The collaboration between DeepSeek and Huawei has moved beyond rumors to become a concrete reality: both Chinese companies have open-sourced part of the software that allows programming the Ascend accelerators. The move isn't so much about competing in the silicon market as it is about breaking their dependence on CUDA , Nvidia's platform that for years has served as the mandatory gateway to artificial intelligence.
What's striking is that the announcement comes via open source. DeepSeek has released tools and libraries so developers can work with Huawei chips without completely rebuilding the core layers of their projects. Huawei has endorsed the work, and the focus is on a complete ecosystem, not a single product.
Open tools for Ascend chips
The released package includes six software components and several libraries for mathematical operations and data exchange between accelerators. The most talked-about component is TileLang, a high-level programming language that the lab has optimized for the Ascend 950 accelerators. The idea is to make writing code for these chips simpler than using Nvidia's tools, without sacrificing high performance.
TileLang didn't emerge from nowhere: it was already used in previous Nvidia cards and is now being brought to Huawei. The company maintains that a common, easy-to-learn language that leverages the hardware is key to building an AI ecosystem that doesn't depend on external providers—a goal that Beijing has been advocating for over two years.
In addition to the language, DeepSeek and Huawei have optimized a supernode of 128 Ascend 950 accelerators . In these types of systems, it's not enough for each unit to be powerful; how data is passed between them is crucial. This is where Nvidia has built part of its advantage with technologies like NVLink, and that's why the Chinese approach is necessarily a comprehensive one.
DeepSeek is no newcomer to this field. It had already developed tools for other AI accelerators, and according to several reports, its V4-Flash model was optimized to run on Ascend. Its next known step is ambitious: it plans to install at least 160.000 Huawei accelerators in a data center in Inner Mongolia, a figure that gives an idea of ​​the scale it aims to handle.
The CUDA wall and the limits of the plan
Huawei hasn't only been held back by the physical components of its chips. For almost twenty years, much of AI software has been written with CUDA in mind , and that installed base is difficult to displace. It's not just a compiler: it also includes libraries, testing tools, manuals, examples, and a very large community. According to various analyses, CUDA boasts around four million developers, a network that Nvidia has turned into its strongest defense.
Huawei has its own alternative, CANN , but that's not enough with just a compiler and a few libraries. To truly compete, you need diagnostic tools, training, and people capable of resolving real-world issues. That's the gap DeepSeek is trying to fill with its expertise in low-level optimization and an existing open-source community.
The political and regulatory context also plays a role. Washington continues to restrict the sale of advanced accelerators and access to key semiconductor manufacturing machines. These obstacles limit Huawei's international expansion, although the market within China is large enough to make the experiment economically viable.
Huawei has admitted that its production is insufficient to meet domestic demand and that, for the time being, it prefers to focus on China rather than expanding internationally. Its roadmap includes the Ascend 960DT for early 2027 and a roadmap through 2029, with the intention of doubling computing power with each upgrade.
Meanwhile, DeepSeek is also eyeing the markets. According to Bloomberg, it's preparing a funding round with a valuation close to $74.000 billion , in anticipation of a potential IPO. This combination of financial muscle, open source code, and domestically developed hardware explains why the operation is being closely followed outside of China.
For a development team, the real dilemma is the cost of switching platforms. Every feature that currently calls an Nvidia library would need an equivalent version for Huawei, and compatibility issues often arise after the system is already up and running. DeepSeek's tools aim to reduce this friction, but they don't eliminate the advantage gained by CUDA, nor do they guarantee that all models will perform equally well on Ascend.
In Spain and the rest of Europe, the implications are less immediate, but no less significant. Most companies and data centers rely on US suppliers, and there is no comparable local alternative to CUDA. The Chinese move serves as a reminder that technological autonomy is not solely determined by chip factories, but also by software, languages, and communities.
The battle won't be won with a press release. What will matter is how many projects abandon CUDA, how many people start using TileLang , and whether Huawei has enough chips to meet demand. If the tools don't convince, many teams will revert to their old ways without much trouble.
DeepSeek and Huawei's move places the competition at a less visible layer than hardware: the software that makes each chip usable. It's not a definitive blow to Nvidia, but it is a step that reduces the cost of leaving its ecosystem and, in the medium term, could change purchasing decisions within China. Outside of China, the effect will depend on Huawei's production capacity and the pressure the United States exerts on its allies.