Scarcity turns the hyperscalers into gatekeepers

Around March, Google told Meta something it is not accustomed to hearing: no. Meta had gone to Google to buy more capacity to run Gemini, the models it leans on for everything from content moderation to internal coding, and Google, according to three people familiar with the matter, said it could not supply all that Meta wanted. The restriction remains in place. It has delayed some of Meta's internal AI work and pushed the company to tell staff to spend "tokens" (the units in which model usage is metered) more sparingly.

The detail worth sitting with is who got turned away. Meta has pledged to invest $600bn in the United States by 2028 and has spent the past year pouring money into talent and data centers in pursuit of what Mark Zuckerberg, its chief executive, calls "personal superintelligence." It is one of a handful of buyers on earth with a budget close to bottomless. And it was rationed anyway: harder, several people said, than any other Google client, because its appetite for Gemini was the largest.

Velvet rope

The consensus read of all this is scarcity, and the consensus is right as far as it goes. Demand for inference — the work of running models after they are trained — has outrun even the most aggressive buildout. What scarcity quietly confers, though, is discretion. When the machines are the binding constraint, the company that owns the machines decides who runs at the frontier and who waits in line. Google rationed several customers this spring and rationed its declared superintelligence rival the hardest. Whether or not that ordering was deliberate, the capacity to impose it is now a fact of the market. A supplier declining to sell more of a scarce good to a competitor is commercially unremarkable; that the good in question is access to the frontier is what makes the moment new.

For three years Washington has been building the other version of this power in plain sight. Export licenses, chip controls tightened repeatedly since 2022, an enforcement apparatus that treats advanced semiconductors as something closer to fissile material than to consumer electronics: all of it rests on a single premise, that frontier capability flows through compute, and that whoever governs compute governs who gets to build. The Gemini cap reveals a second gate the state did not build and cannot easily see into, a line item in a private company's capacity-allocation meeting, pointed at a domestic competitor rather than an adversary abroad. Brussels has been circling the same idea for a year; on June 25th the European Commission reached a preliminary finding that Amazon and Microsoft's cloud arms should be designated "gatekeepers" under the Digital Markets Act. The cap on Meta is that abstraction made operational.

The strangest part is that the gatekeeper is itself a tenant. Google is so short of capacity that in June it agreed to pay roughly $920mn a month to lease GPUs from Elon Musk's SpaceX, capacity meant to bridge the gap until its own buildout catches demand. Anthropic struck a comparable deal with SpaceX last month. A spot market in compute is forming, one in which even the largest hyperscalers arrive as buyers and whoever happens to be holding spare megawatts becomes, for a while, a landlord. At Alphabet's April earnings call, Sundar Pichai, Google's chief executive, said it without varnish: "Obviously, we are compute-constrained in the near term." The company's backlog of signed but undelivered cloud contracts had nearly doubled in a quarter, to more than $460bn: revenue it has booked and cannot yet ship.

Read against that backdrop, Meta's position is a lesson in the difference between owning and renting. Google allocates its own capacity to itself first; Meta, unlike its hyperscaler rivals, has no cloud business of its own and rents from a company it is simultaneously racing to beat. The $600bn now looks less like a growth bet than an insurance premium: the price of not being at a competitor's mercy the next time the meter tightens. In the meantime Meta has begun shifting workloads onto Muse Spark, an internal model, the better to stop asking a rival for permission.

Two gates now stand between a company and the frontier. One was built by a government, and a government can be lobbied, sued, or voted out. The other was built by a competitor, and answers to nothing more public than a quarterly capacity plan. Meta, for now, has been waved through the first and stopped at the second.