Huawei compute limits leave 800B-model training far out of reach, industry source says
ShakeelHashim · x · 2026-07-23
A commenter quotes an industry call saying the gap with the U.S. is fundamentally about compute resources, not just talent or model capability.
- Training a model “as large as theirs” would require about 50,000 GB300s or Huawei 950s, or 200,000 cards.
- The speaker says current resources are only enough to do more experiments at around the 10B-active scale, and that training an 800B model is still far off.
- The bottleneck is framed as compute on both sides: cards can’t be bought domestically, and capital investment is lower than the U.S.
- A highlighted line says the problem is “basically unsolvable” right now because Huawei’s output is also limited, and training an 800B model would need 200,000 of Huawei’s newest cards.
Related event: DeepSeek's Liang Wenfeng on Compute Bottlenecks and US-China AI Gap(2 posts)→
More from Infra
- AI agents are turning web search into a new infrastructure market — Annual_Judge_7272 · 2026-07-23
- Google’s raised $200B capex guide still failed to reassure AI investors — JOBhakdi · 2026-07-23
- vLLM says prime-rl runs trillion-scale agentic RL on 28 H200 nodes — vllm_project · 2026-07-23
- GLM-5.2 adds vision support and is now open source, with SGLang run instructions — baseten · 2026-07-23
- Atomic says its quantized-model runtime cuts KV cache use by up to 6.4x — testingcatalog · 2026-07-23
- DeepSeek says it has about 20,000 H-equivalent compute cards and will keep buying NVIDIA GPUs — ShakeelHashim · 2026-07-23