US Power Supply-Demand Deficit Forecast

Our demand forecast is grounded in the AI 2027 bottoms-up global compute model, which projects AI infrastructure power requirements by modeling GPU deployment, training cluster scaling, and inference demand across major cloud and enterprise platforms. The approach builds from chip-level power consumption through facility-level PUE to grid-level demand.

The model projects global AI power demand growing from ~9 GW in 2024 to ~668 GW by 2035—a compound annual growth rate of ~47%. Growth is supply-constrained through 2030 as new fabrication capacity comes online, after which the trajectory steepens as chip supply catches up to demand and new FABs reach full production.

Forecast certainty decreases substantially beyond 2030. Technology shifts (efficiency gains, new architectures), policy changes, and macroeconomic conditions could meaningfully alter the trajectory in either direction. Our possibility range reflects this asymmetry: the upper bound continues to track the AI 2027 growth trajectory (70% US share), while the lower bound widens to ~200 GW by 2035—reflecting a scenario in which demand growth plateaus due to efficiency breakthroughs, regulatory friction, or macroeconomic headwinds.

Supply capacity is modeled using our bottoms-up Supply Pipeline Model, which tracks nine generation technology categories through a 10-stage deployment pipeline. Each stage—from raw materials through interconnection—represents a potential bottleneck, and the binding constraint for each technology determines its deliverable capacity in any given year.

Annual nameplate capacity additions are adjusted by technology-specific Effective Load Carrying Capability (ELCC) factors to convert to firm deliverable capacity. For example, solar nameplate is adjusted by an ELCC of 18% (reflecting intermittency), while natural gas carries an 87% ELCC and nuclear 92%.

The nine technology categories modeled include natural gas (CC/CT), utility-scale solar, battery energy storage (BESS), onshore and offshore wind, nuclear (restarts and SMR), fuel cells, geothermal, and other resources. Each has distinct supply chain constraints, lead times, and scaling trajectories detailed in our full pipeline model.

The chart below overlays US AI power demand (derived from the global AI 2027 forecast at 50–70% US share) against cumulative firm supply capacity. The shaded bands represent the possibility ranges for each trajectory.

The structural deficit emerges clearly by 2027 and widens through the forecast period. By 2030, the base-case gap reaches 41.7 GW, and by 2035 it grows to 232.7 GW. Even under the most favorable supply assumptions (+15%) and the lowest demand scenario, a significant deficit persists through the forecast horizon.

The key insight is not the precise magnitude of the gap—which is subject to substantial uncertainty—but the structural nature of the mismatch: power generation and transmission infrastructure requires 5–10 year development cycles, while AI compute demand scales on 1–2 year doubling times. This temporal mismatch creates a durable investment window that cannot be quickly closed.

The AI infrastructure buildout has progressed through a series of binding constraints, each resolved only to reveal the next bottleneck. Understanding this sequence is critical to anticipating where capital will earn the highest returns.

As chip packaging capacity (CoWoS) expanded through 2024–2025 and HBM memory production ramped in 2025–2026, power emerged as the next—and most durable—binding constraint. Unlike semiconductor supply chains that can scale with factory investment on 2–3 year timelines, power generation and transmission infrastructure requires 5–10 year development cycles, creating a structural window that cannot be quickly closed.

Our demand forecast is informed by the AI 2027 bottoms-up compute model, which produces estimates above most published industry consensus. The model accounts for next-generation training cluster scaling and inference demand growth that many top-down forecasts do not yet fully incorporate.

The spread across published forecasts reflects deep uncertainty about AI scaling dynamics. Our model sits toward the higher end of the range, reflecting the AI 2027 view that compute demand growth accelerates through 2028–2030 as supply constraints ease and steepens further as new FABs reach full production. If efficiency gains prove larger than expected or AI adoption plateaus, actual demand could track closer to the lower industry estimates.

While the supply-demand chart above presents the conservative scenario (firm generation capacity only), two categories of interventions could materially increase effective grid capacity without building new generation: Grid-Enhancing Technologies (GETs) and flexible load management.

Grid-Enhancing Technologies

GETs—including dynamic line rating, advanced power flow controllers, high-performance conductors, and topology optimization—unlock additional transfer capacity on existing transmission infrastructure. By monitoring real-time conditions and dynamically adjusting power flows, these technologies improve throughput across the grid in three ways:

• Improved power flow: Dynamic line rating and topology optimization allow operators to route power along the most efficient paths, reducing congestion and unlocking latent capacity on existing lines.
• Higher capacity factors for dispatchable assets: By relieving transmission constraints, GETs enable gas, nuclear, and storage assets to dispatch more frequently at rated output rather than being curtailed due to local grid limitations.
• Reduced curtailment of intermittent generation: Stranded wind and solar assets—those unable to deliver power due to transmission bottlenecks—can reach load centers when grid capacity is optimized, improving effective renewable utilization.

Under moderate deployment assumptions, GETs could contribute an estimated ~57.9 GW of effective capacity by 2030 and ~77.5 GW by 2035. The DOE Liftoff Report on Innovative Grid Deployment (2024) estimates that advanced grid solutions could support 20–100 GW of peak demand growth nationwide. FERC Order 881 mandates ambient-adjusted transmission ratings by July 2025, creating a regulatory tailwind for broader GETs adoption.

Flexible Load Management

Data center workloads—particularly AI training—can tolerate limited curtailment during grid stress events. Under conservative assumptions (0.25% annual curtailment tolerance, ~85 hours per year), flexible load could accommodate an additional 76 GW of demand on the existing grid by 2030, according to a February 2025 study from Duke University.

Beyond conventional generation and grid modernization, several frontier technologies are attracting significant capital and research attention. Each could reshape the supply-demand landscape if commercialized at scale.

Small Modular Reactors (SMR)

Companies including NuScale, X-energy, and Kairos Power are pursuing commercial deployment of factory-built reactors in the early 2030s. The DOE’s Advanced Reactor Demonstration Program is funding first-of-a-kind builds. SMRs offer firm, dispatchable, carbon-free power at 50–300 MW per unit, making them well-suited for co-location with data center campuses. However, NRC licensing timelines and first-of-a-kind construction risk remain significant hurdles.

Fusion Energy

The DOE’s Fusion Roadmap (2025) targets commercial fusion on the grid by the mid-2030s. Commonwealth Fusion Systems is building the SPARC demonstration facility, targeting energy breakeven by 2027, with a commercial 200 MW ARC plant planned for Virginia under a Google PPA. Helion Energy is racing toward 2028 electricity delivery to Microsoft and has signed a 500 MW agreement with Nucor. Over $9.7 billion has been raised by fusion startups globally. Despite this momentum, no commercial fusion plant has yet delivered grid power, and regulatory frameworks for fusion licensing remain in development.

Orbital Compute

An emerging concept proposes relocating batch-eligible AI workloads (training, data generation) to space-based platforms with access to continuous solar power and passive thermal management in the vacuum of space. The Astro Compute financial model suggests orbital compute could reach cost parity with ground-based facilities by ~2031 under baseline assumptions, contingent on Starship-class launch economics reaching $200–$10/kg. However, radiation shielding overhead, bandwidth constraints, and unproven deployment at scale make this a speculative scenario for the current forecast period.

Superconducting Transmission

Recent breakthroughs in high-temperature superconductors—including SLAC’s room-pressure stabilization and Penn State’s predictive framework—are accelerating materials research. Microsoft invested in VEIR ($75M Series B) for HTS power delivery systems. Practical grid-scale superconducting transmission could eliminate the ~5% of electricity currently lost during transmission in the US. However, scalable manufacturing of superconducting materials at grid-relevant lengths remains years away from commercial viability.

The near-term window from 2026 through 2030 presents a structural supply deficit under virtually all scenario combinations. Even accounting for GETs and flexible load contributions, the gap between projected demand and firm deliverable capacity remains substantial. Infrastructure bottlenecks—particularly transformer lead times of 3–4 years, gas turbine delivery backlogs, and interconnection queue processing delays—create durable pricing power for developers with secured positions in the supply chain.

Regional divergence is shaping capital allocation decisions. The Central Belt (SPP and ERCOT) offers the fastest interconnection timelines and the most favorable permitting environments for new-build generation. The Southeast (SERC) provides access to vertically integrated utility partnerships with contracted revenue, while PJM remains the highest-value market per MW but faces severe queue congestion. Our state tier analysis provides granular scoring across 50 states on these dimensions.

All long-range forecasts carry substantial uncertainty. The demand trajectory could shift meaningfully based on AI efficiency improvements, policy changes, or macroeconomic conditions. The supply trajectory is similarly uncertain, with technology breakthroughs in SMR nuclear, enhanced geothermal, or long-duration storage potentially accelerating capacity additions beyond our conservative baseline. What appears structurally durable is the multi-year lead time mismatch between power infrastructure development and compute demand scaling.