Breaking

DeepSeek Pushes Toward Huawei Chips as U.S. Controls Squeeze Nvidia Access

September 21, 2026 at 15:33 EDT

Infra & Chips
Foundation Models

DeepSeek is moving to deepen its use of Huawei processors as U.S. export controls keep advanced Nvidia chips difficult to obtain in China, making domestic AI hardware a more central part of the company's training roadmap. The shift, reported by The Information, centers on CEO Liang Wenfeng's expectation that Huawei processors will play a growing role in training DeepSeek's largest models.

September 2026 · DeepSeek

DeepSeek leans into Huawei chips

The company is pushing to reduce dependence on Nvidia as export controls make advanced U.S. chips harder to secure in China. The next test is whether domestic hardware can move from serving and post-training into reliable frontier training.

China AI-chip share is shifting

40%

Nvidia 2025

Nvidia forecast

50%

Huawei forecast

Each block represents roughly ten percentage points, showing the domestic swing in China's AI-chip market.

2.788M

H800 GPU hours reported for DeepSeek-V3 full training

34.22%

MFU reported for DeepSeek-V4 post-training on Ascend infrastructure

Source-level patches needed in an Ascend inference field study

The migration path

Nvidia training base

DeepSeek-V3 relied on H800 compute.

Ascend serving gains

Huawei systems show optimized DeepSeek-R1 serving.

Post-training proof

DeepSeek-V4 family post-training reached 34.22% MFU.

Frontier training test

Reliability, tooling and interconnects remain the key hurdles.

Why developers see promise

DeepSeek's efficiency record and Huawei's faster hardware cadence suggest more workloads can be tuned for domestic AI infrastructure.

Why caution remains

Ascend migrations still face software maturity, operator coverage, parallelism and device-stability issues.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.