The Invisible Infrastructure of AI: Data Centers, Energy, and Water

TL;DR — Data Centers, Energy, and Water

• AI is breaking two decades of flat U.S. electricity demand. Data center power consumption is set to nearly double by 2030 (~945 TWh), growing 4x faster than overall electricity demand.

• The largest new AI campuses draw up to 20x the power of a typical data center, landing on specific grid nodes faster than transmission and permitting (often 5–10+ years) can keep up — already showing up as spiking capacity market prices (PJM up ~10x in one year).

• Cooling forces a tradeoff, not a fix: air cooling uses more electricity, water-based evaporative cooling uses less power but consumes billions of gallons (U.S. data centers: 17.4B gallons in 2023, projected 38–73B by 2028).

• Indirect water use (from power generation) is over 10x direct cooling water use — so facility-level "we solved water" claims (e.g., Nvidia's new cooling system) usually only address a fraction of the real footprint.

• Site selection is now a resource-scarcity decision, not just tax/latency — the cheapest power is often in the driest places, and the wettest places often lack spare grid capacity.

• Latin America shows the same dynamics at earlier scale: 42%/year investment growth, but installed capacity still under a third of Northern Virginia's — and Peru/Ecuador already see hydro-driven blackouts hitting industrial users.

• Bottom line: the constraint on AI's next phase may not be model capability — it may be gigawatts and gallons.

Every AI prompt has a supply chain. It runs backward from the chat window, through a server rack, into a cooling loop, out to a power plant, and often down into an aquifer. Most of that chain is invisible to the person typing the question. It is not invisible to the utility engineers, water regulators, and capital providers who now have to plan around it.

The last two decades of U.S. electricity demand were remarkably boring: nearly flat, growing at well under 1% a year. That era is over. Corporate investment in AI, and the data centers built to supply it, has become one of the fastest-growing categories of grid-connected load in modern power systems, and the resource story behind that growth runs through two commodities that don't scale as easily as capital does: energy and water.

The demand curve is different this time

Global electricity consumption from data centers is projected to roughly double by 2030, reaching close to 945 terawatt-hours, growing at around 15% a year — more than four times faster than global electricity demand as a whole. In 2024, data centers already accounted for roughly 415 TWh, about 1.5% of world electricity use, a figure that has been compounding at 12% annually since 2017.

The scale problem is as much about concentration as volume. A typical data center draws roughly as much power as 100,000 households. The largest next-generation AI campuses now under construction will demand 20 times that amount, landing on specific grid nodes rather than spreading evenly across a network built for gradual, diffuse growth. Some industry projections put AI data center demand in the U.S. growing more than 30-fold by 2035, from a 2024 baseline of about 4 gigawatts to over 120 gigawatts.

Grid infrastructure was not built for that kind of step change. Permitting for new power plants and transmission lines can take well over a decade in the U.S. and EU. The median time from an interconnection request to commercial operation for large loads has stretched past five years. The mismatch is already visible in wholesale markets: PJM's capacity market clearing price for 2026-2027 delivery climbed to roughly ten times its 2024-2025 level, with rapid data center growth cited as a major driver. Similar pressure is building in New York ISO and ERCOT.

The water side of the ledger

Cooling is the other half of the constraint, and it interacts with the power problem rather than sitting next to it. Data centers cool their servers two ways: energy-intensive air cooling, or water-intensive evaporative cooling. Air cooling uses on average about 10% more electricity than evaporative cooling, roughly double that on a hot day. That's the tradeoff in one sentence: reducing stress on the grid tends to increase stress on the watershed, and vice versa.

On direct consumption, U.S. data centers used 17.4 billion gallons of water in 2023. That figure is projected to rise to somewhere between 38 and 73 billion gallons by 2028, with hyperscale facilities — the largest class of data center — expected to account for roughly half of it. A single large facility can consume up to 5 million gallons a day, comparable to a town of 10,000 to 50,000 people. For a more granular sense of scale: researchers at UC Riverside estimate each 100-word AI prompt uses roughly one bottle of water once the full cooling chain is accounted for.

It's worth being precise about what's counted, because this is where a lot of corporate messaging gets slippery. Indirect water use — water consumed generating the electricity a data center draws — typically runs more than ten times higher than direct, on-site cooling consumption. When a hyperscaler announces a facility has “solved” its water problem, it's usually describing what happens inside the building's walls. Nvidia's new warm-water cooling system, for instance, claims to eliminate most water use at the facility level by running coolant at 45°C and rejecting heat without evaporation. That's a genuine engineering advance, but by the company's own framing it addresses roughly a quarter to a third of a data center's total water footprint. The rest — power generation, chip manufacturing — sits outside the fence line and outside the announcement.

Where the two constraints collide

Site selection used to be mostly a tax-and-latency decision. It's increasingly a resource-scarcity decision. Developers are now screening locations for grid interconnection capacity and water availability simultaneously, and the two don't always point the same direction — the driest regions often have the most available land and cheapest power, and the regions with abundant hydro or water access often have the least free transmission capacity.

Policy is starting to catch up unevenly. More than 200 data-center-related bills were introduced across U.S. states in 2025, with over 40 enacted; moratoriums are under active discussion in more than 20 states as of 2026. Most of that legislative energy has focused on disclosure and reporting rather than hard caps, because regulators still lack basic visibility into facility-level consumption — fewer than a third of data center operators historically tracked water use at all.

The industry's own response has bifurcated. Some hyperscalers are moving toward closed-loop and liquid cooling systems that can cut freshwater use by up to 70%, or in some designs approach zero direct consumption. Others are explicitly choosing evaporative cooling where water is abundant, arguing it's the more responsible choice precisely because it reduces the amount of new power generation — and associated indirect water use — a facility requires. Google's own data shows roughly two-thirds of its data centers still use evaporative cooling for this reason, even as the company publishes new voluntary water-stewardship standards.

A regional footnote worth watching

For capital allocators working outside the U.S. hyperscale corridor, Latin America is a useful illustration of the same dynamic playing out at a smaller, earlier-stage scale. Regional data center investment grew at roughly 42% annually between 2022 and 2025, yet the region's installed capacity — around 1,450 megawatts across roughly 500 facilities — remains less than a third of Northern Virginia's alone. The growth is real, but so is the constraint: Peru and Ecuador in particular have experienced seasonal hydroelectric shortages leading to periodic blackouts that already affect industrial operations, a preview of what happens when new, concentrated load arrives faster than firm power supply.

That's the pattern underneath all of this, at every scale from Loudoun County to Querétaro to Quito: AI's appetite for compute is a real and durable demand signal, but the infrastructure that has to feed it — power plants, transmission lines, water rights, cooling technology — moves on a slower, more capital-intensive, more locally contested timeline than the software does. The bottleneck for the next phase of AI growth may not be model capability at all. It may simply be gigawatts and gallons.