The real AI race has shifted underground — and China just proved it's not playing catch-up anymore.
This week brought two signals from China's AI ecosystem that deserve closer attention than they've gotten. 趋境科技 launched ATaaS, positioning itself as a "token factory" capable of processing trillions of tokens daily. That phrasing is deliberate. The company is selling inference infrastructure the way utilities sell electricity — commoditized, abundant, cheap. Meanwhile, a Chinese AI lab released GLM-5.1, a model whose coding benchmarks jumped nearly 10 points over its predecessor, approaching frontier-class performance. Its coding subscription plan sold out within hours of launch.
Separately, these look like routine product releases. Together, they reveal something different: China isn't just building frontier models — it's building the entire stack beneath them.
For developers, the immediate impact is concrete. GLM-5.1 gives them a coding assistant that performs near frontier levels without the waitlist or price tag of Western alternatives. ATaaS gives them the infrastructure access to run AI workloads at costs that make new categories of applications economically viable. The combination — strong model plus affordable, high-throughput infrastructure — is what developers actually need. Neither piece alone delivers the full value.
This vertical integration matters beyond individual product advantages. The traditional Western AI development model built infrastructure first, then trained models on top. NVIDIA built chips, cloud providers built data centers, and frontier labs trained increasingly capable models. This sequencing worked because compute was expensive and abundant in the West. The infrastructure came first.
China's approach is compressing that timeline into months. The result is a different competitive dynamic: infrastructure and model capability are becoming a single vertically integrated system. Traditional Western competitive advantages — where NVIDIA supplied chips, cloud providers built data centers, and frontier labs trained models — can be systematically undercut by a single company owning the entire stack.
This isn't speculative. CAE member Zheng Weimin framed ATaaS's token processing capability as part of a broader shift in AI infrastructure strategy. The sellout of GLM-5.1's coding subscription demonstrates real demand from developers who want both pieces — not just a powerful model, but affordable and fast access to it.
The implications for Western AI companies are stark. They can either build the same integrated stack or cede entire layers of the value chain. The latter is not a viable long-term strategy given how fast this is moving. The token factory is running. The frontier model is shipping. The stack is being built.