Back-of-the-napkin estimate breaks down Kimi K3 training cost
yaroslavvb · x · 2026-07-22
A back-of-the-napkin estimate breaks down the likely training cost of Kimi K3 using chip counts, grid power, and rental-plus-energy costs.
The accompanying chart compares multiple GPU/accelerator options and their estimated costs, highlighting how the economics of large-scale model training depend heavily on hardware availability and power consumption.
More from Infra
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27