GPU forensics suggest DeepSeek's real compute far exceeds disclosed figures
teortaxesTex · x · 2026-11-14
Analyst teortaxesTex reconstructs DeepSeek's true compute scale from multiple threads: Liang Wenfeng told investors on May 20 his fleet is 20K H-equivalent GPUs, with "most" arriving in the prior two months; Flash was served online since Feb 11 and checkpoints released Apr 24. Combining Will's forensics that V4-P used GB2/300 and V4-F used B2/300 (2K each) with multipliers of at least 2.5 and 3, plus the old 2K H800s unlocked to near-parity, that works out to 13/20—more if old 10K A100s or extra Hoppers are counted. He also cites Liang comparing Huawei's 16,000 950DT cards to GB300 at a 1:4 equivalence rate, and DeepInfer's disclosed inference throughput and revenue per GPU, suggesting at least 2.5-3K Grace Blackwells (almost certainly GB200 given reproducibility) by Oct 20. His conclusion: Liang likely didn't lie to investors or exploit technicalities.
More from Infra
- Running DeepSeek 4.1 Flash on a 32GB Mac via SSD streaming with open-source Local Code — alexcovo_eth · 2026-11-14
- SGLang Summit talk details battle-tested NVFP4 4-bit RL across thousands of GPUs — ericzelikman · 2026-11-14
- Will local models keep up with frontier? Users debate a hardware-bound gap — AshAtlasDev · 2026-11-14
- Anthropic IPO roadshow slide leak sparks Blackwell training speculation — and a block — teortaxesTex · 2026-11-14
- New Yorker interview: AI is a brute-force problem that will 'cook the planet' — kevinnbass · 2026-11-14
- Lenore unveils hybrid Token factory in Zhangjiakou: 15,000P cluster, 2 quadrillion tokens a year — 量子位 · 2026-11-14