Who Pays for the Buildout

Issue 06 found the residential bill is not a data-center story over the decade, real but concentrated in the PJM belt now, and modest and forward for everyone as the net-new build is recovered. This companion takes up the two questions underneath that forward cost: whether the demand shows up at all, and who carries it if it does.

Read the companion deck (PDF)→

Issue 06 and this companion are our attempt to understand one piece of a large and fast-moving question, how much of the rising electricity bill traces to AI data centers, and who carries the cost of what the buildout requires. We are not utility regulators, grid engineers, or credit analysts, and a specialist may see nuance we have missed. What we have done is follow the primary evidence where it leads, the demand forecasts, the turbine order books, the behind-the-meter trackers, and the rate filings, and flag every figure that is an estimate or rests on a single source. We hold the conclusions as a working hypothesis open to correction.

The demand forecast has outrun what will connect

The forecast that justifies the buildout has grown faster than anything it describes. Grid Strategies' estimate of five-year nationwide peak-demand growth ran 23, 39, 67, and 166 gigawatts across its 2022 through 2025 reports, more than a sixfold revision in three years on the forecaster's own accounting. Data centers are about 90 gigawatts, roughly 55 percent, of the latest figure (Exhibit 1). The forecaster flags its own data-center number as roughly 40 percent too high against analyst benchmarks.

Underneath the forecast the queue is worse. Developers file five to ten times more interconnection requests than projects that get built, on RMI's estimate. In the separate generation-interconnection queue Lawrence Berkeley finds only 19 percent of requests from 2000 to 2019 reached commercial operation, a different process but the same pattern of speculative filing. LBNL puts data centers at 4.4 percent of US electricity in 2023, rising to somewhere between 6.7 and 12 percent by 2028, a range wide enough to drive a truck through. We mark the net new data-center load that actually connects near 55 to 90 gigawatts by 2030, Grid Strategies' 90 less its own overstatement, and still well under the gross request counts, ERCOT's large-load queue alone reaching about 239 gigawatts by March 2026. The number under every capacity plan is a gross count, and the operators running the grids know it.

The binding near-term constraint is the turbine

The load that is real cannot connect on the schedule the announcements imply, because the equipment is sold out. GE Vernova's gas-turbine backlog reached about 80 gigawatts stretching into 2029, its slots are effectively booked through 2030, and a turbine ordered today starts up around 2031 (Exhibit 2). Siemens Energy says roughly 60 percent of its gas-turbine orders now come from US data centers, and turbine prices have risen about 300 percent in three years on trade-press estimates.

The bottleneck runs past the turbine hall. Large transformers that took months in 2020 now take two to four years, and a new substation interconnection runs five to eight. So the tariff docket, real as it is, is not the near-term limit. A megawatt of firm capacity cannot be conjured by a rate case faster than the supply chain can build it, and the supply chain is the governor into the back half of the decade.

A quarter of the planned load is arranging to leave the grid

A growing share of the buildout is arranging to bypass the grid the tariff fight is about. On Cleanview's tracker, about 90 gigawatts of announced US data-center capacity is behind-the-meter, more than a quarter of everything planned, with 92 percent of it announced since January 2025 (Exhibit 3). The marquee case is Amazon's 1.9-gigawatt power deal at Talen's Susquehanna nuclear plant. Load that self-supplies neither burdens the ratepayer nor dilutes the ratepayer's fixed costs, so the bypass cuts both ways in the who-pays question.

Two cautions temper it. Only about 2 of the 90 gigawatts is operating, so behind-the-meter carries the same forecast-inflation caveat as the queue. And on December 18, 2025 FERC found that netting a data center's on-site generation against its grid draw shifts cost onto other customers, and ordered PJM to write new terms, so the regulator is already moving to close the bypass it reads as a subsidy.

What gets firmed runs on gas

The firm megawatt that serves a data center around the clock is, at the margin, a gas megawatt. Forecasts of the incremental US gas demand from data centers by 2030 run from about 2 to 6 billion cubic feet a day, roughly 2 to 5 percent of national demand, and the dispersion across Enverus, EIA, S&P Global, East Daley, and RBC is wide enough to signal how little is settled (Exhibit 4). Wood Mackenzie counts 58 gigawatts of gas plants in development in Texas alone. The clean-firm alternatives, restarted nuclear and the first small modular reactors, are real but small and priced at a premium, so the near-term firming is gas. That ties the buildout back to gas-turbine and pipeline capacity, the same supply chain that already governs how fast the load can connect.

Regulation puts the stranding risk on ratepayers, the market puts it on equity

The wires, substations, and generation built for a data center are recovered over thirty to forty years, the standard life of a rate-base asset. The commitment on the other end is short, and it can move. A hyperscaler sites its demand where power, land, and latency are cheapest and shifts it when they change, the queued project behind it can be cancelled before a shovel goes in, and the firm contract is measured in a few years. Even AEP Ohio's tariff, written to hold the load in place, pins a departing data center to only about three years of minimum charges on a contract that runs at most twelve. That is the mismatch under the whole who-pays question, forty-year money committed against a counterparty that has firmly promised a few years of it, and a stranded plant for someone if the load never arrives. Which balance sheet that someone sits on depends on which side of the regulatory line the asset falls.

A regulated utility earns a commission-set return on its rate base and recovers that base from customers whether or not the data center it was built for ever draws power, so the utility's shareholders are largely insulated and its ratepayers are the ones left holding it. Harvard's Electricity Law Initiative documents the channels, the special contracts and cost allocations that quietly move data-center costs onto the general ratepayer across several states.

On the other side of the line, a merchant generator or a neocloud that builds against the same load carries the residual itself, in its equity and its debt, with no rate base to fall back on. That is why the retail-price panic points at the wrong balance sheet. The reprice that is coming lands on the merchant and the lender; the regulated utility, whose return is set by a commission, is comparatively insulated. That seam is the one Issue 07 traces into the credit market.

The net load is a fraction of the queue, and the reprice lands on the merchant

The demand worth pricing is the net figure, near 55 to 90 gigawatts of new data-center load by 2030, a fraction of the gross queue the forecaster itself marks down. The binding near-term constraint is physical, since gas turbines are booked through 2030 and large transformers run two to four years, so the load connects on the supply chain's clock and the equipment makers are the balance sheet the buildout most reliably feeds. About a quarter of planned capacity is arranging to leave the grid behind the meter, large and rising but mostly still on paper, and FERC is already moving to charge it. What gets firmed runs on gas, two to six billion cubic feet a day by 2030, with restarted nuclear and the first small modular reactors real but small and dear. The stranding risk sorts by ownership, sitting with ratepayers under regulation and with equity and credit in the merchant world, so the reprice to watch is on the unregulated generator and its lenders, the seam Issue 07 takes up.

The merchant and lender exposure this issue locates is the subject of Issue 07. Issue 07 also carries the power constraint forward, finding that a GPU's economic life is set by the price of power, so where watts are scarce, the condition these issues have traced, the old chip gives way fastest and its residual, the collateral under data-center credit, erodes soonest.

Five-year peak-demand forecasts and the data-center share from Grid Strategies' 2022 through 2025 reports; interconnection-request multiples from RMI; the 19 percent generation-queue completion rate and the data-center share of US electricity from Lawrence Berkeley National Laboratory; the ERCOT large-load queue of about 239 gigawatts as of March 2026 from ERCOT; the gas-turbine backlog and lead times from GE Vernova and Siemens Energy, with turbine-price and transformer and substation lead-time figures from trade-press estimates; the behind-the-meter tally from Cleanview's tracker, and the Amazon power deal at Talen's Susquehanna plant; the FERC order of December 18, 2025 on co-located large loads; incremental gas-demand forecasts from Enverus, EIA, S&P Global, East Daley, and RBC, with the Texas gas-plant count from Wood Mackenzie; and the ratepayer cost-shift channels from Harvard's Electricity Law Initiative.

The demand forecasts are estimates that the forecaster itself revises and marks down, and the 55-to-90-gigawatt net figure is our own read against those benchmarks. Queue and behind-the-meter totals are announced or requested capacity, not committed or operating load; only about 2 of the roughly 90 behind-the-meter gigawatts is running. The gas-demand range is wide and unsettled, and the turbine-price and lead-time figures rest on trade-press reporting. Second Order is an independent research briefing, for discussion only and not investment advice.