Hyperscalers will spend $700B on AI data centers in 2026, but power is the real bottleneck

Money is not the constraint

Hyperscalers are on track to spend $700-730 billion in capital expenditure in 2026, up 70-80% from the year before, and according to a dev.to analysis the sector's binding problem is no longer financing but electricity. The essay, originally published on the AI & Markets newsletter, argues that a data center can be financed in a quarter, while grid connections and reactors operate on multi-year schedules that no amount of capital can compress.

The demand numbers

The projected load is large by any standard. A Lawrence Berkeley National Laboratory update from August 2026 puts US data-center electricity use at 521-843 TWh by 2030, with a reference case of 649 TWh, or 11.8% of all US electricity. Goldman Sachs expects US data-center power demand to grow from 31 GW in 2025 to 66 GW in 2027, going from 4.1% to 8.5% of peak summer demand. Moody's estimates roughly $110 billion is needed for 45 GW of new generation by 2030, and the IEA sees global data-center consumption rising from about 415 TWh in 2024 to roughly 945 TWh by 2030. The dev.to author treats these reference cases as ceilings rather than central estimates.

Nuclear deals: contracted power versus options

The past year's nuclear deal flow splits into firm commitments and development bets. In the first bucket sit Microsoft's 20-year agreement with Constellation to restart Three Mile Island Unit 1 (835 MW, targeted for 2027), Amazon's $650 million campus co-located at the Susquehanna plant (960 MW initially, rising to 1.92 GW through 2042), and Oracle's deal for 125-250 MW from Wisconsin's Point Beach plant. The second bucket holds Google's arrangement with Kairos Power for up to 500 MW with a first 50 MW unit around 2030, Amazon's $700 million stake in X-energy covering up to 12 small modular reactors, and Meta's three agreements totaling up to 6.6 GW across TerraPower, Oklo, Vistra and Constellation, with first power no earlier than 2030. That is more than 10 GW of new nuclear contracted in a year, but the piece's verdict is blunt: the restarts and co-locations are power, while the SMR deals are options.

The SMR race, judged on paper

Among the three public SMR contenders, GE Vernova is the only one building: its BWRX-300 has been under construction at Darlington in Ontario since May 2025, with Unit 1 targeted for 2029, and the NRC issued the first US construction permit for the design to TVA's Clinch River site in September 2026. NuScale holds the only full NRC design certification, granted in January 2023, but its 6-7 GW backlog is concentrated in one utility customer and largely nonbinding, and its Romanian project has slipped to 2033. Oklo headlines the largest backlog, 18.1 GW at the end of 2025 and later reported around 14 GW, yet roughly 12 GW rests on a single largely nonbinding customer. It does claim the only Big Tech check written to an SMR developer, Meta's January deal for up to 1.2 GW in Ohio, and targets first power in 2028 via the DOE pilot track, though its NRC license is still pending and its first-ever revenue, $1.2 million in Q2 2026, came from services rather than electricity.

The rest of the ledger

Skepticism has precedent: NuScale's Utah project with UAMPS was cancelled in November 2023 after costs rose from $4.2 billion to $9.3 billion and its target price from $58 to $89/MWh, despite $1.36 billion in DOE cost-sharing. The bridge fuel is squeezed too: gas turbine costs are up 195% since 2019, industry output of 60-70 GW a year lags more than 110 GW of orders, GE Vernova's gas backlog sits at 116 GW with delivery slots offered into 2031, and operators pay $25 million just to reserve a 2030 slot. Fuel is another chokepoint, with Centrus, the sole US producer of high-assay uranium, having made under 2 metric tons since 2019 and volumes not expected before 2029. Meanwhile, US residential electricity rates are up 42% since 2019, PJM's December capacity auction cleared at its price cap with prices up ninefold since 2023, and Oracle's Wisconsin deal sits at the center of a proposed $176 million rate-hike dispute. The gap to 2030 will be filled mostly by natural gas, more than 30 of Moody's 45 GW, plus grid uprates and behind-the-meter projects; fossil fuels still supply over 60% of AI power today.

Why it matters

The only new nuclear electrons contracted to arrive this decade come from restarting reactors built in the 1970s; everything else lands in 2029-2033 at best, and nuclear schedules have a history of slipping. A $700-billion-a-year construction program is therefore running years ahead of the generation that must feed it, turning AI infrastructure risk from financial into physical. The piece also flags a demand-side wildcard: four rival lab CEOs recently agreed for the first time that frontier scaling should slow, and infrastructure stocks sold off the same day. If that sentiment hardens, today's 2030 demand forecasts become the bull case.

What to watch

The dev.to author closes with a watchlist rather than a prediction: whether Three Mile Island restarts on schedule in 2027, whether Oklo's license survives a regulator that rejected its 2020 application, whether Darlington's 2029 target holds, whether gas turbine slots stretch beyond 2031, and whether the lab slowdown talk becomes more than a press cycle.