Colossus 2's 1.2M B200s could generate ~$400B/year in tokens at DeepSeek Flash rates
bookwormengr · x · 2026-11-16
In part 3 of his compute-economics thread, bookwormengr estimates that at DeepSeek V4.1 Flash off-peak rates and with a suitable architecture, each B200 could yield $38/hour in token revenue. With 1.2M B200s at xAI's Colossus 2, that's $45M/hour, $1.09B/day, and nearly $400B per year of inference output. At Anthropic's rates (roughly 4x higher), the theoretical ceiling would be $1.6T. He notes this requires a specialized model architecture, but argues most labs will adopt it given the profound gains.
More from Venture
- Anthropic's amended S-1 shows slowing growth; full-year revenue seen at $55-57B — bookwormengr · 2026-11-16
- AI compute stocks keep sliding premarket: Nebius -6.9%, CoreWeave -6%, Oracle -3.2% — GaryMarcus · 2026-11-16
- Freshworks alums raise $53M for AI-native CRM to challenge Salesforce, HubSpot — TechCrunch AI · 2026-11-16
- Did AEO really diverge from SEO? QueryBurst audits Profound's 33 recommendations — gaganghotra_ · 2026-11-16
- IDC's 70-page report asks why 350+ national AI models outside US and China failed to compete — perilli · 2026-11-16
- Priors Lab launches agent credit lines and token launchpad on Robinhood Chain — econoar · 2026-11-16