DeepSeek says it has partnered with Huawei to develop programming tools for Huawei's Ascend chips, including TileLang, an open-source CUDA alternative
Reuters
Context & Ripple Effects
DeepSeek’s Ascend work has moved from a constrained deployment strategy—after reported training issues, it had planned to use Huawei chips for smaller models while retaining Nvidia for its largest training runs—to a software effort aimed at the programming layer itself. The earlier training difficulties on Ascend made tooling and operator compatibility a material constraint, not merely a developer convenience.
Huawei had already said its Ascend 950 supernode would fully support DeepSeek V4, while DeepSeek was reportedly planning a large Ascend 950DT cluster. TileLang and associated libraries connect that hardware commitment to a reusable software stack.
First-order effects
DeepSeek and Huawei give Ascend developers an open-source programming path through TileLang and related compute and communication libraries, reducing dependence on CUDA-specific tooling for supported workloads.
Huawei gains software co-development from a prominent model builder as it positions Ascend 950 systems for DeepSeek workloads.
Second-order effects
DeepSeek’s planned Ascend deployments have a clearer route from hardware procurement to usable training and inference workloads, since the partnership targets the software layer that previously complicated Ascend training.
Nvidia faces a more direct competitive challenge in the developer stack: an open-source alternative can make Ascend adoption less dependent on each customer building its own ports and kernels.
Third-order effects
If DeepSeek’s tools gain adoption beyond its own models, AI accelerator competition shifts toward integrated hardware, compilers, libraries, and model-operator support rather than chip performance alone.
The development reinforces a second-source compute strategy in which AI labs split workloads across platforms and invest in portability rather than treating CUDA as the sole deployment target.
The trend:AI infrastructure is becoming a contest over full-stack software compatibility, as model developers help alternative accelerator vendors narrow the CUDA ecosystem gap.
This might be a misunderstanding But it'd be very cool of DeepSeek to release a - single-GPU, screaming fast 1M context model - with extremely intense post-training, agent teams etc - with local-first skills - “trained on Huawei Ascend
DeepSeek and Huawei are building an alternative to Nvidia's CUDA as China ramps up its domestic AI chip industry. Via Bloomberg: DeepSeek has released an open-source toolkit for Huawei's Ascend accelerators, including TileLang support optimized for Ascend 950. The tools simplify …
DeepSeek just ported TileLang, the language behind most of its V4 training operators, to Huawei's AI chips. On September 30, it open-sourced the Ascend version, plus compute and communication libraries for Huawei's Ascend chips. Each component matches one it had already released …
DeepSeek moment 2.0: the thing @JensenHuang warned about just happened. In the last few hours, @deepseek_ai shipped a set of AI infra repos with @Huawei Ascend 950 support. Not a new model. The layer under the model. TL;DR: these kernels are what let the rest of the stack actuall…
🚨DeepSeek has open-sourced its AI infrastructure stack for Huawei Ascend NPUs. The release brings several components to Ascend, corresponding to DeepSeek's existing NVIDIA GPU stack: • TileLang — a high-level DSL for writing optimized AI kernels • DeepGEMM-Ascend — optimized GEMM…
DeepSeek saves Chinese AI, again. Knowing their kernel wizardry, Ascend ecosystem is now fully viable for training. Didn't take too long. 26 months ago Liang Wenfeng said: “Someone must step onto the frontier.