In late May at Meta’s AGM, an investor asked Zuckerberg whether Meta would build a cloud to rival AWS. He said no, they had a use for the compute, though if they ever overbuilt, leasing it out was an option, and that option was part of what gave them the nerve to keep spending. A reasonable hedge. Five weeks later it stopped being a hedge. The exchange is below.
On 1 July, Bloomberg reported Meta is building a cloud business to sell its excess AI capacity, a unit being called Meta Compute. The stock jumped almost 9%. The plans are reportedly early and the shape is still open, whether Meta sells access to its own models the way Amazon’s Bedrock does, or rents out raw GPUs the way CoreWeave does.
The excess-compute story broke on 1 July. A day later, at an internal town hall, Zuckerberg told staff that AI agent development hadn’t accelerated the way he expected over the past four months, and that the restructuring built around it hadn’t come good yet. Put the two together and you have a company that spent the past year chasing chips now sitting on enough spare to sell, while its CEO concedes the AI work is running behind.
The easy conclusion is that this is bad news for everyone carrying a giant AI bill. I don’t see it that way, or at least not evenly spread. If the supply and demand balance slips from here, the hyperscalers aren’t the ones I’d worry about. It’s Oracle and the neoclouds, the ones leveraged up to the eyeballs, that carry the real risk.
Look at the pecking order today. Microsoft is the clearest example: short on GPUs, it leases capacity from the neoclouds and resells it under the Azure brand at a premium to enterprises that will only buy from a name they already trust. Not everyone works this way, AWS mostly builds its own, but where it happens that premium is the hyperscaler’s buffer. The neocloud is the one sitting on the leased hardware and the debt behind it. When someone like Meta turns up with its own surplus to sell, the squeeze lands on the players without the brand or the balance sheet to absorb it.
That’s the risk, not a certainty. The plans are early, and there may be less spare than the headline implies. GPU shortages are still the norm, and by several analysts’ read Meta is still short on compute itself. Zuckerberg says he expects the AI spending to start paying off within three to six months, and if Meta’s own workloads grow into the capacity, there’s nothing to dump.
Even so, you could see it in the tape the day the story broke. CoreWeave and Nebius both fell double digits and IREN dropped too. That’s the market working out who’s actually exposed. A hyperscaler can be wrong about demand for a good while and be fine. The neoclouds don’t have that room.
Compound Steadily,
Thomas
P.S. My book doesn’t ask you to guess who wins the compute arms race. If owning a few good businesses in an hour a day sounds more your pace, The Lunch Break Investor is up for pre-order before it lands 18 August: steadycompounding.com/book.