$125M for 1,000 GB300s: The Brutal Compute Economics of a 'Neolab'
deedydas · x · 2026-09-27
Deedy Das breaks down the economics of a Neolab — a startup of AI researchers raising big money pre-production to finance GPUs against a large problem:
- Compute cost: 1,000 GB300s (14 NVL72 racks) runs $125-150M over 3 years with 15-30% upfront, 2-2.5MW, enough for 10^25 FLOPs per quarter — reaching only a GPT-4-level model, 1-2 OOMs off frontier pretraining.
- Path: post-training on a great open-source model gives a better shot at frontier, but RL environments cost millions more and you risk being lapped by new releases.
- Payback: even at 50% inference margin and $2/M blended pricing, recouping $10M in training means serving 10T tokens — while outcompeting cheaper releases like Opus 5.5. Idle GPUs burn money; below 60% spot utilization you still lose.
- Exits: don't play the model game; play one big labs won't (Jev, World Labs); or acquire a proprietary dataset at scale (Periodic Labs), common in robotics, biology, chemistry. A difficult game overall.
More from Venture
- Indie label built Pitched: AI listens to your track and writes your Spotify pitch in 20 seconds — demian_ai · 2026-11-11
- Similarweb: Claude logs record 1.253 billion website visits in October — gaganghotra_ · 2026-11-11
- a16z Podcast: Mike Volpi on What the 1990s Internet Boom Teaches Us About the AI Buildout — a16z Podcast · 2026-11-11
- Alibaba files to list chip-design unit T-Head in Hong Kong to fund AI push — pstAsiatech · 2026-11-11
- First traffic snapshot of Muse, Instinct, Hark and Fo: 60M visits in October, Muse leads with 45M — agentaya-ltd · 2026-11-11
- AI killed SEO writing moats; first-hand data is the real edge — sven_ai · 2026-11-11