RUMOR: Huawei engineers reportedly work inside DeepSeek on Ascend chips Other

The Information reports that Huawei engineers are working inside DeepSeek's offices to shape future Ascend AI chips, and that DeepSeek plans to start training models on the Ascend 950DT as soon as the end of 2026. Neither company has confirmed this.

This is not verified. The co-design work, the training timetable and the chip allocation all come from The Information's anonymous sources, published on October 19, 2026. Huawei and DeepSeek have not commented publicly. What is confirmed, in public code, is the software side: DeepSeek published Ascend versions of its matrix and MoE communication libraries on GitHub on September 30, 2026.

What the report says

• Engineers on site. The Information reports that Huawei engineers, some of them stationed at DeepSeek's offices, are building DeepSeek's requirements for how models compute and move data between processors into chips due next year and later.
• Training run. The Information reports that DeepSeek plans to begin using Ascend 950DT chips for model training as soon as the end of 2026.
• Scale so far. The Information reports that tests have involved more than 1,000 chips, while a full run could need tens of thousands, and that some software for managing a cluster that large is unfinished.
• 960DT supply. The Information reports that Huawei plans to direct most of its early Ascend 960DT output to DeepSeek when that chip arrives in the first quarter of 2027.
• Earlier attempt. The Information reports that DeepSeek tried to train its R2 model on Huawei chips last year and moved to Nvidia hardware after problems with chip stability, chip-to-chip communication and software support.

What is confirmed in code

• DeepGEMM-Ascend. DeepSeek's port of DeepGEMM to Huawei Ascend. Its README records the initial release on September 30, 2026, with support for Ascend 950 devices. It is API-compatible with DeepGEMM and covers BF16, FP8 and FP4 GEMM, MQA logits and MegaMoE.
• DeepEP-Ascend. A communication library for Ascend NPUs, created on GitHub on September 30, 2026. It provides expert-parallel all-to-all dispatch and combine for MoE models, with FP8 dispatch.
• TileLang for Ascend. TileLang-Ascend is a project of the tile-ai organisation, not DeepSeek. Its README dates the open-source release to September 29, 2025, and lists DeepSeek V4 kernels added on April 24, 2026.
• Hardware used. The DeepEP-Ascend README states its measurements were taken on Ascend 950DT NPUs with CANN 9.2.0, and that the performance runs used a proof-of-concept hardware kit supplied to DeepSeek rather than a public release.

DeepEP-Ascend bandwidth on Ascend 950DT

Figures from the DeepEP-Ascend README: 16,384 tokens per rank, hidden size 7168, top-6 routing over 256 experts, FP8 dispatch and BF16 combine.

Sources

• The Information: How DeepSeek Is Shaping Huawei's AI Chips in a Growing Challenge to Nvidia
• The Information: DeepSeek Open-Sources Software Tools for Huawei AI Chips
• GitHub: deepseek-ai/DeepGEMM-Ascend
• GitHub: deepseek-ai/DeepEP-Ascend
• GitHub: tile-ai/tilelang-ascend