The economics of an AI neolab: 1,000 GB300s cost $125-150M for 3 years at ~2-2.5MW
deedydas · x · 2026-09-27
Investor Deedy breaks down the economics of a "neolab" — an AI-researcher startup raising big pre-production money to fund GPU compute — and invites challenges to his assumptions:
- 1,000 GB300s (14 NVL72 racks) run $125-150M over 3 years, with 15-30% upfront
- 2-2.5MW of power
- Enough for 10^25 FLOPs per quarter, roughly reaching a GPT-4-level model
Related event: Investor Breaks Down the Economics of AI Research Labs(3 posts)→
More from Venture
- Indie dev launches DeliveryKit, a Stripe-native tool for selling digital downloads — ow · 2026-09-27
- Mine your competitors' complaints: the highest-leverage pre-roadmap move for startups — _AustinCalvert_ · 2026-09-27
- Building WhatsApp AI receptionists for restaurants — founder seeks vertical feedback — vxdant23 · 2026-09-27
- 8-year agency vet: small business owners don't buy automations, they buy peace of mind — Warm-Reaction-456 · 2026-09-27
- Dreamina offers first-month Basic plan at $1.5, undercutting stock footage pricing — JaynitMakwana · 2026-09-27
- AI microdrama got 2.3M views in 24 hours; studio produces 50 episodes in 48 hours — Olumide_gbenro · 2026-09-27