Data-center power · U.S. regulation · 2026.
The Megawatt Gate
The line that triggers scrutiny can be as low as 20 MW—or as high as 75 MW.
Electrical demand—not square footage—sets the trigger. DOE’s 20 MW line is an October 2025 proposal, not a FERC standard. New York’s legislature passed a 20 MW bill, but it remained unsigned in the latest reporting; Pennsylvania’s order covers projects above 25 MW.
FERC left “large load” definitions to grid operators; law-firm summaries described 50 MW at 69 kV, non-co-located, as reasonable. PJM counts 50 MW at one site, including affiliated load within one mile; New York EO 62 also uses 50 MW. ERCOT starts at 75 MW. These are different proposals, orders and market definitions—not interchangeable rules.
Forecast data-center load behind PJM’s auctions rose from 7,927 MW in 2025/2026 to 21,470 MW in 2028/2029. For January–June 2026, wholesale power cost reached $114.50/MWh versus $76.16 a year earlier; Monitoring Analytics estimated data-center effects added $11.11/MWh, or 9.7%.
The attribution is a market-monitor counterfactual, not an independently replicated finding. Monitoring Analytics also advocates moving data-center load into a dedicated capacity auction.
“Bring capacity—or face earlier curtailment.”
— Under PJM’s proposal, uncovered load would be curtailed before existing pre-emergency demand response.
California’s September 2026 package requires electricity and water disclosure, payment for certain grid and water upgrades, and state environmental review. SB 886 directs utility rules by January 1, 2028 to prevent cost shifting and stranded costs.
Pennsylvania is seeking data-center-first curtailment unless full capacity is secured—and data-center payment of backstop costs. Arizona froze new data-center sales-tax breaks for three years.
Data Center Watch’s project values combine blocked and delayed outcomes. The figures come from a single report and were not independently reviewed.
Power constraints can force developers to fund capacity, post security, absorb grid upgrades, accept curtailment or wait for permits and energization—material risks to project timing and capital needs.
But the record identifies no project-level loan amounts, lenders, balances, rates, maturities, defaults or loan-to-cost ratios. It also does not consistently separate AI facilities from other data centers—or ERCOT crypto loads. Quantifying lender exposure requires project-level financing, interconnection and construction-timeline data.
“The capital-stack implication is plausible; the loan book is not in this record.”
— The evidence boundary