Excess-Capacity Cloud Arbitrage

The AWS problem

Brad Gerstner names the structural reason that anyone who builds enormous compute is pushed to rent it out, and why consumer-native companies keep ending up in the enterprise and cloud business: "The second you start spending $100 billion on capex annually, you run into the AWS problem. Now I have all this compute, but I don't use it every day equally. Jeff Bezos built AWS because he had to build capacity for Christmas, Black Friday, and the rest of the year half of it sat idle, so he rented it to everybody else. It turned into a blockbuster business and made his core business better, because he could build to Black Friday and nobody else could."¹

The mechanism produces two payoffs from one buildout. Renting peak-sized capacity that sits idle most of the time converts a cost center into a profit center, and because the rental business helps pay for peak capacity, the owner can afford to build to its own peak in a way competitors cannot, a durable advantage in the core business.

Elon Web Services and Meta

Gerstner applies the pattern to 2026. Elon Musk's xAI has effectively launched "Elon Web Services": "Nobody on earth is better at turning electrons into tokens than Elon." xAI signed Anthropic as a major compute customer, and that deal, alongside one with Cursor, "changed the whole tenor of the SpaceX IPO" from mild concern to excitement.¹

Meta is being pushed into enterprise for the same structural reason, despite its consumer-native advantages. At over 100 billion dollars of capex, Meta hits the same AWS problem and must monetize excess capacity, even though taking a heavily consumer-oriented company into AWS-style infrastructure and enterprise agents is, in Gerstner's word, hard.

The economics, made specific

A BG2 episode two days before the SpaceX IPO filled in numbers that earlier analysis had only estimated. The xAI-Google compute deal runs at roughly 50 billion dollars per gigawatt per year in operating profit, the highest of any major AI cloud agreement globally, and the xAI-Anthropic deal runs at roughly 22 to 23 billion dollars per gigawatt per year, also above any competing offer. Both figures run 1.5 to 3.5 times the roughly 14 billion dollar per gigawatt per year rate implied by bank models for a 160 billion dollar 2028 revenue forecast.²

Other data points from the same period: a 55 percent internal rate of return calculated on Colossus 1, straightforward capital allocation math against 6 to 8 percent borrowing costs; 122 days to build what was, at the time, the world's fastest supercomputer at 100,000 GPUs, against a normal timeline of three to four years; and reaching the position of fourth-largest hyperscaler within 30 days of the Google deal, passing Oracle and all neoclouds. The broader dynamic is described as "accidental profitability": the labs expected to be nowhere near breakeven at this stage of the buildout, and actual monetization has far exceeded their models, with demand outstripping supply and pushing monetization per gigawatt up rather than down.²

Why it matters

The pattern predicts a convergence: every hyperscale compute owner eventually competes in cloud and enterprise, regardless of where it started, because the capex math forces it. It also reconciles an apparent contradiction with the idea that there are no dark, unused tokens in the system: capacity is idle only relative to a single buyer's peak, not in aggregate, and the moment it is offered to the broader market it clears instantly rather than sitting unused.

Tensions

Whether the core-business advantage survives once everyone does it is an open question: if every major compute owner rents its excess, the edge of being the only one who can build to peak erodes into a commodity compute market. And the consumer-to-enterprise transition is culturally hard in a way the capex math does not capture; Gerstner's framing explains the incentive to make the move, not the difficulty of executing it.