DeepSeek's TileLang Push Frees AI Chips From Nvidia's CUDA

DeepSeek announced on Wednesday, September 30, 2026 that it is open-sourcing a package of software tools built for Huawei's Ascend AI chips, a step that could loosen Nvidia's grip on the software behind modern artificial intelligence. The company revealed the release in a post on its official WeChat account, according to Reuters, saying the move is meant to help build a software ecosystem that does not depend on Nvidia's CUDA platform.

At the center of the release is TileLang, a high-level programming language and compiler that DeepSeek describes as easier to use than Nvidia's CUDA. The Ascend version of TileLang wraps the low-level instructions of Huawei's Ascend C environment, giving developers high-level programming tools without sacrificing the hardware's performance, the company said. DeepSeek says every operator in TileLang that it uses in its own model training now has a high-performance implementation on Ascend.

What DeepSeek is releasing

The package is not a single tool. Alongside TileLang, DeepSeek is open-sourcing compute and communication libraries that it says map one to one with the versions it previously released for Nvidia hardware. The list includes DeepGEMM for general matrix operations, DeepEP for large-scale cross-device communication, TileKernels for vector computation and memory access, FlashMLA for sparse attention in long-context workloads, and DeepSelect for data filtering.

These are the building blocks of AI training. Matrix operations power the neural-network math, communication libraries move data between chips, and sparse attention trims the cost of processing long documents. In multiple key test cases, DeepSeek said, the computational and communication performance of the new components has approached the hardware's limits.

"To build a new generation of independent, self-controlled GPU software ecosystem, the first priority is establishing a high-level language that is universal, easy to program, and still capable of reaching the hardware's full performance potential," DeepSeek said in the announcement. "TileLang was created precisely to meet this need," the company added, saying it offers "a simpler programming model" than Nvidia's CUDA that improves development efficiency and simplifies code logic.

The open-source move is not entirely new territory. A community-maintained adapter for the Ascend platform, known as tilelang-ascend, has existed since September 2025. What changes now is official backing: DeepSeek itself is releasing and maintaining the Ascend infrastructure, which signals that the company's own training pipeline runs on these tools.

Why the move matters

Nvidia's power in AI rests on more than its chips. For nearly two decades, CUDA, its programming platform, has been the default way developers write AI software. Because their code, libraries, and workflows are built on CUDA, developers tend to stay with Nvidia even when rival hardware looks comparable on paper.

That dynamic has real weight for Chinese companies. US export controls have limited China's access to Nvidia's most advanced chips, pushing firms such as Huawei to build domestic alternatives like the Ascend line. The missing piece has often been software: fast chips with thin tooling lose to slower chips with mature ecosystems. By giving developers a free, high-level programming model tuned for Ascend, DeepSeek and Huawei are addressing the software gap directly rather than waiting for a hardware breakthrough.

DeepSeek shot to global attention earlier with low-cost, high-performing models that drew heavy attention from markets and investors. This release extends that approach from models to infrastructure: open tools, lower costs, and a way for the developer community to build on non-Nvidia hardware.

There is precedent for the strategy. AMD, Intel, and other chipmakers have spent years trying to pry developers away from CUDA with their own software stacks, with limited success. What is different here is the combination: a model lab that developers already trust, a chipmaker with a growing domestic market, and tools released under open-source licenses that anyone can inspect and extend.

Open-sourcing the tools removes one barrier, but an ecosystem also needs documentation, support, and developers willing to write and maintain Ascend code. The WeChat announcement frames TileLang as the starting point for that ecosystem, not the finish line. For now, the release gives Chinese AI developers something they did not have before: a company-backed, open path to program some of the country's most advanced AI chips without touching Nvidia's software at all.

Technical coverage of the release breaks down each library and how it compares with the Nvidia versions (read the full technical rundown); a second roundup lists every open-sourced component in the package (see the component breakdown).