Why Is One Neolab Spending a Billion Dollars on a Single Training Run?

Ramtin Naimi says a neo lab he met a couple of weeks ago was raising $1 billion in a financing he thinks was its second round, and planned to spend the entire billion next year on compute for one training run. The company said it would probably need to raise another $5 billion throughout next year.

People once expected two or three great neo labs. The count is now probably 80 or 90 neo labs that have raised somewhere between $1 billion and $5 billion. The segment is still small beside the hyperscalers, and the combined scale is not small. All of them need to raise tons of cash, and underwriting how well they are doing is very hard. The practical read often tracks how well they are doing and how much money they are raising. They often launch very impressive work, and it is hard to tell a result truly impressive on a generalizable scale from one tuned to perform really well on benchmarks. Benchmark maxing is the split between a launch that is truly groundbreaking and innovative and one that only shows well on the benchmarks and is not commercializable for the near future.

A lot of the neo labs have massive capital needs. Whether all of them continue to raise, whether 50 percent continue to raise the capital they need, or whether about 10 percent do is still open. He leaves open whether capital is the actual constraint, so that a dry-up in the capital markets would be what changes the compute story.