DeepSeek partners with Huawei on open-source programming infrastructure

DeepSeek has announced a new partnership with the Chinese tech giant Huawei to develop programming tools optimised for its AI chips.

Together, the two firms have established a new “supernode” solution formed from 128 Huawei Ascend 950 chips and worked to establish new Huawei-optimised programming tools.

The move reduces DeepSeek’s dependence on Nvidia’s ecosystem, particularly its CUDA parallel computing platform.

CUDA has become one of the core underlying layers of the world’s most popular AI libraries, since it allows developers to use the full power and throughput of a graphics processing unit (GPU) for the mathematical computations necessary in AI training.

DeepSeek and Huawei have developed a new toolkit based on the open-source GPU programming language TileLang, to allow for rapid development of AI for Huawei chips that is not dependent on Western hardware or software libraries.

In a statement, DeepSeek said: “To build a new ​generation of independent, self-controlled GPU software ecosystems, 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.”⁠

The firm added that TileLang meets this need and is simpler to use than Nvidia’s CUDA.

DeepSeek has committed to open-sourcing the programming tools alongside libraries that cover compute and communication. This is important for Chinese AI labs because to make Huawei chips work for training runs they need to efficiently network and share data as a cluster.

Huawei dominates the Chinese chip market and is officially backed by the government as the go-to hardware option for Chinese AI.

It has benefited from strict export controls on Nvidia chips entering China, which Beijing has only recently eased. As one of the most prominent Chinese AI developers, DeepSeek has expressed a public desire to reduce its dependence on Nvidia and in July Reuters reported it could develop its own chips to achieve this aim.

On 17 September, Huawei announced a new roadmap for its upcoming flagship chips the Ascend 960DT and Ascend 960PR. The new offerings, due in the first and third quarters of 2027 respectively, aim to more than double the training and inference power that Huawei can offer domestic AI firms.

In addition to the leap in raw performance it is promising through the chips, Huawei is set to launch its new HiF4 data format and Peerium Computing Architecture with the Ascend 960 family.

The former achieves up to double the tokens per second in AI output, while the latter allows up to a million Ascend chips to work in tandem as a computing cluster for multi-trillion parameter training runs.