On July 1, Bloomberg reported that Meta is preparing to rent AI computing power to outside customers. Investors loved it. The stock gained 9% that day and nearly 17% over the following two weeks.
The conclusion seemed obvious. Meta is spending between $125 and $145 billion on AI infrastructure this year with almost no direct revenue attached to it. If some of that infrastructure can be rented out, an enormous cost center suddenly becomes a new business.
The comparison everyone reached for was Amazon.
Amazon built AWS. Meta, the argument goes, is about to build its own version.
It is an elegant story.
It is also almost certainly the wrong one.
The mistake comes down to a single word that appeared in Bloomberg’s reporting and has been repeated ever since.
Excess.
AWS was infrastructure Amazon intentionally built to sell.
Meta is talking about infrastructure it built for itself and no longer needs.
Those are not the same business.
More importantly, they are not the same signal.
One tells you demand exceeded expectations.
The other tells you expectations exceeded demand.
That distinction sounds semantic.
It isn’t.
It completely changes how investors should interpret Meta’s AI strategy, future depreciation, cloud economics and, ultimately, the companies likely to benefit—or suffer—from the next phase of the AI infrastructure cycle.
Before the July 29 earnings call, there are two facts sitting in plain sight.
The first is public guidance.
Meta expects to spend between $125 and $145 billion of capex this year, almost double last year. After spending just $19.8 billion in the first quarter, that guidance implies quarterly investment accelerating dramatically through the rest of the year.
The second is also public.
Meta has begun talking about renting out unused compute.
Viewed separately, neither fact is surprising.
Viewed together, they create a contradiction.
And it gets stranger.
Four months before the surplus-compute story appeared, Meta committed up to $27 billion to buying even more AI capacity beginning in 2027.
Companies that have too much compute generally don’t commit another $27 billion to buying more.
Unless...
...the compute they are selling is not the compute they are still buying.
That possibility changes almost everything.
It changes what investors should listen for on July 29.
It changes how AI infrastructure should be valued.
And it changes which companies are actually exposed if hyperscalers stop behaving like buyers and start behaving like sellers.
That is what this edition is about.
In the full research edition, I explain:
Why the market’s AWS comparison breaks down once you distinguish planned capacity from surplus capacity.
Why Meta’s March purchase agreement and July rental announcement are not contradictory—and what they reveal about the specific part of Meta’s infrastructure that may already be economically obsolete.
Why the first observable rental prices could become more important than any depreciation assumption buried in the financial statements.
The historical precedent almost no one is discussing, and why it ended with collapsing prices rather than a profitable new industry.
The four questions I’ll be listening for on Meta’s July 29 earnings call—and how each answer changes the investment case.
Most importantly, which companies stand to lose pricing power if hyperscalers become the industry’s largest suppliers of AI compute, and the small group that could actually benefit.
If this thesis is correct, the implications extend far beyond Meta.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.