The Outlook for Data Center Power Demand as AI Token Use Grows

Episode Summary

Executive Summary: The episode argues that AI-driven power demand is surging faster than expected, with U.S. and global data-center load forecasts revised sharply higher. However, the limiting factors have shifted from megawatts to equipment, labor, regulation, community pushback, water, and interconnection delays, making supply growth uneven, expensive, and politically fraught.

Main Topics: AI-driven power demand revisions (Priority: 5/5): Carly Davenport and Brian Singer explain that both U.S. and global data-center power demand forecasts were raised materially, driven by more projects, higher utilization, and stronger hyperscaler spending. Supply-chain constraints beyond generation (Priority: 5/5): The bottlenecks have moved upstream to turbines, transformers, transmission, tradespeople, and cooling/physical site conditions, meaning the challenge is not just building more power but building the right infrastructure fast enough. Regional shifts in U.S. data-center growth (Priority: 4/5): The discussion maps where growth is concentrated, with PJM remaining dominant, MISO emerging as a major second hub, and ERCOT still important but facing more uncertainty and physical constraints. Grid interconnection and behind-the-meter power (Priority: 5/5): Long interconnection queues are pushing more projects toward behind-the-meter gas generation as a bridge solution, though both guests see grid connection as the preferred long-term model. Regulation, affordability, and community pushback (Priority: 5/5): The speakers emphasize that utility tariffs, regulatory approval, moratoriums, and local opposition may become the steepest constraint on data-center expansion, especially as communities weigh jobs and tax revenues against reliability, price, water, and noise concerns. Capital availability and utility balance sheets (Priority: 3/5): Despite rising capital needs, regulated utilities and independent power producers appear financially able to support the buildout, with equity funding and improving credit metrics helping finance the next phase.

Key Arguments: AI demand is outpacing even major efficiency gains: token and server improvements are being overwhelmed by faster growth in workload demand. Hyperscaler spending forecasts have risen sharply, reinforcing the case for higher power demand and more data-center capacity. The U.S. does not face a nationwide power shortage, but specific regions—especially PJM—are tightening quickly. MISO is becoming more attractive because regulated utilities can coordinate generation, transmission, and customer relationships more efficiently than deregulated markets. Behind-the-meter gas generation is increasingly necessary because grid interconnection queues can take years and delay projects. The industry is likely to use a mixed-supply strategy: natural gas in the near term, renewables plus storage in the near term, combined-cycle gas in the medium term, and nuclear longer term. Community resistance is real, but some localities are actively courting data centers for tax base and construction jobs, suggesting geographic concentration rather than a complete slowdown. Regulatory tools like special tariffs and quantified bill-savings commitments may help utilities accommodate growth without harming reliability or affordability. Water and cooling constraints are increasingly central; in the U.S., minimizing water use may take priority even if it raises power consumption. The most important risk indicator is whether state-level moratoriums spread and materially delay projects or reshape where data centers can be built.

Data Points: U.S. power demand forecast CAGR through 2030: 3.5% - Carly said Goldman Sachs raised its U.S. electricity demand outlook from 3.2% to 3.5% CAGR through 2030. Prior U.S. power demand forecast CAGR through 2030: 3.2% - Previous forecast before the recent revision. Year-to-date U.S. power demand growth: over 4% - Carly noted current growth is already running above the long-run forecast pace. 2030 U.S. data-center power demand: 108 GW - Raised from about 83 GW due to new projects and higher utilization. Prior 2030 U.S. data-center power demand estimate: 83 GW - Earlier estimate before the upward revision. U.S. data-center vacancy rate today: 1-2% - Existing data-center markets are close to full, indicating tight supply. Projected U.S. data-center vacancy rate in 2030: 3% - Expected average vacancy as supply expands. Hyperscaler spending for 2027: $1.7 trillion - Brian said expectations rose from $1.2 trillion in March to $1.7 trillion now. Hyperscaler spending for 2027 prior estimate: $1.2 trillion - Earlier projection referenced by Brian. Hyperscaler spending for 2029: $2.1 trillion - Up from a prior estimate of $1.5 trillion. Hyperscaler spending for 2029 prior estimate: $1.5 trillion - Earlier projection referenced by Brian. Global data-center power demand in 2030 vs. 2025: up about 170% - Brian said the global outlook was revised up materially from the March estimate of 117%. Earlier global data-center power demand outlook for 2030 vs. 2025: 117% - Prior estimate at the time of the earlier discussion. AI power demand equivalent: the whole country of Japan - Brian said AI is adding the power equivalent of Japan over the seven-year period through 2030. PJM share of U.S. data-center demand today: 35% - PJM remains the largest data-center market. ERCOT share of U.S. data-center demand today: 15% - Texas is currently the second-largest market. Southeast share of U.S. data-center demand today: 13% - Tied for third place currently. Pacific Northwest share of U.S. data-center demand today: 13% - Tied for third place currently. MISO projected share of U.S. data-center demand by 2030: 16% - Expected to overtake ERCOT, the Southeast, and the Northwest to become the second-largest market. ERCOT projected share of U.S. data-center demand by 2030: 14% - Still a major market, though with more uncertainty. Regulated utility capital investment growth: up 60% over the next five years - Carly said regulated utilities are increasing capital spending materially. Equity funding share for incremental regulated utility capital: 30-50% - Expected to fund growth while preserving balance sheets. Turbine manufacturer order book: halfway sold out for 2031 by end of this year - Brian cited this as evidence of continued turbine constraint. Behind-the-meter gas capacity by 2030: 30 GW - Carly said this could serve about 20 GW of power delivery. Behind-the-meter share of 2030 data-center demand: just over 20% - Represents a meaningful off-grid supplement to grid power. Share of data-center demand expected to be met with natural gas: 60% - Carly said natural gas will likely dominate near-term capacity additions. Share of data-center demand expected to be met with renewables: 40% - Renewables plus storage are expected to play a large supporting role. Interconnection queue delays: 2, 3, 5, or 7 years - Carly highlighted long waits depending on the regional grid. Number of local/regional moratoria: more than 300 - George cited widespread local pushback against data-center development.

Pivotal Quotes: "It's adding the power equivalent of the whole country of Japan, which is the number five power-consuming country." — Brian Singer: A scale comparison used to emphasize how large the AI-driven load increase has become. "The question has now been answered in basically the least comforting way possible, partly, unevenly, and expensively." — Allison Nathan: Her framing of whether the power grid can handle AI growth. "The single most important issue facing power and utilities companies today." — Carly Davenport: How she describes regulatory uncertainty around data-center interconnections.

Implications: AI power demand is likely to keep rising, but execution will be constrained by grids, permits, public acceptance, water, and labor. Expect more behind-the-meter generation, regional winners and losers, and a growing policy battle over where AI infrastructure should be allowed to grow.