The squeeze on AI computing power is falling hardest on the newest research labs, according to reporting by The Information.
These are the neolabs. Research-first outfits founded in the past two years, many of them by people who left OpenAI, DeepMind or Anthropic. Investors have put in sums of $50 million to $500 million before these labs have revenue, or in some cases a product. More than $10 billion has gone into the group.
They are discovering that money and compute are not the same thing.
What the shortage looks like
Computing power is not sold in one market at one price.
The largest buyers sign long contracts with cloud providers years ahead, reserving blocks of chips at set rates. The hyperscalers build their own. Everyone else takes what is left, at spot prices, or rents from third-party providers at a premium.
Neolabs sit in that last group. They have cash but no reserved allocation, no balance sheet to underwrite a multi-year commitment, and no purchasing history to bring a provider to the table. Capital alone does not move them up the queue.
Even the giants are rationing
The pressure shows up inside the cloud providers themselves. Amazon Web Services met engineers in May to clamp down on CPU waste, targeting internal instances left running at low utilisation.
That is a telling detail. When a company the size of AWS starts policing its own idle capacity, spare capacity has stopped being spare.
Add long lead times on hardware orders, rising rental prices and record capital spending by the hyperscalers. The scarce item is no longer chips in aggregate. It is chips available on a schedule you control.
Why timing matters more than price
Training a large model is not a job you can pause and resume without cost.
It needs thousands of chips held together for weeks, in one place, on a fixed schedule. Break the run and you restart or rebuild. Inference is a different problem but no easier. Serving a product means capacity that stays available every day, not capacity you win at auction.
A lab that can guarantee neither has a research plan it cannot execute.
The response so far
Some neolabs have paused frontier training. Others are paying above market for third-party GPU access, which burns runway raised on the assumption of cheaper compute. A third group has moved towards smaller and more efficient experiments, measured in months rather than years.
That last option is a research strategy in its own right, not a consolation prize. Plenty of the interesting recent work has come from architectural change rather than raw scale.
What it means for the shape of the industry
The pitch behind the neolab wave was that the next breakthrough would come from new ideas, not bigger models. Compute scarcity tests that claim early and hard.
If the ideas work at small scale, the funding thesis holds. If they need scale to prove out, the labs face a choice between waiting, paying up, or attaching themselves to whoever owns the infrastructure. That last route has form. Adept, Character.AI and Inflection all ended up inside larger companies.
For everyone watching from outside, the signal worth tracking is not the size of the next funding round. It is who has secured compute, and on what terms.
