DeepSeek Training Math: 2-3K GB200s Would Suffice Even for Rumored 2T-Param V4.1P
teortaxesTex · x · 2026-11-13
Following up his earlier thread, teortaxesTex notes DeepSeek may be experimenting on Huawei Ascend or even training the next generation there — "they clearly intend to in the next few months." But by his estimate, for V4.1F-scale models and even the rumored 2T-parameter V4.1P, 2-3K GB200s would be enough, and DeepSeek has them.
This squares with DeepSeek's own account of holding 20,000 H-equivalent GPUs and buying more whenever prices are reasonable: NVIDIA hardware remains its training backbone.
More from Infra
- kagent 1.0 GA: Agent Substrate brings sandboxed AI agent execution to Kubernetes beyond pod-per-agent — bibryam · 2026-11-13
- 4x R9700 EXL3 serving: 1,094 tok/s Qwen decode, all optimization tricks open-sourced — mikoskinen · 2026-11-13
- Sweden's all-wood EcoDataCenter: 1.15 PUE, just 1.17g CO2/kWh, zero water use — filwi · 2026-11-13
- Strata EXL3 AMD fork gains 10-20% tok/s, runs 3bpw at IQ4-level quality — nasone32 · 2026-11-13
- Swapping one month of Scaleway bills for a 16GB ThinkCentre to run agents locally — jiayuan_jy · 2026-11-13
- From reform-school toilets to $5T: a timeline of Jensen Huang and NVIDIA — tibo_maker · 2026-11-13