AI's Bottleneck Is Now Electricity, Not Chips

When the U.S. Department of Energy announced it would buy and own up to ten nuclear reactors, its chief of staff, Carl Coe, explained the reasoning himself. His explanation was not really about reactors.

The role of having the government involved in private markets is sacrosanct—you just don’t do it. — Carl Coe

The reactors were the instrument. The news was the reason offered for using one: a government that treats entering private markets as an act you just don't commit now says the AI power shortage is a national emergency, to be met with orders and ownership rather than prices. Why command rather than price? Because price cannot move at the speed this shortage does. New generation takes roughly a decade to build, and AI's demand for power is arriving now. Goldman Sachs puts the mismatch in one line.

AI’s insatiable power demand is outpacing the grid’s decade-long development cycles, creating a critical bottleneck. — Goldman Sachs Private Wealth Management

Money can be raised in a year; a power plant cannot be built in one. So a market that would normally signal scarcity through a rising price has nothing to sell at any price, and the states under strain reach for the only instruments that bite immediately: emergency orders, project halts, outright bans, and the purchase of generation itself. That is why, from Washington to Bangkok, the response to this shortage is arriving as orders, halts, and purchases rather than prices. The instrument each state reaches for tracks how much spare electricity it already has — what grid operators call the reserve margin, the cushion of unused generating capacity left after demand is met. China is the one country in this story that directs rather than rations. Decades of state-ordered overbuilding left it carrying reserve margins of 80 to 100 percent, and it has formalized the logic as a policy of "compute follows power," moving data centers from the crowded coasts to western regions with spare land and renewables ^[1][2]. Power is not its shortage; power is what it has in excess to absorb. The United States runs on 15 percent margins or less, and its posture is the opposite ^[1]. The Department of Energy has ordered the PJM grid to force data centers onto backup generation when demand spikes, and it is buying the reactors beside a voluntary pledge asking the big cloud providers to supply their own power ^[3][4][5]. Emergency decree, state ownership, and a request: rationing with a market's manners. Southeast Asia is the sharpest case: it promised a boom and delivered 38 percent of it ^[6]. The International Energy Agency puts the shortfall at a $50 billion annual gap in grid investment, against demand growing more than 5 percent a year ^[7].

The Asean region is facing a widening set of electricity security challenges due to a tightening supply and demand balance, changes in the generation mix with more variable renewable energy (VRE) and constrained electricity networks. — International Energy Agency

Against that backdrop, Thailand has halted dozens of data-center projects, and Malaysia's Johor region has banned the lowest tiers of facilities outright over water and power ^[8][6]. Next door, India is drafting a national policy whose stated purpose is to shield ordinary consumers from the cost of AI's power ^[3]. Its own projections show why: data centers rising from half a percent of national electricity consumption to 2 percent within a decade, with AI alone adding more than 26 gigawatts of demand by the early 2030s ^[3]. Then comes the pointed detail: the states with the least headroom are buying the one instrument that arrives last. Vietnam is building two reactors with Russian help, Malaysia is targeting 2031, the Philippines 2032, Indonesia small reactors by 2034 — all of it explicitly to buy a place in the AI build-out ^[9].

a signal to the world of how civil nuclear cooperation is something that is available. — Marco Rubio

It is the decade-versus-now mismatch made concrete: a shortage dated today, answered by plants dated for the 2030s. The bill lands on people who never ordered the compute. Maryland's consumer agency puts the cost of AI-driven grid upgrades at up to $216 a year per household, and Harvard's Ari Peskoe points to the unsettled question of whose infrastructure is being built ^[10].

There's no agreement as to what infrastructure is for a data center versus for the public. — Ari Peskoe

The same sort shows up at the scale of whole economies. The IMF raised Taiwan's growth forecast to 10.7 percent on the hardware boom while noting that Bangladesh, Cambodia, and Laos face inflation and rising energy prices with no share of the gains ^[11].

Risks remain tilted to the downside. — Argentina and the International Monetary Fund

And even among the winners it stays concentrated: China and the U.S. are projected to hold about 69 percent of the data-center market in 2030 ^[12]. The companies building the compute are buying generation the same way: Microsoft restarting a nuclear unit, Amazon ordering 64 reactors from X-energy ^[13]. North America's grid monitor warns winter demand will run 20 gigawatts higher than last year, more than the grid can reliably deliver ^[14]. Bernstein counts roughly 1,000 terawatt-hours of new U.S. demand through 2030, more than half of it from AI data centers, after two decades of near-flat power growth ^[15]. The honest limit sits in Texas. Its regulator ordered the grid operator to scrap a forecast that peak demand would quadruple past 367,000 megawatts, and the operator's chief executive conceded the number overshot the load growth he actually expects ^[5].

As a result of a changing landscape, we believe this forecast to be higher than expected future load growth. — Pablo Vegas

That correction arrived in April. The Johor bans went out after it, the Thai halts after it, the PJM order after it — each on a projection of the same shape, none of them tested the way Texas's was. The commands are moving faster than the numbers underneath them.