AI Data Centers, Electricity and Water: The 2026 Investor Checklist

An AI data center is a computing facility, but its economics depend on physical infrastructure as much as on GPUs. Servers require dependable electrical supply, network connectivity, cooling equipment, skilled maintenance and a contractual route to sell computing capacity. A campus may be announced years before it can operate at planned capacity. Investors and local communities should therefore distinguish a proposed power requirement from a signed utility agreement, a permitted connection, a commissioned building and equipment serving paying customers.

The International Energy Agency's Energy and AI report analyzes how data-center electricity demand could evolve under multiple scenarios. Its scenarios are not guarantees of what any single operator will consume. A headline about terawatt-hours across the globe does not reveal whether one project has secured transformers, transmission upgrades or affordable power at the right location. RecoupRev's central question is how to read an AI infrastructure announcement without confusing aspiration with execution.

A facility's stated megawatt capacity may refer to information-technology load, total site load, contracted utility capacity or a future build-out phase. These are different denominators. If an operator describes hundreds of megawatts of planned capacity, a reader should ask how much is already energized and occupied, when the remainder is scheduled, and which permits or utility milestones remain outstanding.

Electricity demand is a product of power draw and time. A hypothetical continuously operating one-megawatt IT load consumes 8,760 megawatt-hours in a 365-day year before accounting for the site's other electrical loads. Real facilities vary in utilization and power management. This illustrative calculation does not describe an actual data center. It does show why a small difference in utilization, electricity price or cooling overhead can compound into material annual costs.

Power contracts may be fixed, indexed or exposed to wholesale market prices. They can include minimum usage commitments, construction deposits, interconnection expenses and penalties. A long-term deal labeled renewable does not necessarily mean the facility receives physically uninterrupted carbon-free power every hour. Investors should separate annual renewable-energy accounting from real-time matching and grid emissions at the point of use.

Power usage effectiveness, or PUE, compares total facility energy with the energy delivered to information-technology equipment. An illustrative PUE of 1.2 means total facility energy is one-fifth greater than the IT energy under the measured boundaries. That overhead can include cooling, electrical conversion, lighting and supporting systems. Lower PUE generally signals less overhead, but results depend on climate, equipment density, operating conditions and measurement method.

Comparisons are misleading when one operator reports a best-case design figure and another reports an annual measured figure. A newly opened, partially occupied campus may behave differently from a fully populated campus. The correct question is whether the reported PUE uses consistent definitions over a representative period and whether it includes the same load boundaries.

Even excellent PUE cannot tell an investor whether GPU capacity is profitable. A well-cooled facility with little customer demand can still lose money. Conversely, a facility with slightly higher overhead might perform better economically if it offers superior uptime or access to reliable, competitively priced power. Treat PUE as one efficiency measure, not a valuation multiple.

Water usage effectiveness, or WUE, is a site-based measure of water consumed relative to IT equipment energy, commonly expressed in liters per kilowatt-hour. The US Department of Energy defines and discusses WUE in its guidance on cooling-water efficiency. A site's water profile depends on cooling design, climate, treatment, water source and seasonal conditions. Cooling towers can use water to reject heat; other systems may shift more of the burden to electricity.

For a hypothetical facility consuming 100 million kWh of IT electricity per year, a WUE of 0.5 liters/kWh corresponds to 50 million liters of water annually under that definition. The calculation is illustrative and depends on what the facility measures as water use. Readers should not attach this number to any operator without audited site-level reporting.

A water-efficient system in one region may be impractical in another. Local drought, competing agricultural use, municipal supply constraints and wastewater capacity all matter. Some facilities use reclaimed water. Others may rely on cooling strategies that minimize onsite consumption while increasing electricity demand. The environmental comparison needs both WUE and power-source information rather than a single favorable metric.

Modern high-density computing changes the heat-removal challenge. A traditional air-cooled data hall moves conditioned air through server racks. Liquid-based systems transport heat through water or other engineered fluids nearer to high-power components, with heat exchangers transferring it elsewhere. The detailed design may include facility loops, manifolds, coolant distribution units, pumps and redundancy.

Liquid cooling can support dense accelerators, but it is not free. Installation and maintenance complexity can increase, and leaks or component failures require careful controls. Retrofitting an existing building may be materially different from designing liquid cooling from the outset. Operators should disclose which racks are ready for a specific hardware generation, whether cooling components are available, and whether maintenance procedures have been tested at scale.

The same problem appears in equipment procurement. Buying a large order of GPUs without confirmed rack power, interconnect hardware and cooling infrastructure can strand capital. A model that assumes every purchased accelerator begins generating revenue immediately may overstate cash flow. Investors should inspect commissioning and customer acceptance milestones.

Before a data center draws its planned load, utilities may need substation upgrades, transmission reinforcement, new switchgear and detailed engineering studies. The public announcement date is therefore not equivalent to the date the grid can energize the site. Delays can arise even when the land and server equipment are available.

Power reliability is another consideration. A facility that promises uninterrupted computing needs a strategy for grid interruptions, backup equipment, fuel logistics and recovery procedures. These systems add cost and may face their own environmental permitting. A comparison of two campuses should normalize for the required uptime, redundancy and location rather than compare their cheapest advertised electricity prices.

Ask whether the developer has received a binding interconnection commitment, how much capacity is firm, which upgrades it must fund, and which dates remain conditional. If a company provides only its eventual theoretical demand, future financing requirements may be substantially larger than current capital expenditures suggest.

Announced contracts are not always equivalent to recognized revenue. A customer reservation for capacity may depend on readiness, benchmarks, service-level agreements or staged ramp-up. Some deals involve prepayments or take-or-pay terms, but the exact enforceability matters. Without a disclosure, outsiders should not assume all reserved megawatts will be utilized continuously at headline prices.

A simple earnings model begins with usable IT capacity, expected utilization, contracted pricing and variable costs. It then subtracts power, network, maintenance, labor, depreciation and financing expense. Replacement cycles are material because accelerators and associated networking hardware can become technologically outdated before the building reaches the end of its useful life.

If the company sells capacity through intermediaries, concentration risk matters. One hyperscale client or a single reseller can represent a large share of expected cash receipts. A project also faces the possibility that computing demand changes or newer chips provide more work per watt. Long-duration infrastructure financing against rapidly evolving hardware deserves scrutiny.

The IEA discusses future energy requirements under explicit scenarios and assumptions. Those assumptions can change with AI adoption, hardware efficiency, electricity prices, power availability and technology. It is appropriate to say that the agency projects substantial demand under its analyzed pathways. It is not appropriate to convert a scenario into a guaranteed revenue line for one company.

The agency's energy-supply analysis considers different generation sources and the challenge of meeting growing demand. The important distinction is between global electricity required by all data centers and incremental demand attributable to AI. Even when an article uses the phrase AI electricity, underlying studies may include conventional cloud services and other digital workloads. Verify the accounting boundary before comparing forecasts.

A responsible investment article should provide the forecast year, scenario, energy unit, baseline, and source methodology. Using energy statistics without dates can distort the pace of change. This guide intentionally avoids projecting the share price of any specific data-center operator.

Project A announces 300 MW of eventual capacity but has only 50 MW of firm grid access, limited cooling information and an unsigned customer memorandum. Project B announces 120 MW, has most of its interconnection approved, a commissioned first hall and documented customer contracts. It would be wrong to rank A above B simply because its announced number is larger.

Instead, estimate the expected timeline to cash-generating service, remaining capital spend and probabilities of delay. Treat project A's expansion as contingent rather than realized. Treat project B's customer contracts according to disclosed terms, not their face value. Ask whether the projects face different electricity or water constraints.

The reverse can also be true. Project A may have valuable land, an experienced development partner or favorable future contracts that justify some of its option value. The purpose of the framework is not to assume small is safe. It is to make promised capacity, available capacity and economic capacity explicit.

Ask for the IT versus site power figures, the energized portion, expected load factor, power price structure, PUE measurement method, WUE boundaries, water source, cooling readiness, infrastructure construction schedule and revenue recognition assumptions. Add questions about backup energy and the availability of skilled maintenance workers. A company that can answer clearly is easier to evaluate than one offering only an AI demand narrative.

Local residents may reasonably ask about water consumption, construction traffic, noise, emergency backup systems and the cost allocation for utility upgrades. These are site-specific questions; a global electricity forecast cannot answer them. Public permitting records and utility filings are often more directly relevant than presentations at technology conferences.

Investors should additionally track whether facility spending has outrun customer payments, and whether debt repayment assumes high occupancy immediately after opening. Forecasts can be helpful when their uncertainty is visible.

**Does a lower PUE mean lower water use?** Not necessarily. Some cooling designs reduce electrical overhead through evaporative cooling and consume more water locally. Read PUE and WUE together using the same time period.

**Does a power purchase agreement guarantee 24/7 renewable power?** Not automatically. Energy matching, delivery and emissions accounting depend on contract terms, location and the generation profile.

**Is a planned 1 GW campus already using 1 GW?** A proposal describes potential future capacity unless the operator identifies energized and operating portions. Planned, contracted and drawn power are distinct figures.

**Can efficient new chips reduce electricity demand?** Efficiency per unit of computing may improve, but total demand can still rise if usage expands faster. Neither direction is certain.

**Which numbers deserve the most attention?** Metered IT energy, site energy, water use, firm grid access, utilized capacity, signed revenue and financing obligations. This is an analytical framework, not a claim of audited performance for any company.

A rigorous project model should separate four steps: permitted capacity, grid-connected capacity, physically installed IT power, and revenue-producing utilization. Each step requires expenditure and may slip on a different timetable. If 200 MW is announced but only 40 MW has a funded and dated utility connection, assigning revenue to all 200 MW in year one is not an optimistic sensitivity—it is a category error. Investors can create a schedule with each phase's contracted load, commissioning date, anticipated customer acceptance, power expenses and required capital expenditure.

For an illustrative sensitivity, take 10 MW of sustained IT load, a PUE of 1.3 and electricity costing $0.08 per kWh. The facility-level average load is approximately 13 MW. At constant utilization over a 365-day year, purchased electricity totals about 113,880 MWh and a hypothetical energy bill of roughly $9.11 million. These are arithmetic assumptions, not a reported operator tariff. Raising the unit price to $0.12 increases the annual bill to approximately $13.67 million before taxes and demand charges. A single contract parameter can therefore change the business case by millions of dollars.

A board should run this calculation under reduced utilization, delayed interconnection and changing tariffs. It should examine whether fixed take-or-pay customer contracts offset fluctuating electricity costs and whether an energy price hedge actually matches the facility's hours of use. Gross megawatts sound impressive in a headline. Metered energy, usable GPU hours and cash receipts are what ultimately support debt service and investor returns.

Local authorities may publish environmental reviews, zoning applications, water permits and utility proceedings. Residents should ask which load phase is under review, whether projected water demand is annual or peak-day, and who funds network upgrades. Cooling withdrawals and consumption are distinct: water withdrawn from a source may partly return, while water consumed is not returned to that source. Project documentation should state the boundary clearly.

Readers should also distinguish developer calculations from regulator-verified monitoring. A modeled annual WUE may depend on weather and workload assumptions that actual operations violate. Publishing operating metrics alongside the original design target lets communities assess performance over time. All such comparisons require equivalent periods and measurement boundaries.