Electrical transmission and substation infrastructure on the power grid

AI Data Centers Need More Power. The Harder Question Is Who Pays for It

작성자:

한국어판: AI 데이터센터에 전력이 더 필요해질수록, 더 어려운 질문은 ‘누가 비용을 내는가’다

The AI power debate is usually framed as a race to build more generation.

That is only part of the problem.

For utilities and regulators, two questions are becoming just as important: how quickly can very large new loads connect to the grid, and who bears the financial risk of the infrastructure built to serve them?

Those questions matter because data-center electricity demand is growing fast enough to change grid planning. Lawrence Berkeley National Laboratory’s latest data-center work estimates that U.S. data centers could account for roughly 9.5% to 15.3% of total electricity use by 2030, with a central estimate of 11.8%.

But the most useful insight is not the headline percentage. It is that a data center can create a local capacity problem long before the national power system runs out of electricity.

The bottleneck is not only how many megawatts exist

A large data center needs power at a specific place, on a specific schedule, with a high degree of reliability.

That makes “speed to power” a different problem from simply adding annual electricity generation somewhere on the system.

In a June 2026 report, Berkeley Lab organized large-load connection bottlenecks into five areas: load forecasting, interconnection, resource planning and procurement, markets and operations, and cost allocation and ratemaking.

That last category is easy to overlook. If utilities build substations, transmission upgrades, or new supply for a large customer, someone has to pay for those investments. The risk becomes especially visible if a proposed project is delayed, downsized, moves elsewhere, or never materializes.

In June 2026, the Federal Energy Regulatory Commission opened a series of large-load proceedings covering six regional grid operators. Commissioner David Rosner described “Cost Recovery Agreements” as a way to make large loads bear their fair share of infrastructure costs even if a project does not come online as planned, rather than leaving residential customers exposed to the stranded cost.

That does not settle every jurisdictional or rate-design question. Retail cost allocation remains heavily shaped by state regulators. But it makes one point clear: the AI electricity problem is also a risk-allocation problem.

A proposal to treat homes as part of the energy infrastructure

Ari Matusiak, CEO of Rewiring America, pushes that logic further in a 2026 TED talk.

His proposal, which he calls “Homegrown Energy,” is to redirect some of the capital associated with serving new data centers into energy upgrades for homes in the surrounding region: heat pumps, batteries, rooftop solar, efficiency measures, and virtual power plants that coordinate many small resources.

The idea is not that a house becomes a miniature baseload power plant. It is that reducing peak demand, shifting consumption, and adding distributed generation can sometimes create usable grid headroom faster than a conventional supply project can be built.

Matusiak illustrates the concept with a 200-megawatt data center and roughly 50,000 homes. In his model, upgrades across those homes could create a similar amount of grid capacity. That figure should be read as the speaker’s policy-model estimate, not as an independently established engineering equivalence.

This distinction matters. A megawatt of avoided peak demand, a megawatt of battery discharge, and a megawatt of firm generation do not provide the same service in every hour or at every location.

The strongest objection: “available capacity” is not a single number

The most serious criticism of the Homegrown Energy argument is not that household efficiency or distributed resources are useless. It is that national or regional potential can be mistaken for locally deliverable capacity.

A critical analysis in The Breakthrough Journal argues that the proposal depends on very ambitious deployment rates, can understate geographic mismatches between homes and data-center load, and may overlook distribution-grid upgrades needed as households electrify more equipment. It also notes that some home upgrades add new electric load even while reducing fossil-fuel consumption.

Those objections do not invalidate distributed energy resources. They narrow the claim that can responsibly be made about them.

Rewiring America’s own later response adopts a more careful framing: household upgrades should complement, not replace, new generation, transmission, and grid modernization.

That is a more defensible version of the idea.

The real opportunity is a different way to procure capacity

If both sides are taken seriously, the debate becomes more useful.

The choice is not “build power plants” versus “upgrade homes.” Large data centers may ultimately need a portfolio that combines new generation, transmission, substations, storage, demand flexibility, grid-enhancing technologies, and distributed resources.

The important question is how those resources should be compared.

A simple megawatt count is not enough. A serious comparison needs to ask:

  • Location: does the resource relieve the part of the grid where the constraint exists?
  • Timing: can it be deployed before the new load needs service?
  • Firmness: how reliably is the capacity available when required?
  • Duration: for how long can it sustain the needed reduction or output?
  • Coincidence: does it help during the actual system or local peak?
  • Deliverability: can the grid physically move the benefit to the constrained area?
  • Cost responsibility: who pays if the infrastructure is built but the expected load never appears?

Seen this way, Matusiak’s proposal is most valuable not as proof that homes can replace power plants, but as a challenge to a narrower assumption: that the only way to serve fast-growing AI load is to build new centralized supply and then socialize the surrounding grid costs through conventional rate structures.

The question AI companies cannot avoid

The public debate will keep asking how many gigawatts AI requires. That number matters, but it is not sufficient for planning.

The more operational questions are:

Which resources can create usable capacity where it is needed, when it is needed—and who should pay for each one?

Those questions connect engineering, finance, regulation, and local politics. They also explain why the next phase of the AI infrastructure race will not be decided only by companies that can buy chips or sign power contracts.

It will also depend on whether grids can connect large loads quickly without transferring disproportionate costs and risks to everyone already on the system.

Sources