The people who pay are not the people who decided

Arizona's largest utility wants to raise electricity rates by 45% for data centres and 14.5% for households. The hearing began on Friday. The second number is the one nobody is defending, and it is the one that tells you how the cost of AI's physical build-out is actually being distributed.

Arizona Public Service (APS, the state's largest electric utility) filed its general rate case with the Arizona Corporation Commission (ACC, the five-member elected body that regulates utilities and sets rates) requesting differentiated increases by customer class. Data centres face roughly 45%. Residential customers face roughly 14.5%. The driver, in APS's own filing, is capital expenditure on grid infrastructure required to meet surging electricity demand, and the demand surge is overwhelmingly being driven by hyperscale data centres serving AI training and inference workloads.¹²

The Wall Street Journal called Phoenix "a data-center mecca — and test case for how to pay for AI's power needs."¹ That framing is correct, and it is bigger than Phoenix. No federal framework exists for who pays for AI's physical infrastructure. The venue where that question is being settled, right now, is a state utility commission in Arizona.

The differential is real, and it is not the story. I want to be fair to APS here, because the cleanest version of the contrarian reading deserves engagement. A 45% increase on data centres against a 14.5% increase on households is, by the standards of US utility ratemaking, a substantial attempt to charge the cost-causing class rather than socialise everything. Utilities frequently do worse. Arizona's Attorney General Kris Mayes called the proposal "blatant corporate greed,"³ and the political framing is sharper than the analytical one. The differential exists.

But the differential is not the question. The question is why the residential number is greater than zero.

Cost causation, in plain terms. If a customer class is responsible for the bulk of new load, and therefore the bulk of new grid investment, then the cost-of-service principle that underpins US utility ratemaking says that class should pay for the bulk of the new infrastructure. Cost causation (the regulatory principle that those who cause a cost should bear it) is not a radical idea. It is the textbook standard. Hyperscale data centres in the Phoenix metro are not a marginal new load; they are the load. Microsoft, Meta and Google have all announced significant Phoenix-area expansions in the last two years.³ The residential class did not change its consumption pattern. The grid did not need rebuilding for households.

That is the gap the headline 45% obscures. A genuinely cost-causation-faithful tariff would charge data centres something north of 45% and would charge residential customers very little. The fact that residential customers are being asked for 14.5% means a real share of hyperscaler grid build-out is still being socialised across people who did not ask for the data centres, do not work in them, and will not, in any meaningful sense, be consulted on whether more get built.

Who decided, and who pays. The siting decisions that produced this load growth were made by three companies, in coordination with state economic development officials and the utility. Residential ratepayers in Maricopa County were not part of those conversations. They did not vote on the Microsoft campus. They did not vote on APS's resource plan. They are now the customer class being asked to absorb 14.5% of their electricity bill into the financing of an infrastructure expansion they had no role in authorising.

This is what the consequences question actually looks like once you strip the rhetoric off it. The benefits of the build-out, construction jobs, state tax revenue, the political win of landing a hyperscaler, flow through aggregates: state GDP, the governor's press releases, the economic development authority's annual report. The costs flow through line items on monthly utility bills paid by people whose median household income in Arizona is around $74,000. Aggregate gains, itemised costs. This is a familiar shape, and it is not new to AI. It is how a great deal of industrial-policy externalisation has always worked. AI is making it visible at a scale, and on a timeline, that previous waves did not.

State utility commissioners are now AI policymakers. The ACC has five elected commissioners. Their statutory job is to balance utility solvency against ratepayer protection. They were not designed, resourced, or trained to adjudicate the distributional incidence of a generational shift in compute infrastructure. They are doing it anyway, because no one else is. Virginia, Texas and Georgia commissions will be doing the same thing within eighteen months. The federal government has not produced a framework for who pays for AI's grid, and absent one, the question defaults to fifty separate state proceedings, decided on the procedural rhythms of utility ratemaking — evidentiary hearings, intervenor testimony, final orders late in the year.

I do not think this is where a structural question about who finances AI's physical layer should be settled. I also do not see another venue currently doing the work. The ACC's final order, expected late 2026, will be cited in every subsequent state data-centre tariff case in the country. That is a large amount of precedent-setting weight to put on a state utility commission.

What to watch. Three things. First: whether the ACC's final order moves the residential figure closer to zero, or leaves a meaningful share of the build-out socialised. Second: whether Mayes's intervention generalises — whether AGs in other data-centre-heavy states pick up the same frame, and whether it becomes a coalition position rather than one elected Democrat's filing. Third: whether the hyperscalers themselves respond by signing long-term direct-supply contracts that take their load off the regulated tariff entirely, which would resolve the cost-allocation question by removing data centres from the shared grid as a ratepayer class. Each of these moves the distributional answer in a different direction.

The 45% number will lead the coverage. The 14.5% number is the one I would keep my eye on. It is the measure, in dollars on a household bill, of how much of AI's infrastructure cost the people building AI have so far succeeded in not paying.

Footnotes

¶