Michael JohnsonOn September 30, DeepSeek announced joint work with Huawei on tools for Ascend accelerators and released related software components. This is a step toward developing an alternative hardware platform, but not proof that it is ready to replace CUDA and Nvidia in production systems. The announcement does not fall within the 24-hour window as of the date of this analysis—October 2, 2026—so it is more accurately described as recent news.
Reuters’ report concerns infrastructure for Huawei Ascend: TileLang and libraries for computation and inter-chip communication. Reuters also relays a statement about jointly developing a 128-chip Ascend 950 supernode. This is a statement by the project participants; the sources reviewed contain no independent performance comparison between such a system and a comparable Nvidia configuration.
What was published
The DeepGEMM-Ascend repository states that it supports Ascend 950 and BF16, FP8, and FP4 operations. The project describes the library as API-compatible with DeepGEMM, but also notes differences: for example, Ascend stores scaling factors differently from Nvidia. The documentation lists CANN, torch_npu, and the Bisheng compiler as requirements. Interface compatibility, therefore, does not mean code can be moved between platforms without accounting for their differences.
The same repository publishes the project’s own benchmarks on Ascend 950DT. These are project results, not an independent performance assessment under real-world production workloads.
The DeepEP-Ascend library is designed for data exchange during training and inference, including in Mixture-of-Experts models. Its documentation specifies that the measurements were conducted on a test hardware and software version supplied to DeepSeek—PoC HDK—with additional manual tuning. This configuration was not publicly distributed. Huawei planned to release a commercial HDK around mid-October 2026; this is a plan stated in the documentation, not a confirmed release date.
TileLang did not first appear with this announcement: the TileLang-Ascend history indicates that the project became open source on September 29, 2025. The news concerns a new phase of collaboration and related libraries, not the launch of TileLang from scratch.
What has not yet been proven
A programming language can make it easier to create computational kernels, but replacing a platform in practice also requires compilers, libraries, debugging tools, framework support, and accessible hardware. Open repositories make it possible to study the code and stated interfaces, but do not show how easy or inexpensive it is to move production workloads to Ascend.
The testing conditions are a particularly important caveat for DeepEP-Ascend: the published figures were obtained on a non-public test configuration. And measurements of individual operations, even if they can be reproduced, do not by themselves answer questions about the stability and overall performance of a complete system.
The code release, documentation, and listed environment requirements are confirmed. The performance of the new libraries, broad availability of the stated configuration, and ease of integration into existing workflows have not yet been confirmed by independent testing.
The announcement is therefore best viewed as an effort to strengthen the software ecosystem around Ascend. The practical value of these tools will become clearer once a commercial environment is available, along with independent, reproducible tests and detailed reports on migrating real workloads. For now, this is a concrete step in the competition for the software layer, not a proven replacement for CUDA.