DeepSeek Partners With Huawei to Attack Nvidia CUDA Software Dominance
DeepSeek open-sourced a full programming toolkit for Huawei's Ascend chips, and the target isn't Nvidia's software layer that has kept Chinese developers locked to Nvidia hardware even as domestic alternatives caught up.
DeepSeek announced on WeChat that it partnered with Huawei Technologies to build and open-source programming tools for Huawei’s Ascend AI chips, Reuters reports.
The centerpiece is TileLang, a high-level language DeepSeek says offers a simpler development model than Nvidia’s CUDA, alongside custom compute and communication libraries for the Ascend platform.
Huawei reportedly “provided full support” in developing the infrastructure, and the companies also built a 128-chip Ascend 950 supernode to prove the platform can handle serious training workloads, not just run smaller models.
The Real Target Was Never the Chip Itself
Even when Huawei’s Ascend chips approach Nvidia’s specs, developers must rewrite code built for CUDA, AI’s default programming framework for over a decade.
That massive software migration cost is precisely why Chinese tech companies still rely on Nvidia hardware, a dependency highlighted by Beijing’s evaluation of large Nvidia chip orders for domestic AI labs.
DeepSeek’s toolkit attacks that dependency directly.
Its PyTorch library now converts CUDA code into Huawei’s CANN equivalent, letting developers on Nvidia workflows migrate to Ascend hardware without learning a new system, removing the friction that makes migration practical rather than just theoretical.
DeepSeek Has Been Building Toward This for Months
DeepSeek spent 2026 shifting its infrastructure from Nvidia’s H800 to Huawei silicon amid ongoing U.S. sales blocks. The company optimized its V3.2 model for Ascend hardware, later deploying that same chip architecture across its V4.0 family.
The company is also planning to build an Inner Mongolia data center designed for 160,000 Huawei Ascend accelerators, a commitment of scale that only makes sense if DeepSeek is confident the software stack can actually support it.
Wednesday’s toolkit release completes this hardware strategy: moving from proof-of-concept training with V4 launch to year-long infrastructure expansion, and finally delivering a full open-source programming stack.
Software, Not Silicon, Is Where This Race Gets Decided
Nvidia’s dominance was never just about faster chips; CUDA’s decade-long head start created an ecosystem competitors couldn’t dislodge, slowing Beijing’s chip-substitution efforts despite advancing domestic hardware.
DeepSeek open-sourcing TileLang for free is a deliberate bet that the fastest way to build that missing ecosystem is to hand it to every Chinese developer at once rather than asking Huawei to build adoption one customer at a time.
Whether TileLang can match CUDA’s maturity remains untested at scale, but the core strategy is clear: China’s AI industry is finally targeting the software moat that chip design alone could never breach.
Source: DeepSeek partners with Huawei to develop chip programming tools
