The hardest part of Neolab startups isn't training models — it's the math

FinanceYF5 · x · 2026-10-28

A thread on the brutal economics of "Neolab" startups — top researchers raising huge capital before any product, betting on GPUs to crack big AI problems:

Once you buy compute, an expensive countdown begins. Idle GPUs burn money; teams resell compute, serve open-model inference, or sell spot instances — below 60% spot utilization, they still lose money.
To recoup investment, they must beat frontier models on quality and ideally price. Even at 50% inference gross margin, at $2/M tokens, recovering just $10M of training cost requires serving 10 trillion tokens.
Post-training on strong open models is the likelier path to the frontier, but RL environments cost millions more, and work can be invalidated overnight when a new base model ships.
Escape routes: avoid the general-model race, pick directions big labs won't enter, or build proprietary data moats in robotics/bio/chemistry (e.g. Periodic Labs).

Original post →

The hardest part of Neolab startups isn't training models — it's the math

More from Venture