GenAI's $600 Billion hole
A Sequoia VC peers into Nvidia's revenue & finds something troubling
David Cahn from Sequoia did some basic math on Nvidia’s data center revenue. Here’s the breakdown
Nvidia made: $150B (from datacenters alone!)
It costs Nvidia’s customers (Goog, Meta, Amzn, MSFT) ‘2X the chip cost’ to run the datacenter
So their total expenses: $300B
Amzn/Goog/Meta/Msft need to make profits. Assume 50% gross margin. So 2Xing this again.
Total needed to recoup: $600B
Big question - Who’s making all this money in Gen AI?
He does some generous math and still can’t fill this $600B hole.
His conclusions therefore
there is a massive amount of over-capacity build-up in infra.
There doesn’t seem to be enough killer apps that can make up the $600B.
Ergo, a lot of stock valuations are going to take a hit.
An easy argument to make against him would be
The 2000s saw the same overcapacity in fiber-optic cables
1850s saw the same overcapacity in train tracks
GPU overcapacity is similar. It will be monetized eventually
And herein his counter-point
Datacenters are very different from Railway tracks.
Why?
Physical infra like a railway track from NYC to SF has intrinsic value. As you can’t just lay new tracks between 2 cities that easily.
Nvidia’s B100 will have 2.5X better performance over the H100. Hence older datacenters will depreciate faster
While he believes strongly that AI will be a huge disruption, “Capital Incineration” is a way of life during ‘speculative investment frenzies.
Meanwhile, a top Goldman Sach’s economist has a similar analysis. His numbers are even more startling - $1 Trillion of capital investment needs to recouped.
Great reads
Goldman’s Top Stock Analyst Is Waiting for AI Bubble to Burst
AI’s $600B Question (David Cahn’s latest analysis)
AI’s $200B Question (David Cahn’s older analysis from last year)