DeepSeek reportedly bets on Huawei chips to train next-gen models; Liang says it 'has to work'
kimmonismus · x · 2026-09-22
According to The Information, DeepSeek is betting on Huawei chips to train its next AI models as U.S. export controls restrict access to Nvidia hardware.
- CEO Liang Wenfeng told investors he expects new Huawei training chips in Q4 2026 or Q1 2027, saying the push to train on Chinese processors "has to work."
- DeepSeek is already training a 2-trillion-parameter model and plans an 8-trillion-parameter model.
- The company still uses Nvidia chips; Huawei faces shortages of advanced memory and other components, and Liang previously said Huawei could not supply enough processors.
DeepSeek's ambitions now partly hinge on how quickly Huawei can deliver training hardware at scale.
Related event: DeepSeek Bets on Huawei Chips for Next-Gen Model Training(4 posts)→
More from Infra
- Transformers now runs llama.cpp GGUF quants via ggml kernels, faster local inference on Mac — LysandreJik · 2026-09-22
- MCIO x16 breakout boards tested: 2-slot width cuts risers from 4-8 down to 2 — TheZachMueller · 2026-09-22
- Meta rumored to partner with Oracle to bring Meta AI Platform to Oracle Cloud users — testingcatalog · 2026-09-22
- Astorias AI launches budget inference service with Qwen at $0.1/M input tokens — me_broke · 2026-09-22
- Osborne, now at OpenAI, says UK datacentre NIMBYs threaten Britain's AI sovereignty — nordicinst · 2026-09-22
- What are KV caches really? A storage expert's explainer from prefill to offload — TheZachMueller · 2026-09-22