The Hidden Bottleneck of the AI Revolution: Why Power and Water Decide the Winners

When we talk about the rapid advancement of artificial intelligence, the conversation usually revolves around breakthrough algorithms, massive datasets, and cutting-edge silicon. But behind every large language model lies a far more stubborn constraint — one that can’t be solved with a better architecture or a bigger training run. It’s physical. AI’s real bottleneck is the infrastructure required to feed it: electricity, water, and the industrial capacity to build at a scale most of the tech sector has never attempted.

The companies that win the AI race won’t simply be the ones with the best code. They’ll be the ones who can secure power and water, and build the facilities to use them. Here’s why.

The Exponential Hunger of AI Compute

Traditional cloud computing is relatively efficient. A standard rack of servers in a legacy data center draws roughly 5 to 10 kilowatts (kW). AI is a different beast entirely.

Training and running modern AI models depends on dense clusters of GPUs, and the power density is staggering. An air-cooled rack of NVIDIA H100s runs around 40 kW. NVIDIA’s current GB200 NVL72 rack draws roughly 120–130 kW, the newer GB300 pushes past 160 kW, and the industry is already designing toward megawatt-scale racks. That isn’t a slight increase over the legacy world — it’s a 10-to-20-fold leap in density that older facilities were never engineered to handle.

Scale that up and the numbers leave the realm of facilities and enter the realm of utilities. EPRI and Epoch AI project that the largest individual AI training runs will require 1 to 2 gigawatts of power by 2028, and RAND estimates single runs could reach 8 GW by 2030. Meta’s Hyperion campus is being engineered for roughly 5 GW across eleven buildings, targeting full capacity around 2028. For reference, 5 GW is the output of about five large nuclear reactors — dedicated to a single company’s computer.

The Data Center Dilemma

Building a data center today is less about finding cheap land and more about finding a grid that can actually support it. Three forces are colliding.

Grid capacity. The world’s largest data center hubs are running into hard physical limits. In Northern Virginia, the biggest market on Earth, a developer entering Dominion’s interconnection queue today may not get a grid connection until the early 2030s. PJM, the regional grid operator, approved a record $6.7 billion transmission build-out in 2025 specifically to relieve these bottlenecks, and still projects a sharp rise in reliability risk as demand outpaces supply.
The cooling problem — and where the “tax” really goes. More power in means more heat out, and high-density AI racks run far too hot for conventional air cooling. The industry’s answer is liquid cooling, and here’s the nuance most coverage gets wrong: liquid cooling is the efficiency move, not the power drain. It cuts cooling overhead from roughly 0.5–1.2 kW per kW of compute (air) to 0.1–0.3 kW (liquid), pulling facility PUE from 1.4–1.6 down below 1.2. But that efficiency doesn’t come free — it shifts the burden from electricity to capital, engineering complexity, and, critically, water.
Availability versus innovation. Without secured power, projects stall. Companies are now racing to lock in multi-decade supply: Amazon signed a 17-year deal for 1.9 GW from the Susquehanna nuclear plant; Microsoft committed to a 20-year agreement to restart Three Mile Island. Compute capacity is effectively capped by megawatt availability, and the smartest players are treating power procurement as a strategic asset, not a line item.

The AI Thirst Nobody Budgeted For

Power gets the headlines. Water is the constraint quietly deciding where these facilities can actually go.

Liquid and evaporative cooling consume enormous volumes of fresh water. AI data centers use an estimated 10 to 50 times more cooling water than traditional server farms. A single large facility can consume several million gallons a day; Google’s Council Bluffs campus alone has drawn close to three million gallons daily. Globally, data center water use crossed an estimated 264 billion gallons in 2025 and is climbing fast.

The problem isn’t just the volume — it’s the geography. According to a Tom’s Hardware analysis, roughly two-thirds of the 809 AI data centers planned or under construction in the U.S. are slated for areas already facing water stress. MSCI finds nearly a third of facilities under construction sit in regions projected to grow more water-scarce by 2050. Local governments are responding: Tucson now subjects very large water users to council review and an enforceable conservation plan. As one industry observer put it, for the next wave of AI infrastructure, water — not megawatts — may become the decisive factor.

An AI Data Center Is a Mega-Manufacturing Project in Disguise

Step back and the picture clarifies. A gigawatt-scale AI campus has almost nothing in common with the cloud data centers of the last decade. It has everything in common with a steel mill, a semiconductor fab, or a petrochemical plant.

These are heavy-industrial megaprojects. They require dedicated power generation and high-voltage interconnection, substantial water sourcing and treatment, substations and transmission corridors, multi-year permitting, a supply chain for transformers and turbines that is itself constrained, and thousands of skilled trades on site. The discipline that demands integrated power, water, and industrial delivery — managed against a buildable schedule — is not software expertise. It’s mega-manufacturing and infrastructure expertise. The teams that can orchestrate all of it at once will determine which AI ambitions actually get built, and which stay on a slide.

The Path Forward: Efficiency and New Supply

The industry can’t simply plug into coal plants — the carbon math and corporate commitments rule it out. So the constraint is forcing real innovation on four fronts:

Smarter silicon. Chipmakers are now competing on performance-per-watt, doing more computation per unit of electricity.
Firm, clean power. Tech giants are securing their own zero-carbon baseload: Google’s small modular reactor deal with Kairos (~500 MW), Microsoft’s nuclear restart, and a fast-growing pipeline of geothermal — Google with Fervo in Nevada and Meta contracting 300 MW from next-generation geothermal developers.
Closed-loop water. Operators are moving to zero-evaporation, closed-loop cooling and tracking Water Usage Effectiveness as rigorously as power. Microsoft’s latest designs aim to eliminate cooling-water evaporation entirely.
Algorithmic efficiency. Smaller, distilled models are delivering comparable results with a fraction of the compute — and the resources behind it.

The Bottom Line

Data has been called the new oil. But oil is worthless without an engine to burn it and a cooling system to keep that engine from seizing. In AI, computers are the engine, electricity is the fuel, and water is the coolant — and right now, the fuel and the coolant are scarcer than the engine.

As AI scales from novelty to core enterprise infrastructure, securing power and water is no longer a facility manager’s concern. It’s a boardroom-level strategic imperative, and increasingly a mega-manufacturing one. The winners won’t just write the best code. They’ll be the ones who solved the resource equation and built the thing.

At AuerBridge, we sit at exactly this intersection of power, water, data centers, and mega-manufacturing. No matter the scale, we can deliver valuable research and insights. Research for your firm is available here. If you’re planning the next step in infrastructure for your organization and want to pressure-test the power, water, or mega-manufacturing side before you commit, let’s talk.

Sources: IEA, Epoch AI, RAND, Vertiv (cooling/PUE), Tom’s Hardware (water/drought siting), MSCI (water scarcity), Introl (WUE), Introl (nuclear deals), Canary Media (geothermal).