Xiaomi's Livestreamed RL Training Burns ~$10/sec, Est. 6,000-8,000 H100s
karminski3 · x · 2026-09-17
Blogger karminski3 analyzed Xiaomi's livestreamed training run, concluding MiMo-v2.6-pro and flash are in the Agentic RL post-training phase, benchmarked on DeepSWE (Pro at 62 vs DeepSeek-v4.1-flash's 74.2 — still early).
His back-of-envelope math:
- Stage 10 generated 2.2B tokens in 58 minutes (630K tokens/sec) at 89K mean context length
- At 100-150 tps per H100, that implies 4,000-5,000 H100s for Pro decode plus 2,000-2,500 for flash — roughly 6,000-8,000 total
- Timing logs (58 min generation + 64 min backward pass ≈ 126 min/step) suggest the same GPUs flip between rollout inference and long-sequence backprop
- Over 60,000 concurrent Linux/Docker sandboxes (23,180 for Pro, 37,786 for flash) explain the $10/second burn rate
Related event: Xiaomi live-streams MiMo-V2.6 RL training, burning over $1.08M in 36 hours(14 posts)→
More from Infra
- Tailscale called a 'single point of failure' for startups as acquisition chatter splits devs — TheZachMueller · 2026-11-02
- Gary Marcus: AI compute is headed for a glut, just like fiber in the dot-com bust — GaryMarcus · 2026-11-02
- One H200 Isn't Enough for DeepSeek 4.1 Flash: Two GPUs at $10.59/hr — daluoseo · 2026-11-02
- Databases to distributed systems: every systems skill is merging into AI inference and hardware — Abhishekcur · 2026-11-02
- AWS not "sold out" until 2027: expert says capacity headline needs context — DavidLinthicum · 2026-11-02
- Foxconn chairman Liu: AI is not a bubble, demand for AI products remains clear — pstAsiatech · 2026-11-02