DeepSeek Teams With Huawei to Challenge Nvidia's CUDA

This collaboration focuses on building a software layer for AI chips, specifically through the development of computing and communication libraries and the TileLang programming language. DeepSeek states TileLang is designed to improve development efficiency and simplify code structure, addressing what it calls the priority of building a versatile, high-level language that maximizes hardware performance.

The effort signals a concerted move by Chinese technology companies to reduce reliance on Nvidia's ecosystem, particularly CUDA, which analysts consider a significant barrier to replacing Nvidia hardware. DeepSeek's past models, V3.2 and V4, have already supported Ascend chips. The company reportedly plans to deploy a large number of Huawei Ascend 950DT chips for inference in a new data center.

The reported full support from Huawei and the joint optimization of a 128-chip SuperPod solution suggest a deep integration of efforts. This type of hardware-software co-development is critical for establishing a viable alternative to an entrenched platform like CUDA, as it addresses both the raw processing power and the usability for developers.