AI Compute Is Splitting Into Two Very Different Markets

The AI compute market is starting to split into two very different businesses.
Neoclouds like $CRWV and $NBIS are pricing access to the same Nvidia GPUs well below major hyperscalers such as $AMZN, $MSFT and $ORCL.
But price per GPU is only part of the equation.
Newer generations of AI chips still command higher hourly rates because they can deliver substantially more compute and inference output.
That means customers can pay more for each GPU while still seeing their cost per token decline.
This is why availability is becoming just as important as price.
AI labs want the newest hardware because better performance can materially lower the economics of training and inference.
The providers that can secure those GPUs and bring them online fastest therefore have significant leverage.
Hyperscalers have scale and ecosystem.
Neoclouds compete aggressively on price and specialized AI infrastructure.
And the newest GPUs command a premium because what ultimately matters is not cost per chip.
It is cost per unit of useful compute.