. @dwarkesh_sp just made a massive bull case for the neoclouds $NBIS $AGPU $IREN $SHAZ $WYFI $CRWV

Lab revenue is 10x-ing while lab compute is only 3x-ing. Either margins go to the mid-90s, which he calls crazy and I agree, or compute gets 10-15x more expensive.

The labs can't defend margins that good. The chip is scarcer than the model is unique, so $ will go upstream.

Google is reportedly paying SpaceX 2x spot for 110,000 GPUs, and its going to get worse. People don't like the 1-1 spot vs contiguous training cluster gpu price comparison, but it shouldnt be ignored.

In the end, the labs will drive the price of compute up as a result of increasing revenue generation per chip, chip scarcity and cluster scarcity.

This will pump neocloud revenue as they are able to provide for both training clusters requirements (harder) and inference capacity (easier).

. @dwarkesh_sp just made a massive bull case for the neoclouds $NBIS $AGPU $IREN $SHAZ $WYFI $CRWV

Lab revenue is 10x-ing while lab compute is only 3x-ing. Either margins go to the mid-90s, which he calls crazy and I agree, or compute gets 10-15x more expensive.

The labs can't defend margins that good. The chip is scarcer than the model is unique, so $ will go upstream.

Google is reportedly paying SpaceX 2x spot for 110,000 GPUs, and its going to get worse. People don't like the 1-1 spot vs contiguous training cluster gpu price comparison, but it shouldnt be ignored.

In the end, the labs will drive the price of compute up as a result of increasing revenue generation per chip, chip scarcity and cluster scarcity.

This will pump neocloud revenue as they are able to provide for both training clusters requirements (harder) and inference capacity (easier).

Source:https://twitter.com/mrrydon/status/2082657373685231883