DeepSeek Open-Sources AI Infrastructure for Huawei Ascend in Bid to Cut Nvidia Switching Costs

DeepSeek has open-sourced a suite of infrastructure components for Huawei's Ascend accelerators, targeting lower migration costs for developers seeking alternatives to Nvidia's hardware and software stack. As reported by TechRadar Pro (https://www.techradar.com/pro/huawei-and-deepseek-release-open-source-ai-tools-in-a-bid-to-lower-nvidia-exposure-but-will-programmers-make-the-switch), the release introduces two new GitHub repositories alongside updates to several existing libraries.

The new repositories consist of DeepGEMM-Ascend, which handles matrix multiplication, and DeepEP-Ascend, built for the all-to-all communication required by mixture-of-experts (MoE) architectures. Analysis by FourWeekMBA noted that other packages named in the launch—including TileKernels, DeepSelect, and FlashMLA—are pre-existing projects that received Ascend-specific updates rather than newly created codebases.

International media framed the launch around broader geopolitical and competitive positioning. Reuters described the effort as two Chinese firms deepening collaboration to build alternatives to Nvidia amid trade controls, while Bloomberg highlighted TileLang as the initiative's centerpiece, characterizing it as China's answer to CUDA.

However, TileLang itself is neither newly introduced nor proprietary to DeepSeek. The language was open-sourced in January 2025, and its Ascend adapters were published in a separate repository in September 2025. Rather than serving as an exclusive weapon against Nvidia, TileLang operates as a shared abstraction layer supported across multiple Chinese accelerator vendors, with Huawei acting as one of several participants.

Practical adoption of the newly released tools is constrained by narrow hardware requirements. DeepGEMM-Ascend requires Huawei's Ascend 950 series and the CANN 9.20 toolkit, while DeepEP-Ascend requires Ascend 950 silicon configured with UBMEM connectivity. These prerequisites limit immediate utility to Huawei's internal teams and select domestic labs with access to the silicon.

Rather than making outright performance claims, DeepSeek structured the libraries to minimize switching friction. DeepGEMM-Ascend retains the identical Python package name and API surface of its CUDA counterpart, while DeepEP-Ascend aligns public buffer interfaces with the Nvidia version. The strategy directly targets Nvidia's developer moat by reducing the engineering overhead required to port existing workloads.

Initial developer traction reflects the steep path ahead. GitHub metrics show DeepGEMM-Ascend with 515 stars and 39 forks, compared to 8,522 stars and 1,359 forks for the original DeepGEMM repository. DeepEP-Ascend counts 224 stars, trailing the core DeepEP repository's total of more than 10,200 stars.

The release also navigates persistent usability challenges in Huawei's software stack. Citing reporting from the Financial Times, ChinaTalk previously noted that a Huawei researcher described the CANN toolkit as unstable and difficult to work with, while developers have described porting work on the 910B as fraught with pitfalls. To ease friction, Epoch AI has reported that Huawei routinely sends engineering teams to major clients like Baidu and Tencent to assist with CUDA code conversion.

While DeepSeek's open-source release provides battle-tested production kernels, the software remains tied to specific accelerator families, firmware versions, and localized toolkits that remain largely inaccessible to international engineering teams.