For years, the “cloud” sounded weightless, as if data floated above us and the digital economy existed everywhere and nowhere at once.
But the cloud was never in the sky. It has always been on land, plugged into the grid, cooled by water and air, and built in communities.
That reality has become impossible to ignore. The International Energy Agency projects that global electricity use by data centers will roughly double by 2030. In the United States, the Electric Power Research Institute estimates that data centers could account for 9 to 17 percent of electricity demand by 2030, up from roughly 4 to 5 percent in 2024.
Those projections are not an argument against AI. They are a reminder that AI is physical.
And once a technology becomes physical, it has to contend with the world as it is: land, power, water, permitting, cost, and public trust.
For AI, the infrastructure challenge is not simply securing enough electricity. It is earning and maintaining permission to build: by engaging communities early, allocating costs fairly, and delivering visible local benefits.
I have spent much of my career in nuclear energy, and one thing I have learned is that the hardest problems are often outside the reactor. At Idaho National Laboratory, I helped launch the National Reactor Innovation Center, whose job was to help companies move from promising reactor designs on paper to real demonstrations in the world.
That meant more than engineering. That meant navigating environmental review and working with local officials, state agencies, and communities that wanted to understand not only what might be built near them, but what it would mean for them.
One moment stayed with me. We received a Freedom of Information Act request from a county commission seeking information about projects that could be built in its region. We provided the documents, but did not treat the request as a records exercise alone. We reached out and offered to meet. The conversation that followed was constructive and necessary, and the interaction shifted from a tense transaction towards a collegial relationship.
The lesson was simple: transparency cannot be limited to posting documents after people ask for them. It has to include early outreach, plain explanations, and a willingness to sit down with the communities that will live with the project. AI infrastructure developers should learn that lesson now.
A data center may serve a global digital economy, but it lands in a specific town. Its benefits may reach far beyond the host community, while its most immediate impacts (construction traffic, land and water use, power demand, new substations and transmission lines) are local.
Nuclear energy has lived through this tension for decades. It began with extraordinary promise: a tiny atom producing enormous amounts of energy, with hopes of abundance and national strength. Much of that promise was real. Nuclear power still provides nearly one-fifth of America’s electricity and more than 40 percent of its carbon-free electricity.
But getting the physics right was not enough.
U.S. nuclear construction costs and schedules were already worsening before the Three Mile Island nuclear accident, amid changing designs, regulatory shifts, economic pressures, and growing public opposition. The 1979 accident intensified public fear and distrust. It also exposed real weaknesses and prompted legitimate improvements in regulation and industry practice that materially improved safety, while adding disruption for plants already under construction. One study found that reactors under construction at the time of the accident and eventually completed afterward had median overnight construction costs 2.8 times higher and median construction durations 2.2 times longer than reactors licensed before the accident.1 The study identified licensing delays, regulatory uncertainty, and backfit requirements as significant contributors, but did not attribute the increase to any single cause. The prolonged absence of a sustained U.S. new-build program also left domestic supply chains and construction capabilities in need of rebuilding. More broadly, nuclear’s experience showed how easily trust can fray when people feel excluded from decisions or believe they are being asked to bear risks without a fair share of the benefits.
Yet the aftermath of Three Mile Island also produced a constructive institutional response. The industry created the Institute of Nuclear Power Operations (INPO), recognizing that a failure at one plant could damage confidence in every plant. Through shared standards, peer evaluations, and systematic learning from operating experience, INPO created accountability beyond regulatory compliance. AI companies face a similar collective challenge. They need not wait for a crisis to build credible systems of industry-wide accountability that complement protective, efficient public regulation.
For commercial nuclear projects, NRC licensing provides multiple avenues for public participation, including public meetings, environmental-review comment periods, access to project records, and hearing opportunities. Those formal avenues matter, but they are not a substitute for the broader engagement that earns trust.
The lesson from nuclear is not that we should be less ambitious. It is that engineering success is not enough. Ambition works only when institutions can execute, and people can trust both the process and the allocation of risks, costs, and benefits.
The obligation runs both ways. The nuclear industry has learned a great deal about meaningful community engagement, but maintaining trust requires nuclear developers to keep applying and
strengthening those practices as new projects move forward. Nuclear developers must also be candid about cost, schedule, risk, and local impacts. They must explain who will bear those risks and costs and deliver the benefits they promise.
That shared responsibility matters most where nuclear generation and data centers are developed or expanded together. Their sponsors should coordinate their engagement so communities receive one clear account of what will be built, who will pay for it, how impacts will be managed, and what benefits will remain locally. Communities should not have to sort out conflicting claims or determine which developer is responsible for which commitment.
An emerging shorthand about AI and energy is as follows: AI needs enormous amounts of power; nuclear can provide it; problem solved.
There is truth in that story. Nuclear energy can play an important role in powering the AI economy. Large technology companies need firm, clean power, and nuclear energy is having its most promising moment in decades.
Vogtle Units 3 and 4 in Georgia show both the promise and the difficulty of nuclear’s comeback: they are now producing carbon-free electricity, but only after years of delay and major cost escalation. The lesson is not to give up, but to be clear-eyed about everything required to build.
Choosing the right energy source does not, by itself, solve the broader infrastructure problem. The generation, substations, transmission, water systems, and data-center campus still have to be financed, permitted, constructed, and accepted somewhere.
Technology companies and data-center developers increasingly recognize that access to power is core to their growth. But securing megawatts is not the same as building durable infrastructure. They should bring the same strategic attention to siting, cost allocation, and community engagement that they now bring to energy procurement.
That means engaging early, explaining what a project means for the community as well as the country, and publishing credible estimates of power demand, water use, transmission needs, construction impacts, tax revenues, and permanent jobs.
Developers and utilities should answer early and plainly: Who pays for the new generation, substations, and transmission needed to serve these loads? Those costs should not quietly fall on households and small businesses. Projects also need visible, durable local benefits, not just promises of national competitiveness.
Policy is beginning to catch up. In March 2026, seven leading technology companies signed a White House pledge to secure the new power their data centers require, pay for related grid upgrades, and negotiate separate rates that they would pay whether or not the projected electricity is ultimately used. That is an important statement of principle, but a voluntary pledge is not the same as ratepayer protection. The more consequential work is happening in utility commissions. In 2025, Ohio regulators approved an AEP Ohio tariff requiring large new data centers in the utility’s service territory to support their electricity requests with long-term minimum payments, financial assurances and exit fees.
Speed and public legitimacy are not opposing goals. Early engagement and fair cost allocation are part of building quickly, and building in a way that lasts.
The United States needs AI infrastructure. It needs clean, reliable energy. It needs advanced manufacturing, electrification, climate resilience, and national security. But none of those ambitions will be achieved by software alone.
The future is not actually in the cloud. It is on land, in communities, in steel and concrete, and in the trust people place in the institutions asking to build near them. The real test for AI is whether it can be built responsibly in that physical world.
1 Jessica R. Lovering, Arthur Yip and Ted Nordhaus, “Historical Construction Costs of Global Nuclear Power Reactors,” Energy Policy 91 (2016): 371–382.