Moonshot would need ~59K GB200s and $2.6B/yr just to match one DeepSeek cluster
teortaxesTex · x · 2026-10-24
A detailed compute analysis shared by teortaxesTex lays out the asymmetry between Moonshot and DeepSeek:
- Matching one DeepSeek serving cluster (155T tokens/day) would take 39K–70K GB200s (820 NVL72 racks, 59K GPUs), $2.5–3.7B capex plus 110MW of IT load — or $2.6B/yr rented. That's most of Moonshot's reported $3–5B pre-IPO round plus a $5–10B IPO, and DeepSeek calls it 'a single cluster.'
- The real obstacle is logistics: smuggling 800 racks and securing 100+MW of power.
- Kimi's prices are 20–30× DeepSeek's; matching DeepSeek's volume at DeepSeek prices is impossible because K3 loses money there.
- Reversed, the same fleet lets DeepSeek serve that volume with 2,400 GPUs, leaving 56K spare: a K3-class pretrain in 4 days, a next-gen 300B-active run in 18 days, plus 35T daily RL rollout tokens (5× its current output).
- DeepSeek's architecture converts a GPU into 25× more served tokens — compute spent serving is compute not spent training, which is why 'the sorting will accelerate.'
Related event: Analysis Claims Kimi Serving Costs Over 25x DeepSeek's(4 posts)→
More from Infra
- Claude API Adds Workload Identity Federation to Kill Static API Keys — brada · 2026-10-24
- LLM inference may be shifting from batch to sequence parallelism, setting up diffusion LMs — sedielem · 2026-10-24
- RTX 5090 PSU clicking noise only in ComfyUI Moss Image workflow — 900mV undervolt fixes it — Garito10 · 2026-10-24
- antirez makes the case for local AI as hacker's Samsung TV snooping saga hits HN — antirez · 2026-10-24
- PNG is slow because of deflate, not libpng — zlib-ng's SIMD wins — ssh4net · 2026-10-24
- Micron roadmap: HBM4 at 2.8 TB/s per stack, vertical HBM on GPUs by 2028 — rohanpaul_ai · 2026-10-24