The Biggest Data Center on Earth Comes With an Exit Hatch

Meta is building the largest AI data center in the world: $27 billion in Holly Ridge, Louisiana, with ten new gas-fired power plants and operations expected by 2030 ^[1]. At the same time, the company is developing a cloud division to sell excess AI compute to outside customers ^[2]. These are contradictory bets. One wagers that AI demand will be vast enough to fill the biggest single infrastructure project in Meta's history. The other wagers that it might not be. Mark Zuckerberg has described the compute-selling business in exactly those terms.

It's definitely on the table. — Meta

The conditional matters. Every available market signal says that point has not arrived, and is not close. Northern Virginia, the world's largest data center market, has a vacancy rate of 0.3% ^[3]. Jefferies reports a 12 GW gap between the data center capacity operating globally and what customers are demanding, with a $2 trillion cloud service backlog ^[4]. CoreWeave, a major neocloud provider that also supplies Meta, says it has essentially no 2026 capacity left to sell ^[2]. Amazon is investing $200 billion in data center infrastructure this year, with the company reporting that customers have already lined up to use most of the new capacity before it arrives ^[5]. Record inventory growth is being absorbed the instant it comes online. The supply side is physically constrained, too. Satellite data suggests roughly 40% of U.S. AI data center projects may miss their 2026 completion dates by more than three months — not because tenants have backed out, but because of specialized labor shortages, tariffs on Chinese transformers, and utility infrastructure that cannot keep pace ^[6]. The bottleneck is construction, not demand. On the demand side, something subtler is happening. Enterprises are not retreating from AI. They are pushing back on the cost of running it. PNC CEO Bill Demchak has been the most direct about what that looks like from inside a large company.

Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens. — Bill Demchak

Companies are shifting to smaller AI models and building their own GPU compute to avoid paying per-token premiums to cloud providers ^[7]. This is an efficiency push, not demand destruction. But it is the kind of thing that, if it compounds across thousands of enterprises, could bend the demand curve by the time the delayed construction wave finally arrives. That timing mismatch is what Meta's hedge is really about. The supply currently under construction — delayed, bottlenecked, staggered — will eventually come online, potentially in a concentrated wave. If, by then, enterprises have refined their models enough and built enough of their own compute to need less from the cloud, the market could overshoot after all. Not on the numbers in front of anyone now. But the numbers in front of anyone now are not what Meta is hedging against. The thing to watch is not whether Meta opens the cloud business. It is whether any other hyperscaler follows. Amazon, Google, and Microsoft are all spending at similar scale. If one of them builds the same kind of compute-resale operation, Meta's hedge stops looking like one company's insurance policy and starts looking like an industry read on the shelf life of infinite demand.