Congress punts on who pays for AI's grid, so the states are writing the bill

I.A voluntary bill died, a binding package passed

On 30 September the Senate voted 57–43 to advance the Ratepayer Protection Act, three votes short of the 60 needed to break a filibuster. The House had passed the bill, H.R. 9340, 417–3 on 16 September. Four Democrats, Maggie Hassan, Amy Klobuchar, Jon Ossoff and Raphael Warnock, joined Republicans. The rest of the caucus held.

The bill used existing federal utility law to make state commissions consider a standard under which data centers drawing more than 100 MW pay the full incremental cost of serving them. Consider, not adopt: commissions would have had one year to open proceedings and two to decide, and could decline. Senate Minority Leader Chuck Schumer called it toothless and said Democrats want data centers to cover their costs by statute, not by suggestion.

Nine days earlier, California did exactly that. On 21 September Governor Gavin Newsom signed seven data center bills, three of them on electricity: SB 886, AB 2383 and SB 1168. Together they direct the state utility regulator, the CPUC, to build separate tariffs and rate structures under which data centers pay for the grid upgrades and new supply they trigger, while blocking cost shifts onto low-income customers.

Washington declined to issue even a suggestion. Sacramento issued a mandate. Who pays for AI's grid is now being decided one statehouse at a time.

II.Who pays has become a siting variable

The patchwork already exists, though nobody agrees on its size. The House Energy and Commerce Committee said 24 states already have similar standards. Environment+Energy Leader counted 23. Tech Times reported that only four, Virginia, Oregon, Ohio and Oklahoma, had standards close enough to be exempt from the bill's proceedings. The distance between 24 and four is the distance between having a large-load tariff and having one that bites.

California's version is the strictest on paper. According to a Mayer Brown summary, SB 886 requires new CPUC tariffs by 1 January 2028, with data centers paying infrastructure and generation costs and facing early termination fees if they leave within ten years. AB 2383 caps the minimum-demand threshold at 25 MW and requires ten-year minimum payments backed by collateral.

Sources conflict on how much of that is actually written down. An mgrid analysis published on signing day found no megawatt trigger, no dollar threshold and no filing deadline in the package. We could not reconcile the two from the summaries available. The CPUC proceeding will settle it, and until it does, the operative numbers do not exist.

The stakes are large. Data centers have requested 18.7 GW from California utilities, according to Senator Jerry McNerney's office, against expected data center load growth of 2.3 GW by 2030. The gap between requests and expected builds is the speculative queue that ten-year payment obligations are designed to shrink.

III.The biggest cost sits where state tariffs cannot reach

Both the federal bill and the California package mainly target one cost channel: the wires, substations and transformers built to connect a specific facility. That cost is the easiest to assign because it has a single cause.

The larger number sits in wholesale markets. PJM's Independent Market Monitor estimates that data centers accounted for $6.3bn of the $16.4bn cost of the July 2026 capacity auction, and $29.4bn, or 46%, of $63.6bn across the last four. When new load bids up the capacity price, every customer in the region pays the higher clearing price. A state tariff that bills a data center for its own interconnection does not touch that. The monitor's position is that only a separate auction for large loads would.

The market is tight. PJM's 2028/29 auction cleared at $325/MW-day and 6,831 MW short of its reliability requirement.

The skeptic case is strong and deserves a fair statement. A study backed by EPRI, the utility industry's research institute, found that data centers lowered residential prices between 2015 and 2024: doubling data center capacity cut them by about 3.5%, because fixed grid costs were spread over more kilowatt-hours. A May 2026 E3 report, commissioned by the Data Center Coalition, found no evidence of historical cross-subsidy and put only about half of PJM's 2024/25-to-2025/26 capacity price rise on load growth, the rest on market design and plant retirements. California's own price surge was driven mostly by wildfire costs.

The newer data cuts the other way. In February 2026, average retail prices rose 26.3% year on year in Virginia, 21.9% in Ohio and 19.5% in Pennsylvania, all PJM states, against 9.0% nationally, per EIA data. The backward-looking studies end in 2024; the capacity shock began reaching bills in mid-2025. Both findings can hold: data centers spread fixed costs in a slack system and bid up scarce capacity in a tight one. That reconciliation is our reading, not a conclusion of either study.

IV.The economic ripple

Household prices. A rough sizing: $29.4bn across four delivery years is about $7.4bn a year. Spread across the roughly 67 million people NPR counts in PJM's footprint, that is around $110 per person per year, before costs are divided between homes, businesses and industry. This is our arithmetic and assumes each auction maps to one delivery year. It sits in the same range as the Maryland Office of the People's Counsel estimate, also cited by NPR, of $168–216 a year in extra costs per customer.

Capital and location. Ten-year minimum payments, collateral and exit fees turn a cheap option on grid capacity into a balance-sheet liability. That raises the cost of speculative interconnection requests most, and of firm projects least. If siting responds, marginal load moves toward states with looser tariffs and spare capacity. Virginia and Texas already lead the roughly 5,400 US data centers; California, with 296, ranks third.

Market structure. The largest buyers already agreed to pay. On 4 March, Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI signed the Ratepayer Protection Pledge, committing to cover delivery upgrades and accept separate rate structures. For them, California codifies a promise. The firms hit hardest are the ones that did not sign: colocation developers and site promoters without investment-grade balance sheets. Cost allocation rules act as a fixed cost, and fixed costs favour scale.

Labour. The headline jobs figure states will weigh, 665,500 California jobs in 2024 per the Data Center Coalition as cited by mgrid, is an industry number. Most of the real labour effect is construction and grid work, which follows wherever the tariffs push the load.

V.Second-order: a cleaner queue and a redrawn map

The least discussed consequence may be the most useful. Capacity auctions price forecast load. If ten-year payment obligations purge speculative requests, forecasts fall toward what actually gets built, and clearing prices could ease for everyone. California's 18.7 GW of requests against 2.3 GW of expected growth shows how much air may be in those numbers.

The second is regulatory arbitrage. A patchwork of 24 tariffs with different thresholds lets developers shop. States that compete on cheap, unallocated grid access win buildings and inherit the capacity costs that come with them.

The third is political. The Senate left for the election recess with the issue open, and 53% of Americans told a September poll they were extremely or very concerned about data center effects on power prices and water. Each party can now campaign on the vote.

VI.Precedent: a pledge is not a tariff

Voluntary commitments have now been tried twice. The March pledge promised what the federal bill would have asked states to consider; Roll Call notes the bill codified parts of it. Four months after the pledge, PJM's market monitor said hyperscalers could not honour it under current market rules without a separate auction.

California chose statute instead. Its laws still depend on a CPUC proceeding that has not set its numbers. The precedent so far is that commitments not written into tariffs do not change a single bill.

VII.How it plays out: three scenarios

Probabilities are our analytical judgment, not forecasts from a model, and they sum to 100%.