First TPUv7 Kimi K3 2.8T submission beats GB200/B300 on TCO at high interactivity

woosuk_k · x · 2026-10-19

SemiAnalysis reports that Inferact (makers of vLLM) and Google submitted the first preliminary TPUv7 inference submission for the Kimi K3 2.8T model.

Key points:

• In the ultra-high-interactivity regime, TPUv7 with a megakernel approach delivers strong performance per TCO against GB200 NVL72, B300, and MI355
• Chart shows TCO per chip/hour: TPUv7 $1.21, MI355X $1.5, B200 $1.73, GB200 $1.86, B300 $2.26, GB300 $2.31
• The submission was led end-to-end by Inferact using the TorchTPU vLLM stack, highlighting the value of a strong community for external compute platforms
• Data is an InferenceX official preview; results may change as validation continues

Original post →

More from Infra

• Local Qwen vision model builds a 3D room game in 26 minutes on a single prompt — Beneficial-Ad-8127 · 2026-10-19
• RTX 3090 runs 27B LLM 4.6x faster via pure inference optimizations, no new hardware — rickasaurus · 2026-10-19
• Microsoft scraps UK datacenter deal with Nscale, clouding its IPO hopes — GaryMarcus · 2026-10-19
• NVIDIA DGX Spark rumored to arrive in Europe at around €2,000 — AIFlow_ML · 2026-10-19
• BTIG flags bigger topping structure in SOXX, says rally is a chance to cut exposure — SumitGup · 2026-10-19
• ASUS Ascent GX10 (GB10) drops to ~$4,300 used on Amazon, 3 units left — AIFlow_ML · 2026-10-19

Infra channel →