Most of this week’s AI governance news has pointed upward: the UK sharpening its regulatory posture, the UN pressing for a coordinated international framework, everyone reaching for rules that sit above the technology and tell it how to behave. A Foreign Affairs essay by Sarah Kreps, a technology policy scholar at Cornell, argues that this is governing the wrong layer. The question that will actually decide whether a democracy leads in AI is not what rules constrain the models. It is whether the country can build the physical infrastructure the models run on, in real places, with the consent of the people who live there.
The rules debate governs behaviour; the harder problem is whether you can build at all
Kreps’s argument starts from an inconvenient fact. AI is not an abstraction that lives in the cloud. It is data centres, transmission lines, and energy-intensive computing facilities that have to be sited, permitted, and sustained in communities whose priorities do not always match national ambitions. Those facilities have become strategic assets. But in a decentralised political system, no national strategy can compel a local planning authority to approve one.
She draws the contrast sharply. In China, large-scale digital infrastructure is planned, permitted, and constructed through nationally coordinated processes that subordinate local opposition to state priorities. That model carries real costs, concentrating burdens on communities with no way to contest projects or hold planners accountable. But it confers a strategic advantage that Kreps does not wave away: speed, scale, and predictability. The democratic alternative cannot copy that machinery, and should not want to. What it needs instead is an institutional model that connects national ambition to local legitimacy.
Strategic Reality: The live governance frontier is not how AI systems behave. It is whether the country can physically build the AI economy at the pace competitors set. Rules govern the software layer; institutions of consent govern whether the concrete ever gets poured.
This is the pivot that makes the essay a useful corrective to the week’s headlines. Top-down rules are necessary, but they operate on a layer where a directive can, in principle, be obeyed. Physical build sits on a layer where federal or national authority runs out. Kreps puts the numbers behind the stakes: by some estimates, data centres could consume roughly eight per cent of United States electricity by 2030. Yet no executive order, task force, or incentive programme can override a local zoning board. That gap between national priority and local veto is where AI leadership is quietly won or lost.
| The physical reality Kreps documents | Figure |
|---|---|
| Projected US data centre share of national electricity by 2030 | ~8% |
| Permanent jobs at the proposed Cayuga (Lansing, NY) data centre campus | ~75 specialised roles |
| Data centres built in northern Virginia in the past decade | 250+, hundreds more proposed |
| Google’s statewide Texas cloud and AI investment | $40 billion |
What the fights are actually about
Kreps writes not as a detached analyst but as a resident of Lansing, New York, where an energy and digital infrastructure company is seeking to develop a large data centre campus on the site of a retired coal plant. She conducted an independent assessment of the project’s likely effects on water, power, employment, and the local tax base, disclosing that the company compensated her for the time but had no control over the scope, findings, or final report. Her technical conclusion was that the engineering was sound: a closed-loop cooling system would largely eliminate operational water withdrawals, and the site already sat on a mostly carbon-free grid with surplus capacity.
None of that settled the argument, and this is the essay’s most important observation. Much of the opposition ran on misinformation, conflating a specific project with broad anxieties about AI, corporate power, and surveillance. But the sceptics were not wrong about everything. Permanent employment would likely settle at around 75 specialised roles, meaningful for a town absorbing industrial closures but no jobs engine, and the tax benefit was muddied by corporate exemptions and incentives.
Critical Context: The real disagreement was not about gallons of water or megawatts of power. It was about control: whether decisions affecting the lake, the grid, and the land would be shaped by local institutions or by national-scale companies. The technical details mattered only insofar as they signalled who was trusted to decide.
The same fault line runs across very different places. In Chandler, Arizona, a desert city already facing chronic drought, the council rejected a large data centre complex backed by major technology firms over water and heat concerns. In northern Virginia, home to the world’s largest concentration of data centres, resistance has become a defining local political issue, shaping county elections and mobilising residents over noise, power demand, and water use. What these communities are rejecting, Kreps argues, is rarely computing itself. It is the perceived asymmetry between the residents asked to host the infrastructure and the companies whose decision-making looks opaque and elite-driven.
For a UK reader this should feel familiar rather than foreign. Britain is standing up AI Growth Zones and courting hyperscale data centre investment whilst its planning system remains the slowest link in the chain. The specific grievances differ, but the underlying tension is identical: national strategy needs local land, power, and consent that no strategy document can command.
The stakeholder map beneath the noise
Reading the disputes as simple NIMBYism misses the structure. Each group is optimising for something legitimate, and the conflict comes from the fact that the current process forces those interests into a single, one-off moment of confrontation at the planning hearing.
| Stakeholder | What they are actually optimising for | Where the current process fails them |
|---|---|---|
| National policymakers | Speed, scale, strategic AI capacity | No lever to convert national priority into local approval |
| Local government | Tax base, services, environmental stewardship | Asked to carry risk with thin, contested information |
| Developers | Predictable timelines, operational stability | Treat consent as a one-time hurdle, then face lasting resistance |
| Residents | Control over shared resources, a share of the gains | Consulted once, rarely given standing to hold commitments |
| Independent intermediaries | Credible translation between technical and civic | Rarely funded or built into the process at all |
Competitive Reality: A project rejected at the zoning board is not slower than one built by state fiat. It is a strategic loss. Every campus that dies in litigation or hardens into permanent opposition is capacity a more centralised rival builds without asking.
Building with buy-in: the model Kreps proposes
The essay’s constructive core is that none of the fix requires new federal authority. It requires coordination, shared standards, and credible commitment. Kreps sets out three elements, and they translate cleanly into a build-side operating discipline for any developer or authority working in a democracy.
Standardised transparency, independently reviewed. Before permits are granted, companies should publish verifiable figures on water use, power demand, emissions, and expected tax contribution, with independent analysts reviewing them rather than letting corporate marketing stand unchecked. Kreps frames this as a middle path: more substantive than the minimal oversight of the past, far less onerous than the layered reviews that can make building anything in parts of the United States prohibitive. Transparency does not end disagreement, but it grounds debate in reviewed data instead of worst-case speculation.
A clear mechanism for local benefit. Even a small slice of a multi-billion capital investment can fund technical training at nearby colleges, expand broadband, support energy-efficiency upgrades, or reuse a data centre’s waste heat. The point is not compensation but integration: weaving the infrastructure into the economic life of the community that hosts it.
Standing liaison bodies. Firms should establish permanent forums of residents, officials, and company representatives that monitor compliance, address grievances, and adjust commitments as conditions change. Unlike a one-off hearing, an ongoing forum institutionalises dialogue and stops minor disputes escalating into blanket opposition.
Success Factor: The through-line is that permitting stops being a single moment of consent and becomes a sustained relationship. The firms already treating it that way are not being generous. They are buying predictability.
Kreps’s evidence for the pay-off is concrete. Google paired a Texas expansion with a solar and battery facility, education grants, and a $30 million community fund for workforce training and energy affordability. Microsoft built a technician pipeline with Iowa community colleges. In Umatilla, Oregon, years of Amazon data centre taxes and sponsorships funded robotics and STEM labs until residents came to associate digital infrastructure with reinvestment rather than extraction. These remain exceptions, not the norm, but they are the template that clears sites whilst others stall.
Four ways the model breaks in practice
The framework is sound, but anyone applying it should treat these failure modes as first-order risks, not footnotes.
Transparency can arm the opposition it was meant to reassure. Publishing the real jobs figure, as at Cayuga, hands sceptics an honest number to campaign on. Mitigation: pair disclosure with the benefit mechanism in the same announcement, so the community weighs 75 jobs against a funded package, not 75 jobs alone.
Community benefit reads as a bribe if it arrives late. A fund offered after opposition mobilises looks like hush money; the same fund designed into the project from the start reads as partnership. Mitigation: benefit has to be in the core design, not the crisis response.
Liaison bodies get captured or become theatre. A forum with no teeth is a public-relations exercise that deepens the mistrust it was meant to cure. Mitigation: give it real standing over monitoring and grievance, and independent membership, or do not bother.
Warning ⚠️: The most dangerous misuse of this essay is to invert it. The China-speed argument is not a case for bypassing consent when it gets slow. Kreps’s whole point is that durable consent is the faster route, because litigation and hardened opposition cost more time than engagement ever does.
The legitimacy advantage, and what it means for Britain
Kreps closes on an American precedent that travels. During the New Deal, the Tennessee Valley Authority combined federal investment with local participation, extending electricity and opportunity to rural communities whilst embedding projects in regional institutions. The lesson was not that big things can be built, but that shared benefit and accountability make them politically durable. The physical foundations for the next phase already exist in many places: former industrial sites, legacy transmission lines, surplus generation. What is missing is not engineering capability but the institutional capacity to secure consent for nationally significant projects.
That is the sentence UK policymakers and developers should sit with this week. Regulation of how models behave is real work and worth doing. But it governs a layer where the country still has options. The build layer is where Britain’s constraint actually bites, and no volume of top-down rules will connect a substation or win over a parish council. The governance that decides AI leadership is more prosaic and more local than the summit-level debate suggests.
Take Action: If your organisation is siting, backing, or hosting AI infrastructure, treat local legitimacy as a build requirement, not a communications afterthought. Three tests before you break ground: Have you published independently reviewed figures? Is a community benefit designed into the project, not bolted on? Is there a standing body with real teeth, not a single hearing? A project that fails all three is not cheaper. It is a lawsuit waiting to happen.
The debates playing out in New York, Arizona, and Virginia, Kreps writes, are a test of whether a decentralised democracy marked by declining trust can still coordinate complex, nationally significant projects. Britain is running the same test, on the same infrastructure, at the same time. The country that solves consent, not just rules, is the one that gets to build.
Source and attribution
This analysis draws on “A Better Way to Build AI” by Sarah Kreps, published in Foreign Affairs on 6 July 2026. The original essay makes the case in a United States context; the UK framing, the stakeholder analysis, and the failure-mode assessment are Resultsense’s own. Figures and cases cited are drawn from Kreps’s reporting.
Resultsense makes sense of AI in the UK. For businesses weighing where and how the AI economy gets built, the governance question is no longer only what the rules say. It is whether the institutions exist to turn strategy into infrastructure.