The economics of a Neolab: $125-150M in GPUs buys you GPT-4 level, two orders off frontier
AccBalanced · x · 2026-09-27
Investor deedydas defines a "Neolab" — an AI-researcher startup raising pre-product money to fund GPU compute for a frontier-scale problem — and runs the numbers:
- Compute cost: 1,000 GB300s (14 NVL72 racks) runs $125-150M over 3 years with 15-30% upfront, drawing 2-2.5MW.
- Output: that's 10^25 flops per quarter, enough for a GPT-4-level model — still 1-2 orders of magnitude off the pretraining frontier.
- Post-training route: fine-tuning a strong open model offers a better shot at frontier, but requires millions more in RL environments and risks being lapped by the next release while tied to someone else's base model.
anshublog adds nuance: the Anthropic investor calling Neolobs "dead man walking" has a point, but Anthropic itself was a Neolab relative to OpenAI four years ago — "a 2.7% chance at a trillion dollar outcome will make people write billion dollar checks."
More from Venture
- CoreWeave: 2027 GPU renewals could add ~$4B annualized revenue, analyst skeptical — zephyr_z9 · 2026-11-11
- Polymarket gives 48% odds Anthropic's IPO closes above $1.5T market cap — Polymarket · 2026-11-11
- Anthropic's IPO reportedly oversubscribed multiple times, per Polymarket — Polymarket · 2026-11-11
- Indie-hacking drama: should builders monetize their audience openly? — eptwts · 2026-11-11
- Investor argues analogy-based AI capex bubble takes underestimate LLMs' self-induced demand — menhguin · 2026-11-11
- Nvidia's supply commitments jump from $119B to $279B in one quarter, locking up 2027-28 memory — tengyanAI · 2026-11-11