Back-of-envelope: GB200 cluster should pretrain DeepSeek V4.1F in under two weeks
teortaxesTex · x · 2026-11-13
@teortaxesTex runs the compute math on rumored DeepSeek training runs: Inkling 2S finished its first pretraining stage in just 68 hours, and on a GB200 cluster V4.1F "shouldn't have taken more than 2 weeks" — versus V3's 55 days on 2048 H800s.
He reasons that if there really are multi-billion-dollar expenditures (with a 3-6x smuggling premium), the models could already be in the 200B active-parameter regime, with V4.1 Pro potentially much larger — the only question is whether it scales. Speculation based on public clues, not official info.
More from Infra
- Qwen3.8-27B quantized to run on 12GB RAM; router claims 97.5% of Opus 5.5 at 1/10 cost — Saul_Loveman · 2026-11-13
- Cerebras Q3: cloud revenue up 272%, 700MW data center capacity under contract — Sethwinterroth · 2026-11-13
- Critic faults chip export-control research: ByteDance lines up 36,000 B200s in Malaysia — nitarshan · 2026-11-13
- Tracked GPU acquisition likened to SR-71's Soviet titanium, sourced via hundreds of willing neoclouds — bookwormengr · 2026-11-13
- DynaDFlash: diffusion drafter boosts speculative decoding acceptance length by 5% — charles_irl · 2026-11-13
- Report: Nvidia planning frontier-class open models, sources claim — StewartalsopIII · 2026-11-13