Powering AI Data Centers: From Hourly Demand to Nuclear and Renewable Deals

Physical delivery and contractual procurement answer different questions

Physically, a data center draws power from the grid connection or an onsite system. Grid operators continuously balance injections and withdrawals while transmission constraints determine which resources can respond. A corporate power purchase agreement is a contract: it may settle financially against a generator, transfer renewable attributes or support project financing even when the facility and generator sit in different places. Annual certificates can match yearly consumption while leaving individual hours unmatched. None of these structures means a company directs a dedicated stream of electrons from a distant plant to its servers.

Procurement claims should specify geography, timing, additionality and residual supply. A new wind contract in the same constrained grid has a different effect from buying unbundled certificates created years earlier elsewhere. A colocated generator may still rely on the grid for outages and balancing. A data center advertised as powered by one resource may consume a portfolio in practice. Transparent accounting can recognize contractual progress without confusing it with physical reliability or real-time emissions. Grid planning must use actual load and deliverability, not marketing allocation.

Read the IEA forecast as a scenario with a publication date

The IEA’s 2025 report estimated that data centers used about 415 TWh globally in 2024 and projected around 945 TWh in 2030 in its base case. The report also describes substantial uncertainty: efficiency gains, AI adoption, hardware, utilization and deployment geography can move demand. These figures include data centers broadly, not only generative-AI training. Quoting the 2030 number without its 2025 vintage, scenario and scope converts analysis into false certainty. Forecasts should be updated as new measured consumption and project completions become available.

Per-query energy numbers are especially fragile. The workload may use a small or large model, short or long context, batch or interactive serving, older or newer hardware, and a facility with a different power-usage effectiveness. Training runs also vary by model design, data, hardware utilization and restarts. Operators can measure energy for a defined workload and disclose system boundaries; outsiders should not turn one test into a permanent universal constant. For grids, the more relevant inputs are megawatts, location, ramp, load factor, commissioning schedule and ability to reduce demand.

Worked example one: annual renewable matching can hide an hourly gap

Imagine an illustrative 100-megawatt data center running at a flat 90 MW, or 788,400 MWh over a 365-day year. It contracts enough annual wind and solar generation to equal that energy. On a spring afternoon, contracted output may exceed 90 MW; on a windless evening, it may be far below. Annual totals can match while the facility relies physically on the surrounding grid in low-output hours. Curtailment may occur in surplus hours if transmission or storage cannot use the generation. The annual claim and the hourly operating reality are therefore both true but describe different boundaries.

An hourly strategy can combine geographically relevant wind and solar, batteries for shorter shifts, firm low-carbon resources, demand response and grid supply. It should track unmatched hours rather than averaging them away. Storage duration and cycling limits matter: a battery sized for a four-hour evening peak does not bridge a multi-day weather event. Transmission can diversify weather and access remote resources, but it requires corridors and permitting. The portfolio should be modeled against weather years, outages and network constraints, not one representative day.

Renewables, storage, transmission and flexibility are complementary tools

Wind and solar can contribute large amounts of energy and, in favorable hours, capacity. Their variable output creates a balancing problem rather than a reason to dismiss them. Batteries shift energy and provide fast grid services; long-duration technologies may cover longer gaps if technically and economically viable. Transmission connects diverse supply and demand. Existing hydro, geothermal, nuclear and dispatchable generation can supply firm capability depending on region. Efficiency lowers the entire requirement. The least-cost reliable mix is location-specific and changes as technologies, fuel prices and grid rules evolve.

Data centers can also offer flexibility, though not every workload can move. Batch training, data preparation or backup charging may be scheduled around grid conditions, while real-time inference and critical services have stricter latency and availability needs. Facilities can stagger equipment startup, use thermal storage or participate in demand-response programs if service commitments permit. Flexibility must be measured and contractually available to the grid operator; a theoretical ability to shift work is not a planning resource until controls, communications and customer obligations support it.

Nuclear announcements fall into three materially different categories

A restart involves returning an existing but closed reactor to service, subject to engineering work, licensing, fuel, staffing and economics. A power agreement with an operating plant purchases electricity or attributes under defined terms; it is not the purchase of the plant itself. A small modular reactor announcement often concerns future development, investment or an offtake proposal for a design that may not yet be licensed or built. These categories have different schedules and risks. Headlines that call every agreement “buying a nuclear plant” erase ownership, regulation and project status.

Nuclear generation can provide firm low-carbon electricity, but new projects face capital cost, construction, supply-chain, waste, water, safety and public-acceptance questions. Restarts depend on site condition and regulatory approval. SMRs seek benefits from standardization and modular construction, yet commercial performance must be demonstrated project by project. A buyer’s long-term contract can improve revenue certainty without transferring all project risk. Analysts should read the counterparty announcement, plant status, regulator record, delivery year, volume, termination rights and responsibility for overruns.

Worked example two: compare a portfolio against the hourly load

Suppose, illustratively, the 90 MW facility considers 140 MW of wind, 100 MW of solar, a 50 MW four-hour battery, a 30 MW firm contract and up to 20 MW of curtailable computing for selected hours. Planners simulate hourly production and load across several historical weather years. In many hours renewables cover the load and charge the battery. During a calm winter evening, wind and solar produce 18 MW, the firm contract supplies 30 MW, and the battery can provide 42 MW for four hours. The remaining requirement is met by the grid or a verified workload reduction.

That portfolio cannot be judged by adding nameplate capacities to claim 320 MW for a 90 MW load. Output timing, battery energy, outages and transmission limits decide adequacy. Nor should the example be read as an optimal design; prices and resources are intentionally absent. A real study would compare capital and contract costs, grid upgrades, emissions, reliability contribution, curtailment and community impacts. It would also test a lower-demand case created by efficient chips, cooling and higher utilization. The first megawatt avoided can be faster than the next megawatt built.

Large concentrated loads can require substations, transmission upgrades and new generation. The allocation of those costs affects other customers. A project may create construction work and tax revenue while competing for land, water, equipment or scarce grid capacity. Cooling design changes water demand. Backup generators affect local air quality and noise. Communities need credible load schedules, infrastructure plans and emergency procedures before incentives are granted. Regulators must decide which upgrades are dedicated to the customer and which benefit the wider network.

Developers can improve the bargain by selecting grid-supportive sites, funding attributable upgrades, disclosing water and backup plans, procuring additional supply, committing measurable flexibility and reporting hourly progress. Utilities should avoid assuming every announced campus arrives on time or at full load, while also planning for correlated projects. The durable question is not which single technology wins the AI boom. It is whether an evolving portfolio can serve new demand reliably, affordably and with acceptable environmental and local consequences.

1. Separate physical electricity delivery from contractual purchases and environmental attributes.
2. Date every demand forecast and preserve its scope, scenario and uncertainty.
3. Evaluate annual procurement and hourly matching as different claims.
4. Distinguish nuclear restarts, operating-plant power agreements, proposed SMRs and plant ownership.
5. Model efficiency, generation, storage, transmission, flexibility, grid upgrades and community costs together.

Sources

• IEA Energy and AI (April 2025)
• US Department of Energy report on data-center electricity demand
• Google 24/7 carbon-free energy methodology and progress
• US Nuclear Regulatory Commission reactor licensing