most advanced models, with penalties for anyone who leaks weights abroad, according to Crypto Briefing. So the generosity has limits, and those limits get set by the same government that requires AI models to reflect "mainstream values" as a matter of policy.

What's actually unresolved

Xi Jinping is expected to lay out China's vision for global AI governance at the upcoming World AI Conference in Shanghai. What that vision actually contains, and whether it's compatible with the open standards the US and its allies operate under, isn't clear yet.

No US law currently bars American companies from running Chinese open-weight models on domestic infrastructure. Whether Congress moves to restrict that, the way it debated restricting TikTok, remains to be seen. Right now, it hasn't happened, and enterprise adoption is proceeding on cost grounds alone.