Smash Notes

Back to Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO, with Erik Bernhardsson & Vibhu Sapra.

Why Are AI Hardware Prices Going Up?

Erik Bernhardsson says AI hardware prices are going up because capacity is tight while big buyers take as much supply as they can, even though hardware prices normally fall like gravity. He does not think GPU supply is a fully cornered resource.

NVIDIA is delivering several hundred thousand Blackwells a month. Paying a higher price than everyone else still gets GPU capacity, even as Anthropic and OpenAI buy up as much as they can, and supply keeps coming onto the market. The rise could last another six months or another 36 months. He likens the rush to buy capacity ahead to the great toilet paper shortage of the pandemic and to a run on the bank. Capacity will stay tight for a while. Reservations already locked in at a rate leave his company, to some extent, effectively long GPU prices, and scarcity means customers can be charged more.

He worries more about prices collapsing than about prices going up. He does not think a collapse will happen anytime soon, but it would be quite bad for anyone who locked in a lot of supply, because they paid too much. There is a risk of buying at the top.

Grokipedia·Anthropic·OpenAI

Transcript

Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO ·

Latent Space: The AI Engineer Podcast

More from this episode

  1. What Is Modal's Killer App?Erik Bernhardsson says Stable Diffusion turned out to be the killer app for Modal, the platform he built to make it easy to run code in the cloud.
  2. Why Did Modal Rebuild Its Tools for Generative AI?Erik Bernhardsson says Modal rebuilt its tools for generative AI because Kubernetes and Docker were made for a steady back-end workload.
  3. Why Won't Erik Bernhardsson Buy All the Compute He Can?Erik Bernhardsson will not buy up as much compute capacity as possible just because demand for intelligence looks basically infinite.
  4. How Do Neoclouds Expand So Quickly?Erik Bernhardsson says neoclouds expand quickly by pre-selling chips they have not bought yet.
  5. What Share of Revenue Do Modal Sandboxes Make?Erik Bernhardsson says Modal Sandboxes now make up about a third of revenue, and the team considers them the company's second product.
  6. Is Smaller AI Training Underserved?Erik Bernhardsson says smaller AI training is an underserved market. Huge training runs go to Neo cloud mega deals for 100,000 GPUs.
  7. Why Does Modal Snapshot Sandboxes?Erik Bernhardsson sees snapshotting as the way a long-running agent spins a sandbox up in 100 milliseconds, runs a few commands, shuts it down, and snapshots…
  8. Was Modal Hacked in the Hugging Face Incident?Erik Bernhardsson says Modal itself was not hacked in the Hugging Face incident.
  9. How Does Modal Treat Code Reviews?Erik Bernhardsson treats code review at Modal as optional.
  10. What Is the Best Use of New Cash at Modal?Erik Bernhardsson says the best use of new cash at Modal today is capacity prepayments. Modal is ramping up capacity very aggressively.
  11. How Many People Work at Modal?Erik Bernhardsson says Modal recently crossed 200 people and will probably hit 250 by the end of the year.
  12. Why Does AI Compute Come in Rectangles?Erik Bernhardsson says AI compute comes in rectangles because a typical GPU contract sells a fixed block of chips for about three years, starting three to six…

People

More answers like this

More with Erik Bernhardsson

More with Vibhu Sapra

Related episodes

Plays from this note through the rest of the episode.

Weekly highlights

One letter a week.

The notes worth hearing, with the minute they were said.