Kimi 3.1 on Vera Rubin costs just 0.025 cents per million tokens, per TCO math
bookwormengr · x · 2026-10-28
bookwormengr ran the numbers on inference economics: on NVIDIA's Vera Rubin, Kimi 3.1 works out to 0.025 cents per million tokens of TCO — $1 of compute yields 45M tokens, roughly $90 worth of API-priced output, implying huge margins and room for price cuts. DeepSeek v4.1 Flash could see similar uplift from Grace Blackwell, from 1.6B to 4B tokens per dollar of TCO. The author argues frontier labs may stay wildly profitable through a compute glut. Quoted @tszzl is more bearish: frontier lab equity value is unknowable, and slowdowns or even nationalization are plausible.
More from Venture
- Juan Benet launches PLC Neurotech fund, says neurotech is at AI's 2019 moment — niloofar_mire · 2026-10-28
- Anthropic Taps Goldman Sachs and Morgan Stanley as IPO Lead Underwriters — rohanpaul_ai · 2026-10-28
- Anthropic IPO: Goldman lead underwriter, Goldman and Morgan Stanley could each earn ~$100M — rohanpaul_ai · 2026-10-28
- MicroAGI hits $100M ARR fastest ever; candidates must work 48-72 paid hours to get hired — lukas_m_ziegler · 2026-10-28
- LinkedIn's Feed Now Uses LLMs to Suppress Clickbait, B2B Marketers Told to Name Their Reader — shashib · 2026-10-28
- NYC's 'fastest-growing' early startup adds $260K net new ARR in a single day — julianweisser · 2026-10-28