On 30 September 2026 Chinese firm DeepSeek posted on its official WeChat account that it is partnering with Huawei Technologies to build programming tools aimed at Ascend accelerators. Reuters, Semafor, TechNode and other outlets on 30 September and 1 October agree on the core: open-sourcing of infrastructure, not merely a statement of intent.
The cited package includes TileLang, a high-level language meant to be universal and easy to programme while still reaching near-hardware performance, plus compute and inter-chip communication libraries. Later coverage describes as many as six open-source components spanning programming, compute, communication, attention, kernels and data processing, with names such as DeepGEMM Ascend and DeepEP Ascend.
DeepSeek says Huawei provided full support and that the two advanced a “supernode” solution based on 128 Ascend 950 chips, optimising both computation and communication. That is a different step from the earlier plan to install more than 160,000 Ascend chips at an Inner Mongolia data centre: here the focus is the software that makes that silicon usable.
The geopolitical context is inescapable. U.S. export restrictions have constrained Chinese access to some advanced Nvidia processors. Jensen Huang has said in recent coverage that Nvidia’s share of China’s advanced AI-chip market fell from roughly 95% to effectively zero in the restricted segment. Without CUDA, software becomes as strategic as the wafer.
For Huawei, a denser developer ecosystem is the condition for Ascend to stop being hardware with an incomplete stack. For DeepSeek, which trained on H800s and has been adapting models to Ascend through 2026, owning optimised kernels lowers the toll of migrating workloads and gives it a voice in the domestic software roadmap.
Nuance matters: TechNode stresses the announcement did not provide independent performance results or verified production deployments of the full stack. Comparing Ascend+TileLang with CUDA+Nvidia remains, as of 1 October 2026, an exercise on partial data. Neural Pulse marks those performance figures as verification in progress.
Opening the code can attract contributors and outside scrutiny, but real adoption will hinge on documentation, stability, physical access to Ascend and compatibility with common training frameworks. A language without mature bindings to the tools ML teams already use often stays a symbolic repository.
In parallel, Beijing has considered allowing domestic firms—including Alibaba and ByteDance—to buy Nvidia RTX PRO 5500 chips, per The Information reporting covered in prior batches. That tension—buy what Nvidia exports allow while building the Ascend plan B—defines China’s dual 2026 strategy.
For the rest of the world, the announcement is a decoupling signal at the stack layer: not only chips, but languages, collective-communication libraries and supernode formats. Cloud providers outside China will watch whether TileLang or its ideas leak into international projects, or remain a sovereign silo.
Investors following DeepSeek after the V4 cycle will read the move as compute-margin defence: without reliable access to leading GPUs, algorithmic edge erodes. Open-sourcing tools is also technical diplomacy toward the local data-centre and OEM ecosystem.
Risks include compiler maturity, multi-chip communication bugs, and the chance that a partial easing of export controls reactivates preference for CUDA and leaves TileLang as unused insurance. Invezz summarises that scenario as a key risk to the Ascend thesis.
Neural Pulse concludes that DeepSeek×Huawei on TileLang is the software layer of an infrastructure race we already knew in silicon. The headline is not “China has chips”; it is “China is trying to have its own CUDA.” If the six open-source modules gain traction in 2026–2027, the training-tool map will stop being a U.S. monolith.
