Who Can Block a Data Center

Executives usually discuss AI infrastructure in terms of models, vendors and chips. Those details matter, but they rest on an assumption that is getting harder to defend: that the physical side is simply there to be bought.

AI is turning data centers into visible industrial infrastructure. A site needs land, a substation, transmission, cooling, construction crews, long contracts and some measure of local political support. Capacity arrives as a request for power and space, and requests can be refused.

So a company that treats AI capacity as a procurement line may find it also depends on approvals. A utility commission, town council, governor, grid operator, water district, lender or group of nearby residents can hold up a project when they see no local benefit in it.

Compute now depends on permission from people outside the company, and the C-suite should manage it with that dependency in view.

Demand is rising faster than the approvals behind it

Chips are no longer the only scarce input. Projects now also wait on power, permits and finance, and the demand figures show how much load those approvals have to cover.

International Energy Agency estimates put data-center electricity use at about 415 terawatt-hours in 2024, roughly 1.5 percent of global consumption, with a base-case path toward roughly double by 2030. A more recent agency summary puts 2025 use near 485 terawatt-hours and 2030 demand near 950 terawatt-hours. AI-focused data centers are growing much faster than the total.

In the United States the load is more concentrated. A U.S. Department of Energy (DOE) Lawrence Berkeley National Laboratory report found that U.S. data centers consumed about 4.4 percent of total U.S. electricity in 2023 and could reach 6.7 to 12 percent by 2028. The Electric Power Research Institute (EPRI) 2026 update goes higher for 2030, to 9 to 17 percent of U.S. electricity, with the spread depending on how many projects get through supply-chain and permitting hurdles.

The capital numbers are on the same scale. McKinsey estimates that global data-center demand could almost triple between 2025 and 2030, from about 82 gigawatts to about 220 gigawatts. Goldman Sachs Research expects private-market financing to matter more because hyperscalers may spend roughly $5.3 trillion on AI and data centers by 2030. Another Goldman note put consensus for 2026 hyperscaler AI capex at $527 billion and warned that supply bottlenecks and investor appetite may limit spending before balance sheets do.

Governments have started to respond. The White House's 2026 Ratepayer Protection Pledge shows data-center growth has become a household-bill issue: it asks major AI and data-center companies to build, bring or buy new power supply, pay for power-delivery upgrades, and negotiate separate rate structures so that ordinary households do not end up subsidizing the load. An Associated Press report this week described legislation aimed at forcing AI data centers to get their electricity and water from private sources, and more than 20 Florida counties and municipalities had rejected, delayed, or paused major data-center projects for at least a year. Local planning decisions like these are usually the first place a company feels the constraint.

Where an AI roadmap can be stopped

Boards can leave the utility engineering to others, but they should know which parties can halt the roadmap, and at what stage.

Why this belongs in corporate strategy

The easy mistake is to label this a sustainability issue and pass it to the environmental, social, and governance team. That frame is too small. It is more useful to think in terms of complementary assets.

AI advantage depends on several of them at once: data rights, workflow redesign, risk controls, distribution, talent, and now physical capacity. When one is scarce, the value of the rest shifts. A better model buys little if inference is too expensive to run at volume, and an ambitious agent roadmap stalls when the production capacity it needs sits behind a delayed substation. Even a strong vendor relationship leaves the exposure in place, because the vendor is competing for the same power, cooling equipment, construction labour and capital.

Once capacity is scarce, the possibility of a veto changes build-versus-buy decisions and where workloads run. It also changes which use cases get the capacity that exists and how the company explains its AI plans to regulators and to the communities near its sites.

The companies most at risk hold two beliefs that cannot both be true: that AI will soon touch core operations, and that the physical cost and permissions behind that AI can stay out of sight.

What buyers should ask vendors

Enterprise buyers do not need to audit every megawatt behind a vendor, but they should stop assuming capacity will appear when needed. Before moving a strategic workflow onto an AI platform, ask four questions.

Where is the capacity assumption? If the product roadmap depends on larger models, lower latency or heavier inference, the buyer should know whether the vendor can support that growth without abrupt changes to pricing, availability or service levels.

Who bears infrastructure cost volatility? If power, equipment and financing costs rise, the vendor will absorb the margin pressure, reprice, restrict usage or change architecture. The buyer should know which, in advance.

What is the fallback architecture? Critical workflows need a plan for degraded model quality, lower inference volume, regional capacity limits or a slower rollout. Resilience here covers uptime, and it also means the business process keeps running when compute is constrained.

Which claims rest on demand evidence rather than demos? Heavy capex can signal confidence, or it can signal an industry racing ahead of evidence that the spending will pay. Buyers should ask how much of a vendor's announced capacity is already carrying real workloads.

What boards should require

At board level the rule is simple: every significant AI program names its physical dependency. The dependency may be direct, when the company builds or leases dedicated infrastructure, or indirect, when it relies on a hyperscaler's buildout. In either case the operating plan should answer five questions.

1. Which workflows are important enough to deserve scarce, possibly expensive AI capacity?
2. Which vendors or infrastructure partners control the path to that capacity?
3. What power, cooling, permitting and tariff assumptions sit behind the roadmap?
4. What happens if capacity is delayed, rationed, repriced or politically constrained?
5. What evidence would lead the company to slow its AI buildout instead of spending through the constraint?

I would put most weight on the last one. Infrastructure enthusiasm tends to turn forecasts into commitments before anyone knows whether demand will hold. Commitments should be staged like real options, buying learning and keeping flexibility early, with heavy commitment reserved for cases where usage evidence and operating advantage are both strong.

A first step for this week

Take the three AI initiatives most likely to become operationally important and attach a one-page physical-dependency appendix to each business case. It should record the expected compute pattern, the vendor's capacity assumption, exposure to cost volatility, regional constraints, the fallback mode, and the person responsible for revisiting those assumptions every quarter.

For small pilots this will feel premature, but substations, cooling systems, construction schedules, utility tariffs and voters now sit under every AI roadmap, pilots included. An initiative too small to justify an infrastructure assumption is probably too small to carry a transformation story, and one big enough to matter is exposed to a veto from a utility commission, a water district or a town council.