Back-of-envelope math: Moonshot needs 25x more compute than DeepSeek to serve Kimi
teortaxesTex · x · 2026-10-24
teortaxesTex ran some estimates with Claude: Moonshot could achieve margins comparable to DeepSeek even with subscriptions, given Blackwell compute. The catch is serving Kimi K3 at the same volume as DeepSeek's V4.1F would require 25x more compute, at 60 tps vs 200+ tps — implying tens of thousands of Blackwells and billions of dollars, which he doubts will materialize.
Related event: Analysis Claims Kimi Serving Costs Over 25x DeepSeek's(4 posts)→
More from Venture
- Moonshot's Kimi slashes usage quotas, walks back after backlash — limits now roughly 6x higher — teortaxesTex · 2026-10-24
- Customers prepay billions for AI chip capacity: what the terms reveal about who wins — tengyanAI · 2026-10-24
- Nvidia pays multibillion-dollar advances to lock in substrate capacity through 2028, 2.5x in three years — tengyanAI · 2026-10-24
- Anthropic's binding payments estimated at $56bn in 2027, $86bn in 2028, covered by projected revenue — teortaxesTex · 2026-10-24
- DataFast Turns One: 8 Months to Hit 3K MRR, Now Steady at $30K — marclou · 2026-10-24
- Freepik went all-in on AI despite users hating it — and became Magnific in 3 years — cuenca · 2026-10-24