Power, Water, and Patience: Planning Shared Infrastructure for the AI Data Center Wave

For roughly twenty years, electricity demand in the United States was flat enough that grid planners could treat load growth as a rounding error. That era has ended. The driving force is not population or manufacturing alone, but the concentrated, fast-moving demand of AI-scale computing, where a single campus can request several hundred megawatts and expect to be energized within a development cycle rather than a generation-planning horizon.

The result is an unusual situation for enterprise developers: capital is abundant, sites are findable, and the binding constraint is often a substation queue position or a water rights agreement. Infrastructure planning, not real estate, now sets the timetable.

Demand Growth Arrives All at Once

The defining characteristic of AI load is its shape. Traditional industrial demand ramps gradually and can be forecast with reasonable confidence over five years. AI capacity arrives in blocks, driven by model training cycles and commercial commitments that can compress years of expected growth into eighteen months.

That mismatch has consequences. Utilities that planned transmission upgrades against modest forecasts are now revising load projections upward across entire regions, and interconnection queues that were already slow have become the practical gatekeeper for site selection. In several markets, the decisive question in a site search is no longer land cost or labor availability but how many megawatts can be delivered and by what date.

Three Constraints, One Schedule

Interconnection. Large-load service requests can require new substations, transmission reinforcement, and in some cases network upgrades spanning multiple jurisdictions. Timelines measured in years are common, and the sequence is not fully within the utility's control.

Water. Evaporative cooling remains widespread because it is energy-efficient, but it consumes water in places where water is contested. Municipalities that welcomed data centers a decade ago are now scrutinizing consumption, particularly in drought-exposed basins and in communities where residential growth is also constrained.

Land and neighbors. A campus needs acreage, buffers, fiber routes, and heavy-truck access. It also needs a local government willing to absorb the political cost of industrial-scale development near residential areas. Opposition is increasingly organized and increasingly focused on rate impacts rather than aesthetics.

Regional Pressure Points

Northern Virginia remains the deepest market, and its constraint is transmission and local approval rather than demand. Texas continues to attract capacity through a comparatively fast interconnection process, though the scale of new load is beginning to shape reliability discussions. Growth corridors in Georgia, Arizona, Ohio, and Indiana have drawn projects at a pace that local water and distribution systems were not designed for. In the Midwest, large-load tariffs and capacity obligations have become the central negotiation.

The common thread is that the easiest sites were taken first. What remains requires more engineering, more negotiation, and more patience.

The Queue as a Site Selector

For enterprise developers, the practical implication is that site selection should begin with energy, not acreage. A useful sequence:

1. Screen regions by available capacity and published queue timelines, not by land listings.
2. Engage the utility before the option period, with a defined load profile and a realistic ramp.
3. Identify whether the project can accept phased energization — partial capacity now, full capacity later — which often converts an unviable site into a viable one.
4. Model the cost of network upgrades explicitly. Those costs may land on the developer, the utility's other ratepayers, or both, and the allocation is negotiable.
5. Evaluate behind-the-meter generation and long-duration storage as schedule insurance rather than as permanent strategy.

Designing for Shared Capacity

The most durable projects are those that treat infrastructure as shared rather than dedicated. That can mean co-locating with other large loads to justify a substation upgrade, contracting for water reuse and closed-loop cooling where the technology fits the climate, or agreeing to interruptible service in exchange for faster interconnection. It can also mean funding a distribution upgrade that serves the surrounding community, which converts a local political problem into a public benefit.

Municipalities have an interest in the same design. A data center that consumes a region's remaining water allocation and contributes little to the tax base after abatements expire is a poor bargain. Performance standards — water use effectiveness thresholds, waste heat recovery requirements, phased capacity commitments, and decommissioning plans — give local governments a way to accept the load without accepting unbounded risk.

The Stranded Asset Question

Every capacity decision carries an assumption about how long the demand will last. AI hardware refreshes quickly, model efficiency improves, and the commercial landscape is volatile. A campus designed for a single tenant's peak requirement may sit underutilized if that requirement changes.

The mitigation is modularity. Design in blocks that can be energized, cooled, and leased independently. Avoid single-tenant dependencies where the market permits. Contract for capacity in increments that match build phases. And where possible, retain the flexibility to serve non-AI loads, which broadens the pool of future tenants and reduces exposure to a single demand curve.

What Developers Should Do Now

• Treat utility engagement as a pre-acquisition activity with a named executive owner.
• Require written capacity commitments and milestone dates, not expressions of interest.
• Price the upgrade cost of interconnection into the pro forma at the outset.
• Select cooling strategies against the local water regime, not against a national default.
• Build community and municipal relationships early, with real information about rate and water impacts.
• Structure the asset so that partial capacity, multiple tenants, and phased expansion remain possible.

The AI buildout is not a temporary surge, but it will not remain in its current form. The infrastructure decisions made in 2026 will determine which of these projects are still economically sound when the demand curve flattens — and which become monuments to a forecast that arrived faster than the grid could follow.