The AI tragedy of the commons: hyperscalers racing for 38 GW each are draining one shared grid

Microsoft thinks it will need more than 38 gigawatts of data-centre capacity by 2032—more than triple the roughly 12 gigawatts it runs today and more than New York state draws at its busiest moment. Its rivals are committing in the same register. The four largest hyperscalers, Microsoft, Alphabet, Amazon and Meta, have pledged close to $2.4trn of spending over the coming years, and the sector's capital expenditure for 2026 is expected to approach $725bn, up more than three-quarters on the prior year. Amazon alone has flagged $200bn of investment this year.

Every one of those plans is a sensible act of self-defence: the firm that does not build forfeits the race to deliver artificial intelligence at scale. And every one of them pulls on the same underlying constraint, the electricity grid. The instinct that this is a tragedy of the commons—an uncoordinated stampede that exhausts what is supposed to be shared—deserves scrutiny, because the answer determines who ultimately absorbs the cost.

The grid as a disputed pasture

A commons is a resource that is rivalrous (one user's draw denies another) but hard to exclude. On that test, transmission capacity fits. Each hyperscaler, negotiating separately, asks the same grid operators to deliver power to a data centre. The queue of requests has become the bottleneck that semiconductors used to be. Interconnection timelines in large markets now stretch beyond five years, several times longer than the 18–24 months to build the building itself.

Scarcity is showing up in the market's own language. In PJM, the grid that serves the Mid-Atlantic, capacity prices—the fee paid to guarantee power is available for a future peak—cleared at the regulatory cap of $329 per megawatt-day for 2026–27, roughly ten times the level of two years earlier and a price that would have seemed fanciful in 2024. Because the grid operator runs an auction, it sets a single price for everyone, including households that never signed up for the AI boom. The Natural Resources Defense Council calculates that PJM customers could face bill increases topping $163bn through 2033, or about $70 a month for an average family by 2028. Seven billion dollars, give or take, was added to the region's capacity bill in a single auction, and more than $16bn of generation payments are passed straight to utility customers.

Why it is not quite a pasture

Yet the analogy fails at precisely the point where it is most convenient. A common pasture has no price and no owner, which is why it is overgrazed. The grid prices scarcity, and it can exclude. Capacity auctions clear at their cap precisely because the mechanism is working, telling everyone that power at the peak is expensive. If that price reached the buyers who create the demand, the problem would solve itself in an ordinary market.

Two escape hatches blunt the tragedy. First, data-centre developers are increasingly going behind the meter, building their own gas turbines that never touch the shared wire at all. That is exactly how an actor exits a commons: self-supply instead of a fight over the pool. Second, the cost claims are contested. The Data Center Coalition, an industry body, published a study arguing there is no evidence data centres drive up residential bills under existing rate structures. Householders' rising bills have many causes—an aging grid, storm damage, fuel prices—of which the AI boom is only one.

The real failure is overlapping expectations

The genuine collective-action problem is subtler. It is not that a finite stock of electricity runs out—energy can always be priced or self-generated. It is that each hyperscaler formalises the same projected national demand as if it were racing a rival, and markets respond to the sum of those overlapping claims.

Consider the consequences. Grid operators contract for capacity to meet a worst-case peak that assumes every firm's plan materialises on time. Jurisdictions build transmission for a future that may be double-counted; PJM staff are so worried about duplicated requests that they are considering asking developers whether their proposals mirror others in different regions, an exercise one executive admitted reduces to "thou shalt ask". When the projections prove overlapping or the demand arrives late, the cost of idle capacity does not simply vanish. Under current rules it is socialised to the ratepayers who cannot vote with their feet.

That is why the fight has moved from engineering to politics and pricing. Virginia, the epicentre of the buildout, approved a tariff in 2025 that requires the largest customers to foot at least 60% of generation and 85% of transmission costs, charging them for dedicated substations and high-voltage lines even when they under-use them. Texas, which still socialises costs across all ratepayers, has legislated to force new loads to pay a larger share, though the formula is unresolved. These are not technical disputes. They are the market discovering, tariff by tariff, where the externality of the buildout should land.

Who bears the strain

Divide the sector by exposure and the picture is clean. Regulated utilities are the protected winners: their rate base grows with every gigawatt of demand and every mile of wire, and their returns are guaranteed by law rather than by AI revenue. Independent power producers that sign dedicated contracts with hyperscalers also prosper, because they are selling scarcity itself. The hyperscalers carry the real margin risk: at 47% of sales, Microsoft's capex—already $145bn last year—must eventually translate into revenue as depreciation hits, and the grid is a constraint that delays that revenue without delaying the cost. For them, the danger of the race is not that the grid runs out but that they pay for capacity whose connections slip, or whose demand never shows.

The residual losers are the households. They cannot build their own turbines, cannot delay, and cannot exit. They are the ones billed for a guaranteed future of peak power, whether or not a single promised gigawatt of it is ever switched on. That is the abiding irony of the AI commons: the resource is not, in the end, exhausted. It is simply that the accounting for who caused the scarcity remains unfinished, and until the last regulator finishes it, the public both pays the marginal bill and provides the buffer that lets the race continue.