Usman Salis
The AI race has evolved beyond model size and chip speed. Today, it's fundamentally about controlling the underlying technology stack. Recent moves in China reveal a deliberate plan to build homegrown AI infrastructure independent of U.S. technology.
Phase One: FP8 as the Foundation
A recent analysis by Poe Zhao showed how developments around FP8 precision and TileLang aren't random innovations but calculated steps toward a China-controlled AI ecosystem.
Last August, a brief mention of FP8 precision drove Cambricon shares up 14%. This wasn't market noise—FP8 is becoming the standard data format across China's AI sector. The format cuts memory usage and power consumption while speeding up training and inference. Companies like Cambricon are building FP8 directly into their chip designs, creating tighter software-hardware integration. By standardizing on FP8, fragmented chipmakers now share a common optimization target, reducing wasted effort and accelerating collective progress. FP8 essentially provides the technical foundation for everything that follows.
DeepSeek launched TileLang last week—a Python-style language built for domestic AI accelerators. Immediate backing from Huawei, Cambricon, and Hygon revealed how coordinated the launch was. TileLang matters because it removes friction for developers who can now use familiar syntax instead of learning proprietary frameworks. It creates portability across different Chinese chips, letting the same code run on varied hardware. More importantly, a unified programming layer gives compiler builders, toolchain vendors, and software developers a common standard to rally around. Combined with FP8, TileLang forms the data format and programming layer—the core of an independent AI stack.
Market Response Shows Strategic Weight
Financial markets are treating this seriously. Cambricon's stock jumped on FP8 news as investors recognized its pivotal position. Chinese AI chipmakers saw broad gains as capital flowed toward domestic alternatives. Cambricon recently reported record profits, confirming real momentum behind this approach. Markets clearly see these developments as strategic shifts, not just technical updates.
China's chips still trail Nvidia in raw performance. DeepSeek continues releasing CUDA versions of its models, and H100 throughput remains superior. Coordination doesn't equal parity—not yet. But the blueprint targets long-term independence, not immediate supremacy. By reducing costs and achieving "good enough" performance across a domestic stack, China aims to gradually shift competitive dynamics in its favor.
Usman Salis
