DeepSeek and Huawei Expand Nvidia Alternatives With Open-Source Tools for Ascend Chips
DeepSeek and Huawei are expanding their cooperation in artificial intelligence by developing open-source programming tools optimized for Huawei’s Ascend chips, in the latest effort by Chinese technology companies to reduce their dependence on Nvidia’s dominant AI ecosystem.
DeepSeek said on September 30, 2026 that it was open-sourcing programming infrastructure for the Ascend platform, including computing and communication libraries, while the two companies have also worked on a supernode solution built around 128 Ascend 950 chips.
DeepSeek and Huawei expand work around Ascend
DeepSeek said it partnered with Huawei Technologies to develop programming tools specifically optimized for Ascend processors.
According to Reuters, Huawei provided full support for the development of the new infrastructure, while the two companies jointly improved both computation and communication in a system using 128 Ascend 950 chips.
The announcement comes roughly two weeks after Huawei unveiled its next generation of AI processors and supernode systems and said it expected wider use of its AI infrastructure for model training next year.
Why Nvidia’s software ecosystem matters
Nvidia’s position in artificial intelligence is not based on hardware alone. Its CUDA software ecosystem has become deeply embedded in the way developers and researchers build and run AI workloads on GPUs.
That creates a challenge for alternative chipmakers: producing competitive hardware is only part of the task. They also need programming tools, libraries and developer infrastructure that make it practical to move workloads away from Nvidia systems.
The DeepSeek-Huawei effort addresses that software layer directly.
TileLang moves into focus
DeepSeek highlighted TileLang, an open-source high-level programming language designed to make it easier to build high-performance workloads for AI accelerators.
The company said the goal is to help create a more independent software ecosystem while allowing developers to access the performance capabilities of the underlying hardware without unnecessary programming complexity.
DeepSeek described TileLang as offering a simpler programming model than Nvidia’s CUDA.
That does not mean TileLang has reached CUDA’s scale, adoption or developer ecosystem. Instead, it represents one part of a broader effort to make Ascend hardware easier to program and deploy for AI workloads.
A 128-chip Ascend 950 supernode
The partnership also includes a supernode solution built around 128 Ascend 950 chips.
Such systems connect large numbers of AI processors so they can operate as a tightly integrated computing platform, an increasingly important capability as frontier models require greater processing power, memory and chip-to-chip communication.
DeepSeek said the work with Huawei focused on improving both computation and communication efficiency across the system.
China pushes to reduce reliance on Nvidia
The project is part of a wider effort by Chinese technology companies to develop domestic alternatives across the AI stack, from processors to models and software.
U.S. restrictions on exports of some advanced AI chips to China have increased the importance of domestic suppliers such as Huawei, while DeepSeek has built its profile around AI models and techniques designed to use computing resources more efficiently.
Profile News has previously followed DeepSeek’s growing role in the international AI debate as Chinese developers become more prominent in the global technology race.
Huawei is not China’s only AI hardware effort
China’s push extends beyond the DeepSeek-Huawei partnership, with other companies developing their own chips, models and computing infrastructure.
Profile News has also covered Alibaba’s AI chip and Qwen development strategy, another example of Chinese companies attempting to build more of the AI technology stack domestically.
Does this threaten CUDA?
It is too early to treat the new tools as an equivalent replacement for CUDA. Nvidia’s advantage includes not only its software but also a large developer community, mature libraries and extensive cloud and enterprise support built around its technology.
The significance of the DeepSeek-Huawei partnership is that it targets one of the hardest parts of reducing dependence on Nvidia: making alternative chips practical for developers to program and use efficiently.
If the Ascend ecosystem succeeds in attracting more developers and AI workloads, competition could increasingly move beyond the question of which company makes the fastest chip toward which company can build the strongest combination of hardware, software and developer tools.
Sources
Reuters — DeepSeek-Huawei partnership, open-source Ascend infrastructure, TileLang and the 128-chip Ascend 950 supernode — September 30, 2026.
TileLang-Ascend — Open-source TileLang implementation optimized for Huawei’s Ascend architecture.
