DeepSeek and Huawei co-design Ascend chips to cut Nvidia reliance

DeepSeek and Huawei are co- designing Ascend AI chips to reduce Nvidia reliance, with TileLang compiler open- sourced and Ascend 960DT capacity allocated to DeepSeek.

DeepSeek and Huawei co-design Ascend chips to cut Nvidia reliance

DeepSeek and Huawei are moving beyond standard supplier agreements to co-design future AI chips, a shift that directly impacts how Chinese tech firms reduce reliance on hardware. This partnership signals a structural change in the domestic AI supply chain, where model makers now influence silicon architecture. Buyers monitoring the Nvidia monopoly in China should note that this collaboration accelerates the development of viable domestic alternatives.

DeepSeek and Huawei collaboration logo
DeepSeek and Huawei collaboration logo

DeepSeek engineers work with Huawei on Ascend 950DT and 960DT processors

The core of this effort involves the Ascend 950DT and the upcoming Ascend 960DT processors. Huawei engineers are stationed at DeepSeek to integrate specific model training requirements directly into the chip design process. This close integration aims to optimize performance for DeepSeek's specific workloads rather than relying on generic silicon.

  • Ascend 910C Deployment: Over 1,000 units deployed; architecture supports up to 1 million NPUs
  • Ascend 960DT Capacity Allocation: Most capacity from Q1 2027 launch allocated to DeepSeek
  • TileLang Compiler: Open-source compiler for Ascend platform with high-performance implementations for all DeepSeek training tiles

Huawei has already deployed over 1,000 units of the Ascend 910C, with architectures designed to support up to 1 million NPUs. The Ascend 950DT is currently undergoing testing for large-scale pre-training scenarios that require thousands of chips running continuously for weeks. DeepSeek plans to begin using the Ascend 950DT for pre-training by the end of this year, although this timeline depends on the ongoing testing results.

To support this hardware, DeepSeek has open-sourced TileLang, a compiler for the Ascend platform that provides high-performance implementations for all DeepSeek training tiles. The compiler runs across chips and mirrors Nvidia components, directly challenging the code migration barriers that the CUDA ecosystem uses to retain customers. Most capacity from the Ascend 960DT launch in the first quarter of 2027 will be allocated to DeepSeek.

Market data from IDC shows Huawei captured 28% of China's AI chip market in the first half of 2026, narrowing the gap with Nvidia to just 6 percentage points. This partnership represents a concrete step in reducing dependence on foreign silicon, with DeepSeek and Huawei accelerating their joint efforts to build a self-sufficient AI infrastructure.

Discussion

0 comments

Log in to join the thread with a thoughtful take, question, or correction.

Add to the discussion