AI’s real bottleneck is electricity. Who pays for it?

How two twenty-year nuclear contracts made the grid the story

Within a single week this autumn, two of the world’s three largest cloud companies signed twenty-year contracts for nuclear electricity. Amazon agreed on September 30, 2026, to buy 690 megawatts from Constellation’s Calvert Cliffs plant in Maryland, including 190 megawatts of new generating capacity expected between 2030 and 2032. Six days later, Google signed a twenty-year agreement bringing 890 megawatts of new nuclear capacity onto the PJM grid through equipment upgrades at eleven reactors in Illinois, Pennsylvania, and New Jersey. Neither deal builds a new reactor. Both buy more output from reactors that already exist; some of them older than the companies themselves. Utility Dive (Sept 30, 2026) Google (Oct 6, 2026)

An opinion essay in Al Jazeera on October 30 put the blunt question behind the deals: the bottleneck for artificial intelligence is no longer chips or talent, it is electricity; and the fight is now about who pays for it. Al Jazeera (Oct 30, 2026)

What the ledgers actually say about AI’s power appetite

The International Energy Agency estimates global data centre electricity demand will more than double to around 945 terawatt-hours by 2030, up from about 415 terawatt-hours in 2024, roughly the annual consumption of Japan. Efficiency gains have not dented the total: each improvement in performance per watt has been spent on larger models, longer reasoning runs, and always-on agents, a dynamic the essay traces to William Jevons’s 1865 observation that cheaper power creates new uses rather than savings. Energy Connects (Apr 2025)

In the United States, the arithmetic has reached the household bill. PJM, the grid operator serving some 67 million people across 13 states and Washington, DC, runs capacity auctions whose costs flow to consumers. Its independent market monitor calculates that data centre demand contributed about $29.4 billion, 46.2 percent, of capacity charges across the four most recent auctions, including roughly $6.3 billion in the latest one alone. Three consecutive auctions cleared at the regulatory price cap. PJM disputes the monitor’s framing, noting that the total value of attributed charges does not equal the cost actually borne by load. Energy News Beat (Oct 13, 2026)

The supply side is moving too. Constellation says it will invest more than $4.3 billion upgrading the eleven reactors tied to the Google deal. Microsoft holds a twenty-year contract for the output of Three Mile Island’s Unit 1, which Constellation is bringing back as the 835-megawatt Crane Clean Energy Center, its federal environmental review cleared in September 2026, with a 2027 target that at least one outlet puts at 2028. The US Department of Energy is lending up to $4.4 billion, the Associated Press reported on October 16, toward restarting three shuttered plants totalling about 2.25 gigawatts, including Palisades in Michigan, closed in 2022 and contracted to generate again by March 2027. Business Wire (Oct 6, 2026) EE News Europe (Oct 2026)

Where the estimates break down

Almost every figure in this story is a forecast or an attribution, not a settled bill. The $29.4 billion is the market monitor’s sensitivity analysis. The Natural Resources Defense Council’s estimate that a typical PJM household could pay $70 more per month by 2028 assumes current trends continue; a Synapse Energy Economics counter-estimate puts the cumulative figure closer to $100 billion. Neither is an audited household bill. NRDC (2026)

The contracts carry their own unknowns. Roughly 190 megawatts of Amazon’s Calvert Cliffs deal does not exist yet, it is conditional on Constellation spending the money to create it. Small modular reactors, which several tech companies are backing, are promised for the early 2030s, but very few have been built anywhere in the world. And the chips being installed today could be overtaken by newer technology long before the twenty-year electricity contracts powering them expire, a mismatch the industry’s announcements have not addressed. Enerdatics (Oct 2026)

Who pays: the question regulators could not answer in October

The cost question has become political. Polling in the United States shows growing public dissatisfaction with data centres, with electricity bills among the main concerns. Virginia’s State Corporation Commission approved a separate electricity rate class for data centres in November 2025; customers of 25 megawatts and above face fourteen-year contracts and minimum charges from January 1, 2027. Ohio’s regulator approved an AEP tariff in July 2025 with minimum billing up to 85 percent of contracted capacity over eight years. Illinois is still negotiating its POWER Act in the veto session. American Action Forum (Jan 8, 2026)

At the federal level, the answer was a rejection. On October 9, 2026, the Federal Energy Regulatory Commission threw out most of PJM’s proposal to impose special terms on large new loads, leaving only a 50-megawatt “large load” definition and a registry in place. The industry’s big voluntary commitment, Google’s agreement to bring new generation rather than compete for existing supply, now operates where regulators declined to mandate it. Vector News (Oct 19, 2026)

Why the industry stopped chasing efficiency and started buying supply

A decade ago the industry’s answer to AI’s energy problem was efficiency: better chips, smaller models, smarter cooling. The gains arrived, modern AI chips do far more computation per watt, and total demand kept rising anyway, because every gain was spent on something larger. Even DeepSeek’s demonstration that a competitive model could be trained at a fraction of the expected computing cost did not slow the spending. The essay’s Jevons parallel is the honest one: unable to shrink demand, the industry changed strategy and started buying supply. Al Jazeera (Oct 30, 2026)

The catch is the calendar. A new large reactor takes a decade or more from decision to electricity; the small modular reactors tech firms are funding are promised for the early 2030s, and almost none exist yet. So the “fast option”, restarting shuttered plants and uprating old ones, still means 2027, 2028, 2029, or beyond, while AI capital spending runs on a quarterly clock. Twenty-year power contracts are being signed for chips that may be obsolete in five. Nobody in the announcements has priced that mismatch. Enerdatics (Oct 2026)

What happens when the GPUs are obsolete but the contracts aren’t

Second-order effects cut both ways. If AI demand stalls, the utilities, and, through capacity charges, the ratepayers, are left holding twenty-year commitments built for a boom. If demand keeps doubling, the Gulf’s gamble pays: with electricity now the deciding factor in where computing gets built, energy-rich states can set the industrial tariff, the funding terms, and the contract rules before anyone else does. Qatar’s opportunity cost is real, an estimated 12 percent of national generating capacity, but so is the leverage. Qatar News Wire (Oct 25, 2026)

Meanwhile the constraint is often local, not national. India’s grid authority worries less about gigawatts than about transmission to the right substations, and Malaysia restricted non-AI data centre approvals from February 2026 to keep the load it can actually serve. The Hindu (Oct 10, 2026) Media Selangor (Oct 2026)

What to watch before the next auction

PJM’s December capacity auction for 2029/30 is expected to clear at the price cap for a fourth consecutive time. Illinois lawmakers meet in veto session on November 17–19 and December 1–3 to decide the POWER Act. Virginia’s data-centre rate class takes effect on January 1, 2027. Palisades is contracted to be generating by March 2027, and Constellation’s first Google-linked uprate is due by 2028. In the Gulf, the test is whether Qatar’s $20 billion joint venture publishes real megawatt targets; and whether Abu Dhabi and Riyadh attach the kind of bring-your-own-capacity terms that American regulators just declined to impose.