$2 H100s: How the GPU Rental Bubble Burst - by Eugene Cheah
Swyx’s note: we’re on a roll catching up with former guests! Apart from our recent guest spot on Raza Habib’s chat with Hamel Husain (see Raza’s first pod here).
We’re delighted to welcome Eugene Cheah (see his first pod on RWKV last year) as a rare guest writer for our newsletter . Eugene has now cofounded Featherless.AI, an inference platform with the world’s largest collection of open source models (~2,000) instantly accessible via a single API for a flat rate ($10-$75+ a month).
Recently there has been a lot of excitement with NVIDIA’s new Blackwell series rolling out to OpenAI, with the company saying it is sold out for the next year and Jensen noting that it could be the “ most successful product in the history of the industry”. With cousin Lisa hot on his heels announcing the MI3 25 X and Cerebras filing for IPO, it is time to dive deep on the GPU market again (see also former guest Dylan Patel’s pod for his trademark candid take on the industry of course):
Do we yet have an answer to the $600bn question? It is now consensus that the capex on foundation model training is the “ fastest depreciating asset in history”, but the jury on GPU infra spend is still out and the GPU Rich Wars are raging.
