In a nondescript office park in Beijing's Haidian district, a team of researchers at Baidu is putting the finishing touches on ERNIE 5.0, the latest version of China's most advanced language model. Down the road, teams at Zhipu AI and Moonshot AI are doing the same. In Hangzhou, Alibaba's Qwen team just released a model that outperforms GPT-4 on Chinese-language benchmarks. And in a heavily guarded facility in Shenzhen, Huawei is manufacturing the AI chips that power all of it.
China isn't just participating in the AI race. It's building a completely parallel AI ecosystem — one that doesn't depend on American technology at all.
The Forced Decoupling
This wasn't the plan. Until 2022, Chinese AI labs used NVIDIA chips, American cloud services, and the same open-source tools as everyone else. Then came the export controls. The U.S. government, recognizing that advanced AI could have military applications, banned the sale of cutting-edge chips to China.
The intended effect was to slow China down. The actual effect was to accelerate China's drive toward self-sufficiency. Necessity, as always, proved to be the mother of invention.
The DeepSeek Shock
In January 2025, a relatively unknown Chinese lab called DeepSeek released a model that stunned the Western AI world. DeepSeek-V3 and its reasoning variant R1 matched or exceeded the performance of models from OpenAI and Google — but they did it with a fraction of the computational resources. The implication was clear: China had found ways to do more with less.
DeepSeek's approach — aggressive optimization, novel architectures, and efficient training methods — challenged the assumption that AI progress required ever-larger clusters of the most advanced chips. You could build world-class AI on hardware that the U.S. hadn't banned.
The Domestic Chip Ecosystem
Huawei's Ascend 910B isn't as fast as NVIDIA's H100. But it's good enough, it's available, and it's improving rapidly. SMIC, China's leading chip fabricator, has made surprising progress despite being cut off from ASML's most advanced lithography machines. The chips aren't as small or efficient as Taiwan's TSMC can produce, but the gap is narrowing.
More importantly, Chinese AI labs have optimized their software to squeeze maximum performance from domestic hardware. When you can't get the best chips, you write better code.
The Application Advantage
Where China truly leads isn't in model benchmarks — it's in deployment. Chinese AI applications are embedded in daily life in ways that Western equivalents aren't:
- Alipay and WeChat Pay use AI for fraud detection serving over a billion users
- ByteDance's recommendation algorithms power TikTok and its Chinese counterpart Douyin
- Smart city systems in dozens of Chinese cities manage traffic, energy, and public services with AI
- AI tutoring systems serve hundreds of millions of students
- Factory automation — China deploys more industrial robots than any other country
Two Internets, Two AIs
The global internet is splitting into American and Chinese spheres, and AI is accelerating the split. Chinese models are trained on Chinese data, optimized for Chinese applications, and governed by Chinese regulations that require AI to uphold "core socialist values." Western models are trained on Western data and governed by Western norms.
Neither side fully understands what the other is building. This opacity increases the risk of miscalculation and makes AI governance — already difficult within a single country — nearly impossible at the international level.
The Stakes
The AI race between the U.S. and China isn't just a technology competition. It's a contest over which model of governance — democratic and market-driven or authoritarian and state-directed — can better harness the most powerful technology of the century. The outcome will shape geopolitics for decades to come.