What happened
DeepSeek published two new open-source repositories, DeepGEMM-Ascend and DeepEP-Ascend, specifically designed for Huawei's Ascend 950 series accelerators. These tools provide matrix multiplication and all-to-all communication capabilities for mixture-of-experts models, mirroring the API structure of their Nvidia CUDA counterparts to reduce developer migration friction. The release also includes updates to existing projects like TileKernels, DeepSelect, and FlashMLA for Ascend compatibility. While framed by some media as a direct challenge to Nvidia, the tools are highly specialized, requiring specific Huawei silicon and CANN 9.20 software, limiting their immediate utility to a narrow audience of Chinese AI labs and Huawei engineers.
DeepSeek has released two new open-source repositories, DeepGEMM-Ascend and DeepEP-Ascend, targeting Huawei's Ascend 950 series accelerators. DeepGEMM-Ascend handles matrix multiplication, while DeepEP-Ascend manages the all-to-all communication required for mixture-of-experts models. These are distinct from the existing CUDA-based versions, which have significantly higher adoption rates.
The release also includes Ascend-targeted updates to existing projects such as TileKernels, DeepSelect, and FlashMLA. TechRadar notes that while media outlets like Bloomberg have highlighted TileLang as a key component, it was actually open-sourced in January 2025, and its Ascend adapters were published in a separate repository in September 2025, predating this specific DeepSeek announcement.
The tools are designed to reduce switching costs by maintaining the same Python package names and API shapes as their CUDA counterparts. This allows developers to potentially port code with minimal changes, addressing the primary friction point in moving away from Nvidia's ecosystem.
However, the tools have strict hardware requirements. DeepGEMM-Ascend requires the Ascend 950 series and CANN 9.20, while DeepEP-Ascend requires Ascend 950 silicon with UBMEM connectivity. This limits the immediate user base to Huawei's own engineers and a limited number of Chinese AI labs with access to this specific hardware.
Current GitHub metrics indicate low initial adoption, with DeepGEMM-Ascend having 515 stars compared to 8,522 for its CUDA sibling, and DeepEP-Ascend having 224 stars compared to over 10,200 for DeepEP. This suggests that while the infrastructure is being built, developer confidence and access to the hardware remain significant hurdles.
Source details: techradar.com ↗
Why it matters
This move represents a strategic effort to erode Nvidia's dominance in AI infrastructure by addressing the primary barrier to adoption: high switching costs. By providing production-ready kernels that match the API of popular CUDA libraries, DeepSeek aims to make it easier for developers to port existing workloads to Huawei hardware. However, the practical impact is currently constrained by the limited availability of Ascend 950 chips outside of China and persistent developer concerns regarding the stability and usability of Huawei's CANN software stack. The low initial GitHub star counts compared to their CUDA equivalents suggest that while the technical foundation is being laid, widespread industry adoption remains a significant challenge.
Nvidia's dominance in AI hardware is largely protected by the high cost of switching from its CUDA ecosystem. By providing open-source tools that mirror CUDA's API, DeepSeek and Huawei are attempting to lower this barrier, making it more feasible for developers to consider alternative hardware.
The release is part of a broader trend of Chinese tech firms seeking to reduce reliance on US-made chips, particularly in the face of export controls. Huawei's Ascend silicon is positioned as a domestic alternative, but its success depends on building a robust software ecosystem that can compete with CUDA's maturity.
Despite the technical efforts, the practical impact is currently limited by hardware availability. Access to Ascend 950 chips is restricted, and there are ongoing concerns about the stability and usability of Huawei's CANN software stack, as reported by developers and cited by TechRadar.
The low adoption rates on GitHub suggest that the developer community is still hesitant to make the switch. This indicates that while the strategic intent is clear, the practical execution faces significant challenges in terms of both hardware access and software reliability.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Monitor the growth in GitHub stars and forks for DeepGEMM-Ascend and DeepEP-Ascend to gauge developer interest. Watch for announcements from other major Chinese AI labs regarding their adoption of these tools. Additionally, track any updates to Huawei's CANN software stack that might address the stability and usability issues cited by developers, as this will be critical for broader ecosystem growth.
Track the growth in GitHub stars and forks for the new DeepSeek Ascend repositories to gauge developer interest and adoption over time.
Monitor for announcements from other major Chinese AI labs, such as Baidu or Tencent, regarding their use of these tools or their own Ascend-compatible software stacks.
Watch for updates to Huawei's CANN software stack, particularly any improvements in stability and usability that could address the concerns raised by developers.
Observe any changes in US export controls or Chinese domestic policies that might affect the availability of Ascend hardware or the ability of international developers to access these tools.