Terrestrial Data Center Energy Cost

Answer

The effective blended electricity cost for AI data centers in 2026 is approximately $0.065-$0.090/kWh (pre-PUE), depending on location, procurement strategy, and power mix. This represents a significant increase from the $0.04-$0.06/kWh range that prevailed in 2020-2023, driven primarily by capacity charge escalation in PJM and general grid tightness.

The cost trajectory through 2040 depends critically on the mix of grid-connected, behind-the-meter (BTM), and off-grid power. In the central case, costs peak around 2028-2030 at ~$0.08/kWh as BTM gas generation and grid buildout gradually relieve supply constraints, then decline modestly through the 2030s as BTM renewables+storage and potentially SMRs come online. In the optimistic case, aggressive BTM deployment (solar+storage at $0.057/MWh LCOE already demonstrated) combined with off-grid gas at competitive rates drives costs down to ~$0.045/kWh by 2040. In the conservative case, structural grid constraints, carbon pricing, and gas price volatility push costs above $0.10/kWh before moderating.

Central estimates ($/kWh, pre-PUE, 2025 USD):

• 2026: $0.075
• 2030: $0.080
• 2035: $0.075
• 2040: $0.070

Evidence

A. Current Grid-Connected Electricity Costs

A1. [evidence] The U.S. average industrial electricity rate was approximately $0.047-$0.15/kWh in 2025, with enormous regional variation. Iowa and Texas offered industrial rates around $0.06/kWh, while California exceeded $0.176/kWh. Virginia, hosting 26% of U.S. data centers, averaged $0.1594/kWh residential but data centers paid significantly less through negotiated commercial/industrial tariffs. (Sources: CNBC state data center concentration article; RealClearEnergy datacenters-electricity-costs-2026)

A2. [evidence] Target wholesale electricity rates for major data centers have historically been below $0.05/kWh, with some facilities securing rates below $0.04/kWh in regions with abundant hydroelectric power (e.g., Pacific Northwest, Quebec). However, these rates are pre-capacity-charge and pre-transmission. (Source: solartechonline.com data center electricity guide)

A3. [evidence] Between 2020 and 2024, residential electricity prices rose 25%. Commercial prices rose only 3% over two years (2022-2024), and industrial prices fell 2% over the same period. Data centers are consuming more power but paying proportionally less, thanks to negotiated PPAs and industrial tariffs. (yale-dc-electricity-rates)

A4. [evidence] U.S. residential electricity prices rose from $0.1276/kWh in 2020 to $0.1744/kWh in February 2026, a 36% increase. They are projected to reach $0.1901/kWh by September 2027. (cnbc-footing-ai-bill)

B. Capacity Charges (PJM)

B1. [evidence] PJM capacity prices jumped 9.3x from $29/MW-day (2024/25 delivery year) to $270/MW-day (2025/26), with constrained zones in Virginia and Maryland seeing $400+/MW-day. Subsequent auctions hit the price cap of $329-$333/MW-day for 2026/27 and 2027/28. (semianalysis-dc-electric-bills; rmi-pjm-speed-to-power)

B2. [evidence] At $329/MW-day with a 40% load factor, the capacity charge alone translates to approximately $0.034/kWh (3.4 cents/kWh). This is layered on top of wholesale energy costs. For PJM-connected data centers, the all-in effective rate is therefore wholesale energy ($0.03-$0.05/kWh) + capacity ($0.034/kWh) + transmission/distribution ($0.01-$0.02/kWh) = approximately $0.074-$0.104/kWh. (semianalysis-dc-electric-bills)

B3. [evidence] The PJM Independent Market Monitor found that removing all data centers from the forecast reduced peak load by 7,927 MW and would have cut total capacity payments by $9.33 billion (64%). Data centers were 40% of costs in the December 2025 auction for 2027/28. (semianalysis-dc-electric-bills; volts-pjm-explainer)

B4. [evidence] PJM's total capacity bill rose from $2.2 billion (December 2022 auction) to $16.1 billion. Pennsylvania Governor Shapiro called this "the largest unjust wealth transfer in the history of US energy markets." (rmi-pjm-speed-to-power)

C. ERCOT (Texas) Pricing

C1. [evidence] ERCOT uses real-time scarcity pricing (ORDC) instead of a capacity auction. Normal prices are $10-$50/MWh ($0.01-$0.05/kWh), with a scarcity cap of $5,000/MWh. Forward prices for 2026-2030 contracts increased 11-17% year-over-year but saw no PJM-style 9x surge. (semianalysis-dc-electric-bills)

C2. [evidence] ERCOT projected 77.9 GW of potential data center load by 2030 but applied significant haircuts (49.8-55.4% discount) to developer claims. The grid handled record peaks of 90+ GW in summer 2024 without brownouts. Real-time prices during Winter Storm Fern (January 2026) peaked around $300/MWh, while PJM averaged $700/MWh with Virginia's Dominion zone spiking to $1,800/MWh. (semianalysis-dc-electric-bills)

C3. [evidence] Wholesale electricity near data center clusters cost as much as 267% more than five years ago, per Bloomberg analysis (September 2025). (cnbc-electricity-prices-inflation)

D. Behind-the-Meter Generation Costs

D1. [evidence] Lazard LCOE+ 2025: CCGT $48-$107/MWh (midpoint ~$72.5/MWh). Gas peaking $149-$251/MWh. CCGT costs at a 10-year high due to turbine shortages and rising costs. (lazard-lcoe-2025)

D2. [evidence] BNEF 2025 data: 4-hour battery storage LCOS fell 27% YoY to $78/MWh globally, a record low. Combined solar+storage delivered at $57/MWh average in 2025 (87 GW deployed). Fixed-axis solar benchmark rose to $39/MWh. BNEF forecasts 30% solar LCOE reduction and 25% battery storage reduction by 2035. (bnef-lcoe-2026; bnef-battery-costs-2025)

D3. [evidence] ScienceDirect comparative analysis found solar+battery storage as the lowest-cost option for data centers at $25.11/MWh ($0.025/kWh), though this is sensitive to CAPEX and capacity factors and does not account for firmness requirements (data centers need 24/7 power, not intermittent). (sciencedirect-dc-lcoe-comparison)

D4. [evidence] 56 GW of BTM generation planned for U.S. data centers, representing ~30% of projected DC capacity. ~90% of BTM announcements came in 2025. McKinsey estimates 25-33% of incremental DC demand through 2030 will be met by BTM solutions. (latitude-btm-traction)

D5. [evidence] Major BTM gas announcements: Stargate (7 GW planned, gas-fired BTM); Joule 1.3 GW fully islanded in Utah; VoltaGrid/Energy Transfer 2.3 GW for Oracle; FO Permian/Hivolt 5 GW off-grid gas in Texas; International Electric Power 944 MW gas in Pennsylvania. (latitude-btm-traction)

D6. [evidence] Google acquired Intersect Power for $4.75 billion: 640 MW solar, 1.3 GWh battery, gas backup. Sites designed as self-contained microgrids with solar providing 50-70% of annual energy. Hyperscalers signed 40+ GW of solar PPAs in 2025. (google-intersect-acquisition; hyperscaler-solar-2025)

D7. [evidence] Solar PPA prices for data centers are rising: P25 solar prices rose 3.2% in Q4 2025, up ~9% year-over-year, as hyperscaler demand compresses available supply. (pv-magazine-solar-ppa-playbook)

E. Off-Grid Gas Generation

E1. [evidence] FTAI Power (launched January 2026) converts surplus CFM56 aircraft engines into 25 MW power units. FTAI owns or controls 1,000+ engines and plans 100+ turbine units annually. No public pricing disclosed, but positioned as "cost efficient." (ftai-power-launch)

E2. [evidence] Boom Superpower: 42 MW natural gas turbine. Crusoe paying $1,033/kW of generating capacity (capital cost only, not LCOE). Targets 39% efficiency; maintains full nameplate capacity at 110°F+ without water. (boom-superpower-launch)

E3. [opinion] At $1,033/kW capital cost, 39% efficiency, and $3.50/MMBtu gas: fuel cost alone ~$0.031/kWh. Adding O&M ($0.005-$0.01/kWh) and amortized capital over 15-20 years (~$0.008-$0.012/kWh at 90% CF), the all-in LCOE for BTM gas turbines is roughly $0.045-$0.055/kWh. This is competitive with grid power in PJM but does not include carbon costs. (Derived from boom-superpower-launch data and standard engineering economics)

E4. [evidence] SemiAnalysis reports 15 different equipment manufacturers now have >400 MW orders for onsite gas generation. Diesel generator manufacturers are pivoting to onsite gas, adding 10+ GW of manufacturing capacity. The market constraint is shifting from energy availability to construction capacity. (semianalysis-dc-electric-bills)

F. Nuclear and SMR Costs

F1. [evidence] SMR LCOE estimates: First-of-a-kind (FOAK) ~$180/MWh; Wood Mackenzie projects 40% reduction to ~$100/MWh by 2030. NuScale target rose from $58/MWh (2021) to $89/MWh. Rolls-Royce SMR targets below $80.9/MWh (GBP 70/MWh). Competitive range for grid deployment: $52-$119/MWh. (Multiple SMR analysis sources)

F2. [evidence] Talen Energy sold a 960 MW nuclear data center campus at Susquehanna to AWS for $650 million. Constellation Energy's 20-year PPA with Microsoft supports Three Mile Island Unit 1 restart. Aalo Atomics and Valar Atomics broke ground on pilot reactors at DOE sites in 2025. (balerion-kilowatts-to-compute)

F3. [opinion] SMRs are unlikely to materially affect data center electricity costs before the mid-2030s at earliest. Even optimistic timelines show first commercial deployments in 2030-2032, with costs remaining at $80-$120/MWh for initial units. Nth-of-a-kind costs may reach $60-$80/MWh by late 2030s, competitive with CCGT but requiring significant regulatory and manufacturing scale-up. (Synthesis of F1, F2)

G. Structural vs. Cyclical Factors

G1. [evidence] Goldman Sachs projects household electricity prices to rise an additional 6% through 2027, then slow to 3% in 2028 on lower natural gas prices. Data centers make up 40% of electricity demand growth. (cnbc-electricity-prices-inflation)

G2. [evidence] PJM's interconnection queue has been closed to new entry from 2022 through spring 2026. It now takes 8 years to bring new generation online, but capacity auctions plan 2-3 years ahead. This structural mismatch drives the capacity price spike. (rmi-pjm-speed-to-power; volts-pjm-explainer)

G3. [evidence] ERCOT moves much faster than PJM (single-state, not subject to FERC jurisdiction). Forward prices show traders believe ERCOT can absorb growth via supply expansion and SB 6 curtailment authority. (semianalysis-dc-electric-bills)

G4. [evidence] The constraint is shifting from energy to construction. SemiAnalysis reports more "serious" datacenter supply available in 2027 than actual demand for AI and non-AI compute, as onsite gas has dramatically alleviated energy constraints. (semianalysis-dc-electric-bills)

G5. [evidence] Duke Energy study found the grid could integrate 76-126 GW of flexible DC load with only 22-88 hours/year of curtailment, suggesting the physical grid can handle growth if interconnection/regulatory barriers are addressed. (duke-flexible-load-study)

H. Benchmark: Project Suncatcher Comparison

H1. [evidence] Project Suncatcher (2025 white paper) estimates terrestrial data centers spend $570-$3,000 per kW per year on power, depending on local costs and PUE. SpaceX/Starlink satellites deliver energy at $14,700/kW-year, roughly 5-25x more expensive. (techcrunch-orbital-ai-economics)

Analysis

Decomposing the Effective Rate

The effective electricity cost for a grid-connected AI data center has three main components (all pre-PUE):

Post-PUE (applying a PUE of 1.2 for modern liquid-cooled AI facilities), the IT-load-adjusted cost is 20% higher. A $0.075/kWh pre-PUE rate becomes ~$0.090/kWh at the chip.

Why Grid Costs Are Elevated

The current cost spike is driven by a convergence of factors:

1. PJM capacity market dysfunction (B1-B4): An 8-year interconnection queue combined with a 2-3 year forward capacity auction creates structural inability for supply to respond to demand signals. This is partly cyclical (queue reopening in 2026) but will persist as long as demand growth outpaces interconnection speed.

2. Wholesale energy inflation near DC clusters (C3): Bloomberg found 267% wholesale price increases near data center concentrations, reflecting localized transmission congestion.

3. National demand-supply imbalance (G1): Data centers drive 40% of demand growth while generation additions lag. Goldman projects this persists through 2027-2028.

The BTM Revolution

The most important trend for cost trajectory is the massive shift to BTM and off-grid generation (D4-D5). Key dynamics:

• Gas BTM ($0.045-$0.065/kWh all-in): Simple-cycle and reciprocating engine installations at $800-$1,200/kW capital cost, 35-45% efficiency. Fuel cost of ~$0.03/kWh at $3.50/MMBtu gas. Already being deployed at multi-GW scale. This is the bridge solution (2025-2032).

• Solar+storage BTM ($0.025-$0.060/kWh LCOE): Already the lowest LCOE option where solar resource is strong, but intermittency means it supplements rather than replaces firm power for data centers. The $57/MWh solar+storage figure from BNEF represents co-located projects, not standalone DC power. Data centers need 24/7 power, so the effective cost of solar+storage as a complete solution requires significant overbuild or gas backup, raising the blended cost to $0.04-$0.06/kWh.

• Nuclear BTM ($0.06-$0.12/kWh): Existing nuclear PPAs (Constellation/Microsoft, Talen/AWS) offer $0.06-$0.08/kWh with zero carbon. SMRs will be more expensive initially ($0.10-$0.18/kWh FOAK) but could reach $0.06-$0.08/kWh at scale by late 2030s.

K-Shaped Pricing

Evidence strongly supports K-shaped pricing divergence (A3, A4, yale-dc-electricity-rates):

• Residential rates: rising 5-7%/year, projected to continue
• Industrial/DC rates: flat to modestly increasing (3% commercial increase vs 25% residential over 2020-2024)
• Large data center operators with negotiated PPAs and BTM generation are insulated from capacity and transmission cost spikes
• This divergence is likely to persist and widen as more DCs go BTM, effectively opting out of the capacity market

Scenario Trajectories

Optimistic ($0.065 -> $0.045/kWh):

• Assumes aggressive BTM deployment succeeds at scale
• Solar+storage LCOE continues declining (BNEF: -30% solar, -25% storage by 2035)
• Gas BTM bridges to renewables+storage
• PJM reforms accelerate interconnection; capacity prices moderate
• ERCOT-like markets gain share of DC location decisions
• SMRs reach competitive costs by late 2030s
• No significant carbon pricing imposed on BTM gas

Central ($0.075 -> $0.070/kWh):

• BTM gas handles 25-33% of DC load through 2030 (per McKinsey)
• Grid costs peak 2027-2029 as PJM queue reopens and new supply comes online
• Solar+storage costs decline but intermittency limits penetration to 30-40% of DC energy mix
• Modest carbon pricing ($25-$50/ton) adds $0.01-$0.02/kWh to gas generation by mid-2030s
• SMRs contribute marginally by 2038-2040
• Net effect: costs stabilize and gradually decline, reaching $0.07/kWh blended by 2040

Conservative ($0.090 -> $0.095/kWh):

• Grid constraints prove more persistent than expected
• Gas price volatility (LNG exports, cold winters) drives fuel costs to $5-$7/MMBtu
• Carbon pricing accelerates: $50-$100/ton by 2035, adding $0.02-$0.04/kWh to gas
• BTM deployment faces permitting/community opposition
• SMRs delayed to 2035+; costs remain above $0.10/kWh
• Electricity demand growth from electrification (EVs, heat pumps) compounds DC demand
• Effective rate stays elevated at $0.095-$0.11/kWh through 2035, declining only as solar+storage matures

Why Costs Don't Simply Collapse

Several factors prevent electricity costs from falling to the theoretical minimum of solar+storage LCOE ($0.025-$0.04/kWh):

1. Firmness premium: AI training requires 24/7 power at >99.99% reliability. Achieving this with intermittent sources requires massive overbuild (3-5x) plus storage duration beyond current 4-hour lithium-ion batteries, or gas/nuclear backup.

2. Land and transmission constraints: Even BTM solar requires ~5-7 acres per MW; a 500 MW DC needs 2,500-3,500 acres of panels plus battery banks.

3. Construction bottlenecks: The constraint is shifting from energy availability to construction capacity (G4), which means costs reflect construction labor and materials inflation.

4. Policy risk: Carbon pricing, methane regulations, and community opposition to BTM gas create upward cost pressure that partially offsets technology learning curves.

Summary Table

All values in 2025 USD, pre-PUE. To convert to post-PUE (at-the-chip), multiply by PUE (1.15-1.25 for modern AI facilities).