DeepSeek Releases Huawei-Compatible AI Tools to Challenge Nvidia CUDA
Chinese artificial intelligence startup DeepSeek has released an open-source software suite designed to program Huawei's Ascend AI accelerators, marking a significant step in China's effort to build a domestic alternative to Nvidia's dominant CUDA platform. The toolkit, which includes the custom programming language TileLang, allows developers to write instructions for AI chips without relying on American software.
Building an Independent AI Ecosystem
The release includes six software modules that mirror tools previously developed by DeepSeek for Nvidia hardware. Among these is an Ascend-compatible version of TileLang, which supports native code generation, automatic scheduling, and synchronization for Huawei's Ascend 950 chips. DeepSeek stated that the suite provides computational and communication performance approaching hardware limits, a result achieved through close collaboration with Huawei.
According to Tom's Hardware, huawei's Ascend accelerator clusters compensate for lower per-chip memory bandwidth and density by scaling supernode card counts, but consume roughly four times the electrical power of comparable Nvidia Blackwell architectures.
According to Hugging Face, for Chinese AI Startups Ecosystem, SiliconFlow and Zhipu AI publicly committed to and deployed DeepSeek model architectures across Huawei Ascend chips, creating an industry-wide push for domestic software-hardware integration.
According to Pandaily, tileLang serves an architectural role on Ascend chips comparable to OpenAI Triton on Nvidia GPUs, abstracting low-level hardware assembly into Pythonic tile primitives while achieving near-metal performance.
DeepSeek's broader software package includes specialized libraries such as DeepGEMM Ascend for matrix multiplication and DeepEP Ascend for large-scale cross-device communication. These tools are intended to streamline the development of high-performance kernels, which are essential for training and running AI models on graphics processing units.
Strategic Partnership and Infrastructure
The collaboration underscores a deepening alliance between DeepSeek, which develops open-weight models to compete with global leaders like OpenAI and Anthropic, and Huawei, which is spearheading China's domestic production of AI accelerators. DeepSeek reported that Huawei provided unreserved and vigorous support during the research and development process. The two companies are currently working on a supernode architecture powered by 128 Ascend 950 processors.
This move aligns with Beijing's broader push for technological self-reliance amid ongoing US export restrictions. DeepSeek is actively integrating domestic hardware into its operations, with plans to deploy at least 160,000 of Huawei's most powerful accelerators at a new data center currently under construction in Inner Mongolia. According to Pandaily, other Chinese chip firms, including Cambricon Technologies and Moore Threads, have also confirmed compatibility with DeepSeek's models. Meanwhile, Huawei has accelerated its hardware roadmap, recently announcing that its Ascend 960DT training chip is expected to be ready in the first quarter of 2027, three quarters ahead of its original schedule.