For US AI data centres, the wire still wins: once reliability charges and firm-grid relationships are included, on-site generation costs exceed the delivered grid in all modeled scenarios—combined-cycle gas costs about $114/MWh complete-site versus a $92/MWh delivered grid, while hydrogen-based options remain two-to-six times costlier even under aggressive electrolyser learning.

Main Finding

On a complete-site basis (including retained grid capacity/network charges and measured GPU training load), no on-site supply technology in the paper’s sample beats the delivered grid in the modeled U.S. 2030–2040 scenarios. Even though some generators (e.g., combined‑cycle gas) have low busbar costs, once the facility keeps the grid for peaks, outages, and reserve capacity the total delivered cost to the data center exceeds the grid tariff. Hydrogen pathways only approach competitiveness—at roughly 2× the delivered grid—after massive electrolyser deployment and large public/private capital (hundreds of billions to trillions USD), and storage is physically limited in annual energy share and often dirtier than the grid when charged at the margin.

Key Points

• Accounting choice matters: comparing busbar LCOE to a fixed grid price answers a question developers don’t face. This paper prices technologies against an endogenous delivered-grid tariff and on a complete-site basis (so fixed grid charges largely remain).
• Measured AI load: accelerator telemetry (A100-class RL training) shows fast, high-amplitude ramps (majority of samples exceed ±200 W/s). Rather than exclude slow generators, the model prices a small battery buffer for generator response time (an adder of ~$0.9–$3/MWh).
• Endogenous grid: projected data-center demand is routed into the power sector (60% of incremental load → gas burn) which raises gas prices and escalates capacity/network charges; the delivered grid price rises with the modeled demand (reference: ~$74/MWh in 2026 → ~$92/MWh in 2030).
• Complete-site vs busbar: example—combined-cycle gas busbar LCOE ≈ $47/MWh (2030), but the complete-site cost with retained grid ≈ $114/MWh, i.e., about $21–$22/MWh above the delivered grid ($92/MWh).
• Storage limits: 4‑hour lithium BESS has a practical annual-energy cap (~17.8% / 16% CF at realistic cycling) and, when charged at the margin, can be dirtier than the grid. Corrected incremental-charging cost ≈ $118/MWh (at physical ceiling).
• Hydrogen economics: electrolytic hydrogen requires very low-priced power (~$20/MWh) to approach competitive production cost. Under an optimistic S4 scenario (co‑located electrolysis on curtailed renewables + tax credit), delivered H2 can fall to ~$3.8/kg and hydrogen power ~1.9–2.2× the grid, but that requires large-scale electrolyser deployment and policy support.
• Investment inversion: the paper computes the capital required (via Wright’s‑law learning) to push costs down. Key result: conversion-hardware learning alone cannot buy parity—global electrolyser deployment on sub-$20/MWh power would still leave hydrogen power ~2× the grid, and bringing parity is not purchasable at any realistic budget in the model.
• Policy mechanics: Section 45V (clean-H2 PTC) and Section 48E (investment tax credit for zero-GHG assets) materially affect hydrogen and storage economics, but eligibility and tiering limit their reach; they improve but do not create grid parity.

Data & Methods

• Technology set: nine supply technologies — combined-cycle gas, gas peakers, 4‑hour Li-ion storage, SOFC (pipeline NG & RNG), PEM fuel cell (delivered H2), H2 reciprocating engine, RNG reciprocating engine, HVO genset.
• Parameter sources: published cost surveys, vendor deployment data, national-lab lab costings, lifecycle-assessment literature. Minimum practical unit sizes enforced.
• Load data: measured accelerator power telemetry from reinforcement‑learning training runs on A100-class hardware. Facility load constructed via bootstrap with appropriate block length; generator ramp gaps bridged by a priced battery buffer sized to generator response time.
• Cost model: annuity LCOE at 7.7% WACC with technology-specific lifetimes, O&M, delivered fuel prices, compliance carbon pricing applied to stack emissions only; lifecycle carbon reported separately. Storage modeled as a system component; charging uses incremental (marginal) tariff.
• Endogenous coupling: incremental DC demand routed to power-sector gas (60% of increment) with supply elasticity 0.35; peak-demand growth escalates capacity/network charges at 5× proportional rate up to a cap (1.6× 2026 charge). This produces delivered-grid prices that rise with DC demand.
• Scenarios: four main scenarios (S1 Reference; S2 Grid-constrained/high gas; S3 Carbon‑priced; S4 Hydrogen-scales with $20/MWh electrolyser power and production tax credit). Hydrogen price is derived (not assumed) from electrolyser capital costs under Wright’s law plus electricity price.
• Representative quantitative outcomes (2030 reference):
  • Delivered grid (endogenous): ~$92/MWh
  • Combined-cycle gas busbar: $47/MWh; complete-site: $114/MWh; lifecycle CI ≈ 431 gCO2e/kWh
  • 4‑hr BESS incremental charging LCOE: $118/MWh; complete-site priced at its physical annual-energy cap (~17.8%)
  • PEM fuel cell (delivered H2): busbar ≈ $490/MWh; complete-site ≈ $556/MWh; lifecycle CI ≈ 915 gCO2e/kWh
  • Hydrogen delivered prices vary across scenarios (e.g., ~ $3.8/kg in S4; ≈ $7–8/kg in reference)
• Open reproducibility: model, revision log, and workbook reproducing tables are released with the paper.

Implications for AI Economics

• Siting and total cost of ownership: developers cannot treat generator busbar LCOE as the facility-level electricity cost. Complete-site accounting is necessary for realistic comparisons; availability of the wire is an economic as well as access constraint (interconnection queues drive decisions).
• Investment priority: the most valuable investments for reducing AI data-center electricity cost are not necessarily conversion hardware (electrolysers, fuel cells). Instead, addressing interconnection bottlenecks, expanding transmission and local distribution capacity, and securing access to very low‑cost curtailed renewables (for cheap electrolysis) provide more leverage.
• Role of natural gas: on-site combined-cycle gas is the lowest-cost on-site generation in many scenarios but still typically more expensive than the retained-grid delivered energy. Where the wire isn’t available (interconnection scarcity), on-site gas may be the pragmatic option despite higher complete-site cost.
• Storage and demand-shaping: four‑hour batteries have constrained contribution to annual energy (~≤18%) and, when charged at the incremental margin, can increase facility carbon intensity. However, small buffer batteries are inexpensive relative to energy and necessary to bridge generator response for AI loads—so operational design (power control, flexible scheduling) that allows buying curtailed renewables or shaping training load matters more.
• Hydrogen as deep decarbonization path: hydrogen-based on-site power can provide deep carbon reductions (model shows ~85% lifecycle reduction relative to grid-power production in optimistic deployments) but at very high cost and very large capital deployment. Hydrogen can be part of a long-run decarbonization strategy, but it is not a near-term route to grid parity for AI loads without unrealistic budgets or systemic changes to power prices.
• Policy design considerations: tax credits and production incentives (45V/48E) materially affect cost trajectories but do not eliminate the fundamental gap for most technologies; careful eligibility, tier design, and duration influence whether incentives meaningfully alter siting decisions.
• Modeling guidance for AI economists: (1) use measured load profiles for LCOE comparisons and price buffers for generator response times; (2) compare against an endogenous delivered-grid tariff when demand is large enough to affect the market; (3) conduct investment-inversion analyses to answer “how much money and where must it flow to change the ranking?”

Limitations noted by the paper: the gas-price fit is the weakest model element (limited data for long-run exports), solar logistic ceiling is a key driver of grid carbon outcomes, and elasticities/judgments (e.g., 60% of incremental DC load → gas, network escalation multiplier) are transparent but uncertain—sensitivity analysis is provided.