DeepSeek teams up with Huawei to challenge NVIDIA's AI computing hegemony: participating in next-generation chip design and pushing ahead with pre-training of domestic chips by the end of the year.

Beating AI News (Dongcha) reports that last year, DeepSeek attempted to train its R2 model using Huawei Ascend chips but failed. The Financial Times previously disclosed issues with chip stability, inter-chip communication, and software support. Even with Huawei engineers on-site assisting, DeepSeek ultimately switched back to Nvidia chips to complete the training. Over a year later, the two parties have deepened their collaboration to include hardware design. The Information now reveals that Huawei engineers are permanently based at DeepSeek’s offices, integrating the practical demands of model training directly into the design of next-generation Ascend chips—including how to schedule computing tasks and transfer data between chips. At the end of September this year, DeepSeek open-sourced the Ascend version of TileLang, plus underlying tools for matrix computation, MoE expert communication, etc., to reduce the development work required for migrating models to domestic chips. However, getting the software to run is still a long way from full pre-training. Tens of thousands of chips need to operate continuously for weeks; any single chip failure or data transmission delay could slow down the entire cluster. Now, the two sides are starting to tackle these issues at the chip design stage. Huawei plans to prioritize allocating most of its supply of Ascend 960DT, set to be ready in Q1 2027, to DeepSeek rather than distributing it to other model companies. DeepSeek, for its part, is preparing to start pre-training as early as the end of this year using the 950DT. So far, related tests have used over 1,000 chips, but full pre-training may require tens of thousands, and some cluster management software remains incomplete. Meanwhile, DeepSeek is still seeking more Nvidia chips.