Kimi K3.1 Pro Day 0: SemiAnalysis TCO chart shows MI355X beating B200/B300 on tokens per dollar
AnushElangovan · x · 2026-11-07
Anush Elangovan (NVIDIA) posts "Day 0: K3.1 Pro" with a SemiAnalysis InferenceX benchmark of Kimi K3.1 Pro inference TCO (updated 2026-11-06).
Key points from the chart:
- At hyperscaler-volume ownership, AMD MI355X (vLLM) leads on total tokens per $1 TCO: up to 45M tok/$ at low interactivity, with the lowest TCO at $1.5/chip/hr;
- B300 (vLLM) runs at $2.26/chip/hr and B200 at $1.73/chip/hr, with B200 well behind at low interactivity;
- At high interactivity (220 tok/s/user P90), all chips converge to 4-5M tok/$.
The real story is K3.1 Pro's cross-vendor inference economics, with MI355X winning on cost-efficiency.
More from Infra
- Intel CEO: Agentic workloads could 5x server CPU TAM to $200B in 5 years — Beth_Kindig · 2026-11-07
- Personal AI agent rig: local 40 TOPS Arduino board + Tailscale/Cloudflare/Azure — unixterminal · 2026-11-07
- Big Tech capex up 5-8x since Q1 2023 while Apple spends 20% less; only Mag7 name beating Nasdaq — a16z · 2026-11-07
- Prefill-dominated endpoints coming for extraction and summarization workloads, researcher predicts — sh_reya · 2026-11-07
- SurrealDB 4 ships as 'the database for every agent' with PostgreSQL/MySQL support — mattturck · 2026-11-07
- India pledges indigenous GPU in 24 months; analyst says EUV access is the real clock — SumitGup · 2026-11-07