Nvidia's real moat has never been only its chips — it is CUDA, the software layer that a generation of AI developers builds on. On July 18 at WAIC in Shanghai, Alibaba's T-Head chip unit moved directly at that moat, open-sourcing SAIL, the foundational software stack for its Zhenwu-series AI processors.
What was released
SAIL is the toolchain that lets developers program and deploy on T-Head's own silicon. By making it freely available the same day, Alibaba is inviting international developers to target its chips without paying the switching cost that normally keeps them inside the Nvidia ecosystem. T-Head claims programmers can adapt the SAIL stack to mainstream AI frameworks in under seven days — a pitch aimed squarely at the friction of migrating off CUDA.
The CUDA problem
CUDA's dominance is self-reinforcing: because everyone builds on it, everyone keeps building on it. Chinese chipmakers, cut off from Nvidia's most advanced hardware by export controls, cannot win developers on raw performance alone, so they are attacking the software lock-in instead — trying to make their chips easy enough to adopt that the ecosystem advantage erodes.
Not the first move
T-Head (also known as PingTouGe) joins a coordinated national effort. Huawei open-sourced its CANN compute toolkit in 2025, and Moore Threads has pushed its own CUDA-compatible path. The common thread is a bet that open, freely available software is the fastest way to build a domestic alternative to Nvidia's platform.
The catch
Open-sourcing a stack is necessary but not sufficient: developers still need the underlying chips to be available, performant and well-supported. A seven-day porting claim is only as good as the tooling behind it. Still, releasing SAIL turns Alibaba's in-house silicon program into an open platform play — and signals that the CUDA fight is now being waged in software, not just fabs.