The Power Constraint

August 6, 2026

For the decade before generative AI reached the market, efficiency gains absorbed most of the growth in digital demand. Global data centers consumed roughly 415 TWh in 2024, about 1.5% of world electricity.

Commercialization broke that pattern, in part because those levers had already been pulled. Global data-center demand grew 17% in 2025 to roughly 485 TWh, with AI-specific facilities up about 50% in a single year. The IEA’s April 2026 update has consumption roughly doubling to about 950 TWh by 2030, around 3% of global electricity, close to Japan’s entire consumption today. EPRI’s US estimate is wider still: 9–17% of national electricity by 2030 against 4–5% now.

Global Data Center Electricity Consumption (TWh)

The aggregate share understates the problem, because the load is extraordinarily concentrated. Data centers already take over 25% of Virginia’s electricity; EPRI projects 41–59% by 2030, with seven further states plausibly above 20%. In Ireland, 23% of metered consumption. Sites are scaling too: from 40–60 MW for a large facility to campuses approaching a gigawatt.

The interconnection queue tells a different story from headline volumes. At the end of 2025, roughly 8,200 projects representing 1,312 GW of generation and 749 GW of storage were awaiting connection, a total that fell 10% on the year as withdrawals outran new requests. What rose was gas: active gas capacity in the queues increased 86% to 253 GW. The binding constraint is time rather than volume. For projects reaching operation in 2025 the median wait exceeded five years, against a one-to-three-year construction cycle for data centers.

Utilities have responded with rate-base growth rather than redesign. Capex for the 46 utilities tracked by S&P Global RRA is forecast at a record $1.295 trillion for 2026–2030, including roughly $103 billion at Duke, the largest regulated capital plan on file and $81 billion at Southern.

The response from technology buyers has been bypass. Roughly 82 GW of behind-the-meter generation, overwhelmingly gas, has been announced since the start of 2025. Blocked permits in New Mexico and New Jersey have already slipped flagship projects, and Cleanview models only 5–13 GW online by end-2027. They are becoming energy developers anyway: the grid’s timetable does not match theirs.

This is the largest privately financed infrastructure program on record. The four largest US hyperscalers spent roughly $410 billion in 2025 and have guided to more than $725 billion in 2026: Amazon about $220 billion, Alphabet $195–205 billion, Microsoft approaching $190 billion, Meta $130–145 billion. Goldman Sachs models $5.3 trillion for those four through FY2030, up from $4.5 trillion three months earlier.

2026 Guided Capex Distribution ($ Billions)

Combined capex now absorbs close to all operating cash flow, against a ten-year average nearer 40%, and the strain is visible in the accounts. Alphabet reported negative free cash flow of $5.9 billion in the second quarter, its first since the 2004 listing and paused buybacks. Meta’s quarterly free cash flow fell 91% to $784 million; Amazon’s trailing twelve-month figure turned negative. Microsoft stayed positive at $19.6 billion, down 23%. The July reporting season turned from rewarding announced spending to interrogating its returns: Alphabet fell about 7% on its capex raise, Meta about 10%.

How it is being financed

The gap is being closed in credit. Hyperscaler bond issuance topped $100 billion in 2025, roughly four times the five-year average, and AI-related debt accounted for about 30% of net investment-grade dollar issuance. The five largest issuers reached roughly $194 billion in the first half of 2026 alone, with Goldman projecting about $250 billion for the full year, a third of capex. Appetite is thinning: order books that covered jumbo deals five times over in February covered them roughly twice by July.

Off-balance-sheet structures are the template: Meta’s Hyperion vehicle raised $27 billion of A+ rated notes and $2.5 billion of equity with Blue Owl and PIMCO, leaving Meta operational control and no consolidated project debt at a 6.58% coupon to 2049. Data-center ABS and CMBS issuance was roughly $27 billion in 2025. Morgan Stanley put the requirement plainly in mid-2025: some $2.9 trillion of global data-center investment through 2028 against roughly $1.4 trillion of Big Tech operating cash flow, a $1.5 trillion shortfall it stresses is a funding gap, not a debt forecast.

Three things to follow for anyone underwriting this build.

Gas is the bridge, and turbine scarcity is the real capex clock. GE Vernova’s gas turbine backlog and slot reservations, firm orders plus paid options on future manufacturing capacity, moved from 83 GW at the end of 2025 to 116 GW by mid-2026, with 125 GW targeted by year end and reservations taken into 2031. Output is the ceiling: 20 GW a year from the third quarter of 2026, 24 GW targeted for 2028.

GE Vernova Gas Turbine Backlog (GW)

Existing nuclear delivers firm clean electrons soonest, which is where the largest commitments have gone. Microsoft’s twenty-year Constellation agreement underwrites the $1.6 billion restart of Three Mile Island Unit 1. Meta contracted more than 2.6 GW from Vistra’s Ohio and Pennsylvania plants, uprates included, within a January 2026 package of up to 6.6 GW by 2035. None of the advanced-reactor capacity in that package, an Oklo campus and up to eight TerraPower Natrium units, is contracted to deliver before 2030.

Everything else is slower, smaller or unproven. Next-generation geothermal is the clearest genuine addition: Fervo’s $1.89 billion May 2026 listing was the largest renewable energy IPO on record, and its fully contracted 500 MW Cape Station expects first power late in 2026. Build cost stands near $7,000 per kW. Lazard’s 2026 estimates put utility solar at $40–98 per MWh and solar-plus-storage at roughly $49–140, against $51–129 for a combined-cycle plant whose turbine slot may not exist before 2031.

The Frontier