Terrestrial Energy Supply Constraints

Answer

If AI compute demand grows as projected (3-4x by 2030), terrestrial data center electricity costs will likely rise modestly in the central case but could increase substantially in supply-constrained scenarios. The outcome depends on the race between demand growth and the supply response across multiple generation technologies.

Scenario ranges for effective data center electricity cost ($/kWh, blended grid + BTM, pre-PUE):

The central case shows a temporary "supply squeeze" from 2028-2032 where demand outpaces both grid interconnection capacity and BTM build-out, pushing costs ~5-10% above current levels. Costs moderate afterward as grid infrastructure catches up and solar+storage costs decline. The conservative case reflects a scenario where multiple supply constraints compound -- turbine shortages, regulatory pushback on BTM gas, and slow grid buildout -- creating a sustained cost premium of 30-50% above current levels that persists into the mid-2030s. This is the scenario most favorable to the orbital argument.

However, even in the conservative scenario, terrestrial electricity costs remain in the $0.095-$0.110/kWh range. To test whether any energy price could make orbital competitive, we derive the break-even terrestrial energy cost from the model: holding all terrestrial TCO parameters at central except energy cost, and setting orbital TCO to the optimistic scenario, the energy cost that equalizes the two is:

break_even_energy = (orbital_tco_optimistic − terrestrial_non_energy_tco_central) / (8760 × PUE)

From 2028 onward, the break-even is negative — meaning even if terrestrial electricity were free, orbital would still be more expensive in the optimistic scenario. This is because the orbital cost premium is driven by the effective lifetime penalty, cost of capital spread, and GPU space adaptation, not by energy savings. Energy cost is only ~6-7% of terrestrial TCO; eliminating it entirely cannot close a ~1.4x gap that is structural.

The previously asserted $0.15-$0.25/kWh threshold was not derived from the model and overstates the plausibility of energy-cost-driven parity. The actual conclusion is stronger: no terrestrial energy price makes orbital competitive from 2028 onward, and even in 2026 the required price ($0.31/kWh) is implausible at the systemic level. Note: this finding is conditional on the model's time-invariant orbital parameters (WACC, effective lifetime, platform manufacturing cost). If these improve over time — as would be expected with operational experience — the break-even energy cost could become positive in the 2035-2040 window. See the WACC analysis for the estimated magnitude of WACC compression.

Analysis

1. Demand Projections: How Much Power Does AI Need?

Multiple institutional forecasts converge on rapid growth but diverge on magnitude:

Current baseline (2024-2025):

• Global data center power capacity: ~30 GW for AI specifically, ~55 GW total including cloud and traditional workloads [epoch-ai-power-30gw.1, gs-dc-power-demand-2025.1]
• Global electricity consumption: ~415 TWh (1.5% of global electricity) iea-energy-and-ai-2025.1
• U.S. data center demand: ~25 GW in 2024 mckinsey-dc-power-2030.1

2030 projections:

The forecasts agree that data center power demand will grow 2-4x by 2030 [gs-dc-power-demand-2025.1, mckinsey-dc-power-2030.1, iea-energy-and-ai-2025.1]. McKinsey's 219 GW global figure is the highest, driven by aggressive AI workload assumptions (156 GW for AI alone, a 3.5x increase from 2025's 44 GW AI load) mckinsey-dc-power-2030.1. Goldman Sachs is more moderate at ~122 GW. The IEA's 945 TWh implies roughly 100-110 GW at typical capacity factors.

2035 projections:

• BNEF: 106 GW for U.S. alone, up 36% from their April 2025 forecast bnef-dc-power-106gw.1
• IEA: ~1,300 TWh global consumption iea-energy-and-ai-2025.2

Key uncertainty: AI computing capacity is currently doubling every ~7 months epoch-ai-power-30gw.2, but this growth rate in chip deployment may not translate linearly to power demand growth because (a) chip efficiency improves with each generation, (b) models like DeepSeek demonstrate efficiency gains in inference workloads, and (c) actual utilization rates may be lower than capacity. The forecasts above implicitly assume some efficiency offset but still project rapid demand growth.

For scenario construction, we use:

• High demand case: 80+ GW U.S. by 2030, 120+ GW by 2035 (McKinsey trajectory)
• Central demand case: 60-70 GW U.S. by 2030, 90-100 GW by 2035 (Goldman/BNEF blend)
• Low demand case: 45-55 GW U.S. by 2030 (IEA-consistent)

2. Supply Responses and Their Costs

A. Grid Connections

Grid power remains the lowest-cost option for data centers where available. However, interconnection has become the primary bottleneck:

• Average interconnection timeline: rose from <2 years (2008) to >8 years (2025) camus-grid-interconnection.1
• PJM alone faces 31 GW of projected data center load over five years bnef-dc-power-106gw.2
• Utilities are experiencing "more growth in a single year than they used to see in ten or twelve" camus-grid-interconnection.2
• Over 700 GW of interconnection requests received by utilities in 2025 -- more than total U.S. consumption eesi-dc-energy-bills.1

Cost impact of grid constraints: PJM capacity market prices rose ~10x from $28.92/MW-day (2024/25) to $269.92/MW-day (2025/26) to $329.17/MW-day (2026/27) ieefa-pjm-capacity-prices.1. Data centers were responsible for 63% of the 2025/26 price increase, adding $9.3 billion in costs to PJM ratepayers ieefa-pjm-capacity-prices.2. The 2027/28 auction hit the $333.44/MW-day price cap ieefa-pjm-capacity-prices.3.

For data centers specifically, grid wholesale electricity costs have risen dramatically in concentrated markets. Virginia prices increased 267% over five years eesi-dc-energy-bills.2. The EIA forecasts a 45% wholesale price increase at the ERCOT North hub in 2026, with a high-demand scenario showing 79% higher prices ieefa-pjm-capacity-prices.4.

Flexible interconnection as an alternative to grid upgrades: A joint Camus/Princeton Zero Lab/Encord study modeled optimal power flow for 500 MW data centers at six sites within a single utility's territory. At four of the six sites, total annual curtailment required to avoid transmission upgrades was just 7, 11, 13, and 35 hours respectively — at most 0.4% of hours per year volts-dc-flexibility-2026.1. The longest individual curtailment events were 5-16 hours — well within battery ride-through capability volts-dc-flexibility-2026.2. Two sites had no constraints at all. This approach trades minor operational flexibility for eliminating years-long transmission upgrade delays. ERCOT is developing formal rules for flexible load interconnection, and PJM is actively debating similar frameworks volts-dc-flexibility-2026.3. The opportunity cost of delayed data center deployment is estimated at ~$7 billion per GW per year, making even expensive flexibility solutions (e.g., batteries instead of gas) worthwhile if they accelerate interconnection volts-dc-flexibility-2026.4.

Grid reform progress: PJM has processed 170,000 MW of generation requests since 2023, with new interconnection agreements targeted at 1-2 years going forward. FERC issued rules in December 2025 establishing data center colocation options at power plants. These reforms are real but will take years to clear the existing backlog.

B. Behind-the-Meter Gas Turbines

BTM gas generation has emerged as the primary short-term solution to grid constraints:

• At least 46 data centers with 56 GW combined BTM capacity identified grist-btm-gas-2026.1
• 1,000+ GW of gas-fired power in development globally, 31% increase year-over-year; U.S. accounts for ~250 GW grist-btm-gas-2026.1
• Texas alone: 58 GW of gas in planning/construction, nearly half exclusively for data centers grist-btm-gas-2026.2
• McKinsey estimates 25-33% of incremental DC demand through 2030 will be met by BTM solutions latitude-btm-traction.1
• Major projects: Stargate (7-10 GW), Joule (1.3 GW fully islanded), VoltaGrid/Oracle (2.3 GW), Project Horizon/CoreWeave (2 GW) latitude-btm-traction.2

Technology options and costs:

• Boom Superpower: 42 MW turbine at $1,033/kW_gen, 39% efficiency; $1.25B deal with Crusoe for 29 units (>1 GW), deliveries starting 2027 grist-btm-gas-2026.3
• FTAI Power: 25 MW aeroderivative adapted from the CFM56 engine, leveraging 22,000+ existing engines; capacity to deliver over 100 units annually [ftai-power-launch]
• Reciprocating engines: Meta's El Paso facility uses 813 modular mini-turbines; faster ramp-up (~1 min vs ~1 hour) but lower efficiency than CCGT grist-btm-gas-2026.6
• Simple-cycle gas capex: $800-$1,800/kW_gen [terrestrial-power-asset-capex evidence A1-A4]

Constraints on BTM gas:

1. Gas turbine manufacturing bottleneck: GE Vernova (20 GW/yr by mid-2026, stretch to 24 GW by mid-2028), Siemens Energy (record backlog of $148B), and Mitsubishi Power (deliveries not until 2028-2030) [rmi-gas-turbine-constraints.1, ge-vernova-backlog-2025.1]. Two-thirds of U.S. gas project developers have not yet identified their turbine manufacturer grist-btm-gas-2026.7. GE Vernova expects turbine reservations sold out through 2030 by end of 2026 ge-vernova-backlog-2025.2.
2. Regulatory risk: BTM operations circumvent state climate regulations. New Mexico's Project Jupiter ($165B, 2,880 MW) would "outweigh the actions that New Mexico has taken to lower emissions" in recent years grist-btm-gas-2026.8. Cornell estimates 44 million metric tons CO2 from DC gas buildout by 2030 grist-btm-gas-2026.9.
3. Environmental opposition: Environmental groups including Center for Biological Diversity are opposing BTM gas facilities over air quality and climate impacts grist-btm-gas-2026.10.
4. Cost escalation: New gas plant construction costs tripled since 2022, from $785/kW to $2,000-$3,000/kW for utility-scale CCGT eesi-dc-energy-bills.3. EPC costs up 20-30% since 2021 gasturbinehub-market-2025.1. BNEF records CCGT LCOE at all-time high of $102/MWh bnef-lcoe-2026.1.

Important nuance: Despite these constraints, BTM gas is being deployed at speed. The 56 GW already identified represents a massive commitment, and aeroderivative/reciprocating options partially bypass the large-frame turbine bottleneck. The constraint is not that BTM gas cannot scale, but that it scales at higher cost and with growing regulatory uncertainty.

C. BTM Solar + Battery

Solar+storage is the fastest-declining cost option and is increasingly competitive with gas for data centers:

Current economics:

• Solar+storage LCOE: $57/MWh average in 2025 (87 GW deployed globally) bnef-lcoe-2026.2
• Battery storage LCOS: $78/MWh (4-hour), down 27% year-over-year bnef-lcoe-2026.2
• Utility-scale solar LCOE: $38-78/MWh, average $58/MWh, down 4% year-over-year (Lazard LCOE+ June 2025) [lazard-lcoe-2025]
• Solar+4hr battery capex: ~$2,175/kW_gen (EIA 2023 benchmark), declining rapidly [terrestrial-power-asset-capex evidence B1]

Projections:

• By 2035: BNEF forecasts 30% solar LCOE reduction, 25% battery storage reduction bnef-lcoe-2026.5
• Solar+storage could reach $1,400-$1,600/kW_gen by 2030 [terrestrial-power-asset-capex analysis]

Constraints:

1. Intermittency: Solar+4-hour battery does not provide 24/7 baseload for data centers. Longer-duration storage (8-12 hours) or gas backup is required, adding cost.
2. Panel supply concentration: China holds 80%+ of global polysilicon, wafer, cell, and module manufacturing capacity through 2026. Module shortages reported in the U.S. through year-end 2026 due to tariffs and supply chain restrictions.
3. Land requirements: At 5-8 acres/MW for solar, a 1 GW solar installation requires 5,000-8,000 acres, constraining co-location with data centers.
4. Battery supply: While battery costs are plunging globally, U.S.-specific costs remain significantly higher — the NREL 2025 U.S. benchmark is $334/kWh for a complete 4-hour system [nrel-battery-cost-2025], versus ~$125/kWh globally (ex-US, ex-China) per Ember [ember-battery-cost-2025] — due to tariffs and domestic content requirements.

Despite these constraints, solar+storage capacity is not fundamentally supply-limited in the way gas turbines are. Global solar module manufacturing capacity reached 1.8 TW in 2025, far exceeding demand. The constraint is more about deployment speed, interconnection, and 24/7 reliability than about physical supply limits.

Portfolio / "power park" approach: Rather than building all generation on-site, data centers can assemble portfolios of solar, wind, battery, and VPP resources distributed across the broader grid region — once flexible interconnection unlocks grid access for 99%+ of hours. This approach leverages hundreds of GW of renewable capacity already in interconnection queues for 2028-2030 deployment volts-dc-flexibility-2026.5. Companies like Firma Power (founded by Jesse Jenkins) are developing optimization software to assemble these portfolios, matching capacity accreditation, deliverability constraints, and data center demand profiles. The portfolio model makes solar+storage more viable for data centers by combining it with grid power and demand flexibility rather than requiring it to provide standalone 24/7 baseload.

D. Nuclear (Existing Plants and SMR)

Existing nuclear colocation/restart:

• Constellation plans TMI Unit 1 (renamed Crane Clean Energy Center) restart (835 MW) by 2027 at ~$1.6B cost, with $1B DOE loan (closed November 2025). 20-year PPA with Microsoft for carbon-free electricity [constellation-tmi-restart].
• Meta signed deals with Oklo, Vistra, and TerraPower for up to 6.6 GW of nuclear power by 2035 — including PPAs for 2,176 MW from existing Vistra plants, up to 2.8 GW from TerraPower Natrium reactors, and up to 1.2 GW from Oklo Aurora fast reactors [meta-nuclear-deals-2026].
• Existing nuclear PPAs reportedly priced at nearly double the standard wholesale rate.

SMR timeline and costs:

• NuScale: first module planned for 2029, 12-module plant at 720 MW. Only SMR with NRC design approval fervo-geothermal-2025.1.
• Google/Kairos Power: 500 MW across 6-7 reactors, first unit by 2030, full deployment by 2035.
• FOAK costs: ~$14,600/kW (vs. $2,800/kW projections), with 7-10 year timelines introl-smr-timeline.1.
• NOAK projections: $5,000-$8,000/kW_gen by 2035-2040, with LCOE of $60-$100/MWh [terrestrial-power-asset-capex evidence C1-C3].

Realistic contribution: SMRs will not materially contribute to data center power supply before the mid-2030s. Even optimistic timelines show first commercial deployments in 2030-2032 with small capacity (hundreds of MW, not GW). Existing nuclear restarts and colocation PPAs can contribute 5-10 GW by 2030 but face their own permitting and grid constraints. GlobalData forecasts at least 3 GW of data-center-linked SMR capacity commissioned in the next three years, with nuclear deployment for data centers peaking between 2031 and 2035.

E. Geothermal

Fervo Energy (enhanced geothermal):

• Cape Station: 100 MW by 2026, 500 MW by 2028 fervo-geothermal-2025.2
• Per-well costs halved from $9M to $4.5M; 35% learning rate across first 8 wells fervo-geothermal-2025.3
• Current LCOE: ~$88/MWh with federal tax credits; projected $50-60/MWh by 2035 dc-geothermal-frontier.1
• U.S. geothermal potential: ~3,400 GW with current drilling technology dc-geothermal-frontier.2
• 80% of oil/gas workforce skills transferable dc-geothermal-frontier.3

Limitations: Geothermal is geographically constrained to the western U.S. for enhanced systems. While technically promising with massive potential, current deployed capacity is minimal (<1 GW) and scaling to tens of GW will take a decade or more. Geothermal is best understood as a post-2030 contributor.

3. Supply-Demand Dynamics: Constraint Thresholds

The key question is: at what demand level does each supply source become constrained, and what happens to marginal costs?

Supply capacity by source (U.S., annual addition capacity):

Total feasible annual supply addition: ~30-55 GW/yr by 2028-2030.

Demand for new supply: If U.S. data center demand grows from 25 GW (2024) to 60-80 GW (2030), that requires 35-55 GW of new capacity over 6 years, or ~6-9 GW/yr -- within the feasible supply range.

Compute flexibility and actual utilization: Demand projections typically assume nameplate capacity, but most data centers operate at ~40% average utilization. Training workloads peak at 100% but cycle on and off; inference and cloud workloads have more organic variability volts-dc-flexibility-2026.6. Google can shift load between data centers in seconds for reliability purposes, and some providers are commercializing compute flexibility as a grid service (e.g., Emerald AI). This means effective demand growth on the grid is lower than nameplate additions suggest.

However, data centers are not the only source of demand growth. NERC projects 150 GW of total peak demand growth over the next decade, alongside 120 GW of retirements, creating a 270 GW total generation gap rmi-gas-turbine-constraints.3. Data centers compete with electrification, industrial load, and EV charging for the same supply. This is where the supply squeeze emerges: total demand for new generation far exceeds data center demand alone.

Marginal cost dynamics:

When demand exceeds readily available supply, the marginal source of electricity becomes more expensive:

1. First constraint hit (already occurring): Grid interconnection. Data centers that cannot get grid connections turn to BTM gas at a premium. Grid wholesale prices in concentrated markets (PJM, Northern Virginia) have already spiked. PJM capacity costs rose 10x from 2024/25 to 2026/27 ieefa-pjm-capacity-prices.1.

2. Second constraint (2027-2028): Large-frame gas turbine supply. With GE Vernova's backlog extending to 2029 and turbine reservations sold out through 2030 ge-vernova-backlog-2025.2, operators shift to higher-cost aeroderivative and reciprocating alternatives. Turbine capex rises 20-30% above 2021 levels gasturbinehub-market-2025.1.

3. Third constraint (2028-2030): Cumulative competition for all power sources. As multiple demand drivers (DCs, EVs, electrification, manufacturing reshoring) compete for the same gas turbines, solar panels, battery cells, and grid interconnection capacity, prices for all components rise. The BNEF record CCGT LCOE of $102/MWh (16% increase in one year, driven by data center demand) illustrates this dynamic bnef-lcoe-2026.1.

4. Easing of constraints (2030-2035): Gas turbine manufacturing expands (GE Vernova targeting 24 GW/yr by mid-2028), grid interconnection reforms take effect (PJM targeting 1-2 year processing), flexible interconnection models unlock existing transmission capacity with minimal curtailment volts-dc-flexibility-2026.1, solar+storage costs continue declining (30% and 25% reductions by 2035), and first SMRs come online. However, new constraints could emerge if demand exceeds even these expanded projections.

4. Scenario Construction

Optimistic Scenario (for Terrestrial)

Assumptions: Supply keeps pace with demand. Grid interconnection reforms succeed and processing times drop to 2-3 years by 2028. BTM gas scales rapidly using diverse turbine sources (aero, recip, Boom Superpower). Solar+storage costs decline as projected. Battery costs continue plunging. Some demand moderation from AI efficiency improvements (DeepSeek effect). Gas prices remain moderate ($3-4/MMBtu).

Cost trajectory (blended $/kWh, pre-PUE):

• 2026: $0.065 (current low-cost markets)
• 2030: $0.055 (solar+storage at $40-50/MWh displaces more expensive sources)
• 2035: $0.050 (mature solar+storage, early SMR)
• 2040: $0.045 (diversified cheap generation mix)

Why this scenario might occur: Solar+storage economics are genuinely extraordinary -- $57/MWh combined in 2025, heading toward $40/MWh by 2030. If battery duration extends to 8-12 hours at declining cost, solar+storage can provide near-baseload power at costs below gas. Meanwhile, grid reforms and the sheer number of BTM options (Boom, FTAI, recip engines, fuel cells) prevent any single bottleneck from choking supply.

Central Scenario

Assumptions: Supply partially constrained from 2028-2032 as demand growth outpaces grid buildout. BTM gas fills the gap but at elevated costs. Gas turbine shortage causes a 2-3 year cost premium. Grid interconnection improves gradually but doesn't fully resolve until early 2030s. Solar+storage grows but doesn't dominate until mid-2030s. Gas prices at $3.50-5/MMBtu.

Cost trajectory (blended $/kWh, pre-PUE):

• 2026: $0.075 (current blended rate)
• 2028: $0.080 (beginning of supply squeeze)
• 2030: $0.080 (peak squeeze; BTM gas + grid premium)
• 2032: $0.078 (grid reform beginning to ease constraints)
• 2035: $0.075 (supply catching up; solar+storage share growing)
• 2040: $0.070 (diversified mix driving costs down)

Why this scenario is most likely: The supply response is real but takes time. The gas turbine bottleneck is genuine (sold out through 2030), but aeroderivative and reciprocating alternatives partially fill the gap. Grid reforms are underway but will take several years to clear backlogs. Solar+storage is cost-competitive but cannot yet provide 24/7 baseload without gas backup. The net effect is a modest cost plateau during 2028-2032 rather than a dramatic spike.

Conservative Scenario (for Terrestrial / Optimistic for Orbital)

Assumptions: Multiple supply constraints compound. Gas turbine shortage persists beyond 2030. BTM gas faces growing regulatory friction as methane regulations tighten and community opposition mounts. Grid interconnection remains 5+ years. Tariffs constrain solar/battery imports. Gas prices spike to $5-8/MMBtu. Demand growth at the high end of forecasts (McKinsey trajectory).

Cost trajectory (blended $/kWh, pre-PUE):

• 2026: $0.090 (tariff impacts, early supply tightness)
• 2028: $0.105 (full supply squeeze emerging)
• 2030: $0.110 (peak constraint period)
• 2032: $0.108 (slow easing)
• 2035: $0.105 (persistent premium; SMR delays)
• 2040: $0.095 (constraints gradually ease but remain above historical)

What would need to go wrong: This scenario requires several things to go wrong simultaneously: (1) gas turbine manufacturing fails to scale beyond current plans, (2) regulatory pushback successfully blocks or delays a significant share of BTM gas projects, (3) solar/battery import tariffs remain high, (4) grid interconnection reform fails to materially reduce timelines, and (5) demand grows at the high end of projections.

Even in this scenario, the electricity cost ceiling is bounded. At $0.110/kWh, data center operators would aggressively pursue every available alternative -- solar+storage with longer duration, geothermal in western states, existing nuclear PPAs, demand flexibility, efficiency improvements. The market would not passively accept $0.11/kWh when $0.05-0.06/kWh solar+storage is physically available; it would build out alternatives even if grid interconnection is slow.

5. Implications for the Orbital Argument

The proponent argument for orbital data centers rests on the premise that terrestrial electricity costs will rise to levels where space-based solar becomes competitive. Based on this analysis:

The argument has some basis in reality: Supply constraints are genuine. Grid interconnection queues are 8+ years. Gas turbine manufacturing is at ~90% utilization and sold out through 2030. PJM capacity prices have risen 10x. These are not speculative -- they are current conditions.

But the argument overstates the severity: The supply response is diverse and vigorous. BTM gas (56 GW already identified), solar+storage (at $57/MWh and falling), aeroderivative turbines, reciprocating engines, existing nuclear PPAs, grid reform, flexible interconnection (requiring only 7-35 hours/year curtailment to bypass transmission constraints volts-dc-flexibility-2026.1), and demand-side efficiency all provide alternatives. The terrestrial market has many ways to add supply, even if no single path is unconstrained.

The cost ceiling is irrelevant because energy cost cannot close the gap: Even in the conservative scenario, blended costs peak at ~$0.11/kWh. But the model-derived break-even analysis (see Answer section) shows that from 2030 onward, no terrestrial energy price — however high — would make orbital competitive. The orbital cost premium is structural (effective lifetime, cost of capital, GPU adaptation), not energy-driven. Energy cost is only ~6-7% of terrestrial TCO.

Localized vs. systemic constraints: The strongest version of the terrestrial constraint argument applies to specific geographies (Northern Virginia, parts of PJM) where demand is concentrated and grid infrastructure is genuinely maxed out. Virginia prices have risen 267% in five years eesi-dc-energy-bills.2. But data center operators are responding by relocating to less constrained markets (Texas, Ohio, the Southeast), which disperses demand and prevents systemic price spikes.

Evidence

Demand Projections

1. Goldman Sachs Research projects data center power demand will grow 165% by 2030 vs. 2023 levels, reaching approximately 122 GW globally. Current global usage is estimated at ~55 GW, split among cloud computing (54%), traditional workloads (32%), and AI (14%). AI share expected to reach 27% by 2027.

2. McKinsey projects 3.5x increase in data center capacity demand from 2025 to 2030. By 2030: 219 GW total global demand, with 156 GW for AI workloads (up from 44 GW in 2025). U.S. to grow from 25 GW (2024) to 80+ GW by 2030. $5.2 trillion in investment required for AI data center capacity alone.

3. IEA estimates global data center electricity consumption at ~415 TWh in 2024 (1.5% of global electricity). Base Case projects doubling to ~945 TWh by 2030 (~3% of global), and ~1,300 TWh by 2035. Data center electricity grows ~15% per year, 4x faster than all other sectors combined.

4. IEA notes 85%+ of new data center capacity additions over the next decade are expected in the United States, China, and the European Union. A tripling by 2035 represents less than 10% of total global electricity demand growth, but is highly concentrated geographically.

5. BloombergNEF forecast (December 2025): U.S. data center power demand could reach 106 GW by 2035, a 36% increase from their April 2025 forecast. Driven by 150+ significant projects announced in the prior year, over a quarter exceeding 500 MW each.

6. Epoch AI (January 2026): total AI data center power capacity reached ~30 GW by Q4 2025, comparable to peak power usage of New York State. Based on chip sales data with a 2.5x multiplier for facility overhead (servers, cooling, networking).

7. Epoch AI: global AI computing capacity has grown ~3.3x per year since 2022, with capacity doubling every ~7 months. NVIDIA accounts for over 60% of total compute, with Google and Amazon making up much of the remainder.

Grid Constraints

8. Grid interconnection timelines rose from an average of less than 2 years in 2008 to over 8 years by 2025. Utilities are experiencing more load growth in a single year than they used to see in ten or twelve.

9. PJM capacity market prices: $28.92/MW-day (2024/25), $269.92/MW-day (2025/26), $329.17/MW-day (2026/27) -- approximately a 10x increase in two years. The 2026/27 price would have been higher without a price cap.

10. Monitoring Analytics (PJM's independent market monitor) estimated data centers were responsible for 63% of the price increase in the 2025/26 capacity auction, translating to $9.3 billion in additional costs recovered from PJM ratepayers in higher electric rates.

11. In PJM's Dominion Zone (including Northern Virginia), the 2022 load forecast showed ~5,700 MW growth by 2037. The 2025 forecast shows over 20,000 MW from data centers alone by 2037 -- a nearly 4x upward revision.

12. Utilities received more than 700 GW of power connection development requests in 2025, exceeding the 477 GW total U.S. electricity consumption in 2023.

Electricity Price Impacts

13. U.S. average electricity price rose from ~13 cents/kWh (flat for over a decade pre-2019) to 19 cents/kWh by end of 2025, a 27% increase. Virginia prices increased up to 267% over five years due to data center concentration. Residential prices rose 11.5% in 2025, outpacing inflation. Prices expected to increase by up to 40% by 2030 compared to 2025.

14. Cost to build new gas plants tripled since 2022. Natural gas plants expected to enter service in 2030 or later report costs of $2,000/kW, which may soon rise to $3,000/kW. Utilities requested more than $29 billion in rate increases in H1 2025, double the amount from H1 2024.

15. Residential bill impacts in PJM: Washington D.C. (Pepco) saw $21/month average increase starting June 2025, approximately $10/month attributed to capacity market price spike. Western Maryland $18/month, Ohio $16/month. Starting June 2026, additional $1.4 billion in capacity market costs region-wide.

16. IEEFA notes "strong reasons to believe that PJM's 20-year forecasts of data center growth are inflated. But in the short term, markets are responding as though these forecasts are going to materialize."

Gas Turbine Supply Constraints

17. Three OEMs (GE Vernova, Siemens Energy, Mitsubishi Power) supply 75%+ of gas turbine projects. Mitsubishi: deliveries not until 2028-2030. Siemens: record backlog of EUR131 billion ($148B). GE Vernova: new turbines unavailable until late 2028 at earliest.

18. Utility planned gas capacity doubled from 25 GW (end of 2021) to over 45 GW (end of 2024) for deployment by 2030. Duke Indiana's Cayuga CCGT plant: $2,340/kW, 36% higher than resource plan estimate, translating to a $900 million cost overrun on a 1,476 MW facility.

19. NERC projects national peak demand to increase by 150 GW over the next decade (18% growth), while 120+ GW of existing generation (primarily coal) is expected to retire. The projected 270 GW gap requires rapid new construction.

20. GE Vernova targets 20 GW annualized turbine production by mid-2026, potentially stretching to 24 GW by mid-2028 at two existing facilities. Expects to end 2025 with an 80 GW gas turbine backlog stretching into 2029.

21. GE Vernova expects gas turbine reservations to be sold out through 2030 by the end of 2026. Siemens Energy nearly doubled units sold from 100 (2024) to 194 (2025) and committed $1 billion to U.S. manufacturing expansion.

22. Industry analysis: 6-7 year OEM delivery horizons. 20-30% EPC cost rise since 2021. Gas turbine manufacturing at ~90% capacity utilization in 2025. Data centers described as "new dominant offtaker" replacing traditional utilities.

23. BNEF LCOE 2026: combined-cycle gas turbine LCOE rose 16% to $102/MWh, the highest level on record. Data center demand has doubled U.S. gas turbine capex in just two years and pushed CCGT costs far above global averages. In the U.S., wind power has regained its position as cheapest new electricity generation, overtaking gas for the first time since 2023.

BTM Generation

24. At least 46 data centers with combined capacity of 56 GW are using behind-the-meter power generation. Over 1,000 GW of gas-fired power in development worldwide (31% increase year-over-year). More than one-third of new U.S. gas capacity targets data centers.

25. Texas has almost 58 GW of natural gas in planning and construction -- more than the next four states combined and more than every country except China. Nearly half of Texas power plants under construction provide power exclusively to data centers, without connecting to the grid.

26. Boom Supersonic: $1.25B deal with Crusoe for 29 jet-engine gas turbines (>1 GW). Turbine deliveries starting 2027. Superpower is a 42 MW natural gas turbine at $1,033/kW_gen, targeting 39% efficiency.

27. BorderPlex's Project Jupiter (New Mexico): $165B proposed data center campus with 2,880 MW of simple-cycle gas turbines. Behind-the-meter operation allows project to avoid state climate regulation compliance. Emissions could outweigh New Mexico's recent emission reduction actions.

28. McKinsey (via Jefferies) estimates 25-33% of incremental data center demand through 2030 will be met by BTM solutions, implying ~25-33 GW of BTM deployment over 5 years.

29. JLL sees utilities requiring million-dollar application deposits, letter of credit obligations, and take-or-pay policies covering 85% of requested power consumption, driving BTM adoption. But Schneider Electric CTO notes most major DC companies do not want to own/operate power systems long-term. AWS has stated its goal is to "get back to being grid-tied."

30. Cornell University analysis projects data center gas buildout could add 44 million metric tons of CO2 by 2030, equivalent to annual emissions from approximately 10 million passenger vehicles.

Solar + Storage Costs

31. BNEF: 4-hour battery storage LCOS fell 27% year-over-year to $78/MWh in 2025 -- a record low. 87 GW of combined solar+storage deployed in 2025 at average cost of $57/MWh.

32. BNEF: solar benchmark at $39/MWh (6% increase from prior year). By 2035, BNEF forecasts 30% solar LCOE reduction, 25% battery storage reduction, 23% onshore wind reduction, 20% offshore wind reduction. [Nuclear LCOE claim of $258/MWh removed -- not found in BNEF source]

33. [evidence:source-tbd] Global utility-scale BESS installed cost: ~$125/kWh outside U.S./China (~$75/kWh equipment, ~$50/kWh installation). U.S. costs higher due to tariffs and domestic content requirements (~$334/kWh NREL benchmark). [Source TBD -- previously misattributed to bnef-lcoe-2026; $334/kWh figure is from NREL]

Nuclear / SMR

34. [evidence:source-tbd] Constellation plans Three Mile Island Unit 1 restart (835 MW) for 2027-2028 at ~$1.6B total cost. $1B DOE loan secured November 2025. 20-year PPA with Microsoft. Not a colocation arrangement -- power will match Microsoft DC consumption in PJM. [Source TBD -- previously misattributed to eesi-dc-energy-bills]

35. [evidence:source-tbd] Meta signed deals with Oklo, Vistra, and TerraPower for up to 6.6 GW of nuclear power by 2035, including 20-year agreements for 2.1 GW from Vistra's Beaver Valley, Perry, and David-Besse nuclear plants. [Source TBD -- previously misattributed to eesi-dc-energy-bills]

36. SMR FOAK capital costs estimated at ~$14,600/kW (vs. ~$2,800/kW projected NOAK), with development timelines of 7-10 years. Costs typically decline 30-50% for subsequent units, but NOAK economics remain unproven.

Geothermal

37. Fervo Energy's Cape Station: 100 MW of carbon-free power by 2026, 500 MW by 2028. Largest next-generation geothermal development in the world. $462M Series E raised December 2025.

38. Fervo tripled drilling speed and halved per-well costs from ~$9M to $4.5M, demonstrating a 35% learning rate across first 8 wells. 80% of oil/gas workforce skills are transferable to geothermal.

39. Enhanced geothermal LCOE: currently ~$88/MWh with tax credits; projected $50-60/MWh by 2035 after learning-curve improvements. Below nuclear costs while maintaining clean energy profile. Project InnerSpace estimates ~3,400 GW of U.S. geothermal potential accessible with current drilling technology.

40. Analysis of 1 GW geothermal-powered data center: up to $3.2B savings over 30 years by using excess geothermal thermal energy for direct cooling (cooling represents 30-40% of total DC energy consumption in high-density AI facilities).

41. PJM faces 31 GW of projected data center load growth over five years, the largest single source of demand growth in the interconnection queue.

42. Utilities are experiencing more load growth in a single year than they used to see in ten or twelve years, driven primarily by data center buildout.

43. Meta's El Paso facility uses 813 modular mini-turbines (reciprocating engines). Reciprocating engines offer faster ramp-up (~1 min vs ~1 hour for CCGT) but lower fuel efficiency.

44. Two-thirds of U.S. gas project developers have not yet identified their turbine manufacturer, indicating severe supply chain risk in BTM gas buildout.

45. New Mexico's proposed Project Jupiter ($165B, 2,880 MW BTM gas) would "outweigh the actions that New Mexico has taken to lower emissions" in recent years, according to state officials.

46. Cornell University estimates data center gas buildout could add 44 million metric tons of CO2 by 2030 from BTM gas generation.

47. Center for Biological Diversity and other environmental groups are opposing BTM data center gas facilities over air quality degradation and climate impacts.

48. BNEF forecasts 30% solar LCOE reduction and 25% battery storage cost reduction by 2035, driven by continued technology learning rates and manufacturing scale.

49. Fervo Energy halved per-well costs from ~$9M to ~$4.5M and demonstrated a 35% learning rate across the first 8 wells at Cape Station.

50. Approximately 80% of oil and gas workforce skills are directly transferable to enhanced geothermal development, enabling rapid labor force transition.

Flexible Interconnection and Data Center Flexibility

51. Camus/Princeton Zero Lab/Encord study: optimal power flow modeling across six 500 MW data center sites within a single utility's territory found total annual curtailment of 7, 11, 13, and 35 hours respectively at the four constrained sites (two sites had no constraints at all). Maximum curtailment is 0.4% of hours per year. — volts-dc-flexibility-2026

52. At three of the four constrained sites, the longest individual curtailment events were 5 hours; at the fourth, 16 hours. Only 3-4 events per year. These durations are suitable for battery ride-through. — volts-dc-flexibility-2026

53. ERCOT has a "connect and manage" model for generation and is developing corresponding rules for flexible load interconnection. PJM is actively debating conditional high-impact load additions. SPP has proposed "hills, chills and spills" rules for conditional load. — volts-dc-flexibility-2026

54. Jesse Jenkins (Princeton Zero Lab) estimates the opportunity cost of delayed data center deployment at approximately $7 billion per GW per year. This makes even expensive flexibility solutions (e.g., substituting batteries for gas) cost-effective if they accelerate interconnection. — volts-dc-flexibility-2026

55. Jesse Jenkins: once flexible interconnection unlocks grid access, data centers can assemble portfolios of wind, solar, battery, and VPP resources already in the interconnection queue for 2028-2030 deployment, rather than building all generation on-site. His company Firma Power is commercializing this approach. — volts-dc-flexibility-2026

56. Most data centers operate at ~40% average utilization vs nameplate capacity. Training workloads cycle between 100% usage and off periods; inference and cloud workloads show more organic variability. Google can shift data center load between sites in seconds using built-in reliability infrastructure. — volts-dc-flexibility-2026

57. Halcyon tracks 85 GW of gas plant additions currently planned across the U.S. — a "potpourri of different designs" including combined cycle, simple cycle gas turbines, and reciprocating engines. A rapidly growing share is behind-the-meter for data centers. — volts-dc-flexibility-2026

58. Astrid Atkinson (Camus CEO, former Google reliability engineering): on-site fully off-grid natural gas generation is "not great from a data center perspective" due to poor reliability profile, need for significant redundancy, vulnerability to gas price fluctuations, and deliverability challenges. Utilities report "there's no such thing as a fully off-grid data center" in practice. — volts-dc-flexibility-2026