Huawei and DeepSeek are building programming tools for Huawei's Ascend processors, opening the platform's computing and communications libraries to developers in a direct strike at Nvidia's deepest competitive moat: its CUDA software ecosystem. DeepSeek announced the collaboration on its official WeChat account, and Reuters reported the news Wednesday.

The move matters because Nvidia's dominance in AI acceleration rests on more than GPU throughput. Nearly two decades of developer tooling, libraries and frameworks built around CUDA make Nvidia hardware the default target for AI software. Building a competitive accelerator is one engineering challenge; equipping it with a programming environment that developers can actually use efficiently is another.

DeepSeek and Huawei are attacking both sides of that problem simultaneously. The two companies have developed a supernode computing system that links 128 Huawei Ascend 950 processors, with Huawei providing computation and communication optimization for the cluster. The Ascend 950 belongs to the next generation of Huawei AI processors, which the company unveiled two weeks ago alongside its new supernode computing systems. At that launch, Huawei predicted its AI technologies would see widespread use for model training next year — a roadmap claim, not yet a confirmed deployment figure.

The second pillar of the software push is TileLang, an open-source, high-level programming language designed specifically for AI chips. DeepSeek claims TileLang offers a simpler programming model than Nvidia's CUDA, potentially boosting development efficiency and simplifying code. The language was highlighted in the same WeChat announcement as part of the effort to make the Ascend platform accessible to outside developers.

The commercial stakes for Nvidia are straightforward. Most discussion of Chinese competition in AI accelerators has focused on hardware: whether domestic chips can match the performance of Nvidia's data-center GPUs under U.S. export controls. But Nvidia's pricing power and market share also depend on switching costs — developers trained on CUDA, and codebases written for it. If open tools lower those switching costs for Ascend hardware, Chinese cloud operators and AI labs gain a credible path to a full-stack alternative rather than a drop-in chip substitution.

For investors, as GuruFocus notes, the development exposes a new dimension of Chinese competitiveness. The attack on Nvidia comes not from a faster chip but from the software layer that makes chips useful. Chinese technology companies are working to build an AI computing environment independent of Nvidia, and DeepSeek — whose own models already demonstrated that competitive AI systems can be trained on constrained hardware budgets — brings developer credibility to that effort.

Huawei's timeline adds urgency. If its prediction of broad adoption for model training next year holds, the Ascend ecosystem could move from defensive substitute to genuine contender for Chinese AI workloads, with TileLang and the opened Ascend libraries determining how quickly developers make that shift.