K3 vs DeepSeek serving math: 21,600 GB200s needed vs DeepSeek's 2,400
YouJiacheng · x · 2026-10-25
Responding to "fakenews" about Kimi K3's serving efficiency, YouJiacheng clarifies: at EP64 and 77k average context (DeepInfer's published values), both K3 and DeepSeek V4.1 Flash can reach compute-bound batch sizes, even though K3's KV Cache is 16x larger. But the hard math remains: K3's active params are 6x V4.1F's and it lacks CED/YOCO efficiency designs, so serving DeepSeek-equivalent traffic would take 21,600 GB200s vs DeepSeek's 2,400 — cards Kimi doesn't have. His takeaway: efficiency targets must be baked into the backbone at training time; retraining on a new architecture takes too long to keep pace with monthly iterations.
Related event: Kimi K3 serving cost claimed 25x DeepSeek's, disputed by YouJiacheng(5 posts)→
More from Infra
- Bindu Reddy: GPU arbitrage isn't a technology or a defensible business — bindureddy · 2026-10-26
- S&P Global: Big Tech to spend over $7 trillion on data centers and AI capex by 2030 — Beth_Kindig · 2026-10-26
- Kimi K3's ~3T parameters spread across 59 Cerebras WSE-3T wafers, weights stay resident — petrusenko_max · 2026-10-26
- OpenAI's agent built the same site 8 times — every deployment landed on Cloudflare — FinanceYF5 · 2026-10-26
- DIY LLM-designed 3D-printed chassis housing four AMD Instinct v620 accelerators — PraxisOG · 2026-10-26
- Building AI infra may be easier than selling AI outcomes with clear ROI — seankinneyRCR · 2026-10-26