The Glut Everyone Is Waiting For

It is a fair question. The four largest US technology companies are on track to spend something like $700 billion on AI infrastructure in 2026. NVIDIA is doing roughly $75 billion of data-center revenue in a single quarter. Memory is sold out. Power contracts are being signed years ahead of the concrete.

Every investor who has lived through a capital cycle knows the ending. Everyone builds at once. Supply catches demand. Utilization falls. Pricing collapses. The last cohort of buildings sits half-empty while the depreciation runs anyway.

Railroads did it. Telecom fiber did it. A lot of that capital never came back, even though the technology did.

I want to put one number against that movie.

GPT-4-class output cost $60 per million tokens at launch in March 2023. Three years later the same quality of work runs about forty cents.

If intelligence were fiber, cheaper units would mean idle capacity. That is not what happened. Cheaper tokens pulled in work that was not worth doing at the old price.

That is the support. The rest of the argument, including why the largest buyer of the next unit of compute may be the machines themselves, is on the site.

Full Article: There May Never Be Enough Compute

A company can overbuild. That is not the same claim as civilization running out of useful things for cheap intelligence to do.

Thank you for reading.