DeepSeek Open-Sources a Toolkit to Help Huawei Chips Do What Nvidia's Software Does, for Free

The most durable moat in the AI chip business is not a chip. It is a programming environment called CUDA, which is what happens when a generation of engineers learns to write fast code for one company’s hardware and then, understandably, declines to learn it again. On Wednesday, DeepSeek did what a large number of people have been waiting years for someone to try: it published, for free, a set of tools for making the other guy’s hardware behave.

Specifically, DeepSeek open-sourced infrastructure components for Huawei’s Ascend accelerators, covering Ascend support for TileLang, a kernel programming language, along with compute kernel libraries and a distributed communication library. Per Startup Fortune, DeepSeek says it can serve as a replacement for CUDA. Per common sense, the word “can” is doing something heroic.

The package, as listed by the coverage: TileLang; DeepGEMM for matrix operations; DeepEP for large-scale communication across chips; TileKernels for vector computations; FlashMLA for long-context attention; and DeepSelect for data filtering. That is a toolbox with the modest names of a hardware store and the ambition of a border crossing.

The key design choice, according to KuCoin’s summary, is that each component maps one-to-one to libraries DeepSeek previously released for Nvidia GPUs, so developers can use the same tools on both. It is a translation service where the source language is CUDA and the target language is a domestic supply chain.

The company reported that benchmarks show computational and communication performance “approaching hardware limits,” and Startup Fortune says TileLang supports most of the operators used in training DeepSeek’s V4 series. Huawei, for its part, collaborated on a 128-card super-node design built on the Ascend 950. Cryptonomist adds that DeepSeek is separately building a data center in Inner Mongolia designed to hold no fewer than 160,000 Ascend accelerators. That is a very large building for a chip you have just finished writing the manual for.

Why this matters is not complicated. U.S. export controls limited what Nvidia hardware Chinese firms could buy, which turned the software layer into the bottleneck: it is one thing to make a competitive chip and another to make a chip that a graduate student can, by Thursday, actually use. As Cryptonomist notes, nearly every model, training pipeline and inference system depends on CUDA. Which means the moat is not the silicon. The moat is muscle memory.

You can see the sales pitch coming from across a parking lot. It is “yes, it is a different chip, but the tools are the same,” which is the same promise made by every airline that has ever told you the new terminal is basically the old terminal.

Startup Fortune is refreshingly cautious about it, and we will follow its lead, because being funny about a precise fact beats being funny about a sloppy one. It says the Ascend 910C delivers about 60% of the inference performance of Nvidia’s H100. It says major Chinese companies, ByteDance, Tencent and Alibaba, have placed hardware orders. And it says the real limit is that tools mirroring another company’s stack are only as good as the number of developers who actually port their code. Whether that happens is, in its words, the thing worth watching next.

There is an old software adage that the best API is the one that already exists. DeepSeek is betting on the inverse, that the best API is the one that already exists and also runs on the thing you can buy. It is a shrewd angle. It also lands DeepSeek in the relatively new and slightly awkward position of being a model lab that is now, in practice, a chip-ecosystem company. It began by writing very good models. It is now writing the manual for the machine that runs them.

For Nvidia, the news is a study in ambiguity. In the narrow sense, this is a competitor’s toolkit. In the broader sense, it is a testimonial: a full suite of open-source libraries, each explicitly modeled on the Nvidia version, is about the sincerest form of flattery available in the semiconductor industry. When someone spends a year building a translator for your language, it is because your language won.

And a note of honesty on what this doesn’t establish. It does not say Ascend is a match for the newest Nvidia hardware. It does not say developers are leaving. It says that a cheaper, closer-to-home option has, as of today, a much better manual. Manuals are underrated. Most of the history of technology is a history of the thing with the better documentation.

We will end where Startup Fortune does, which is with the boring, correct, and slightly ominous observation that adoption is the whole game. TileLang could be a masterpiece and still sit unused, like a superb piano in a flat with no one who plays. Or a large fraction of the world’s kernel writers could spend the next year quietly porting things over, and the moat could become a puddle.

We will know which by watching what developers do, rather than what companies announce. Developers, historically, do the easiest thing available, with a stubbornness that is the closest thing this industry has to a law of nature. DeepSeek has just made a second thing easy. It is now up to gravity.

Sources: