An AI campus does not become reliably powered by announcing a large annual renewable purchase. Electricity must reach the site at the hours it operates, with reserves for stress and failures. That physical requirement explains why solar and natural gas can expand together: one supplies energy when available, while the other can provide controllable output.
The Department of Energy’s cited forecast places U.S. data-center use in 2028 between 325 and 580 terawatt-hours, or 6.7% to 12% of consumption. The wide range matters as much as the upper number. It describes data centers, not AI alone, and it does not mean every proposed campus will connect.
Separate energy totals from dependable hourly supply
Annual renewable certificates or contracts need not match the electricity physically serving a facility overnight. Hourly and location-aware accounting asks a different question. Solar with batteries can shift output, but a few hours of storage are not the same as coverage through several cloudy days.
Gas can fill gaps, while creating combustion emissions, methane, fuel and long-lived asset risks. Calling a plant a temporary bridge does not establish an affordable conversion plan. Evaluate the present plant and permits on their actual terms, with any future alternative supported separately.
Verify the load before building around it
Large campuses request power in blocks that differ from gradual growth across many customers. Developers may inquire in several locations before choosing one, so summing requests can overstate committed demand. Distinguish inquiry, land control, financing, equipment orders and binding service agreements.
Underbuilding threatens reliability; overbuilding can leave others paying for underused assets. Deposits, minimum payments and dedicated construction contributions should address delays, changed ownership and lower use. A short customer commitment may not cover a network asset lasting decades. Clear cost allocation matters more than a headline megawatt figure.
Treat the grid as part of the project
Generation and servers still need transmission, substations, transformers and interconnection approval. Good national supply can coexist with severe local congestion. A complete solar project or server building can therefore wait for usable grid access.
Site selection should reflect available power and delivery timing alongside land, fiber and water. Onsite generation may shorten a connection dependency but adds fuel, permitting, noise and backup requirements. It changes constraints rather than removing them. Communities also need visibility into who bears infrastructure costs and local impacts.
Test flexibility with operating evidence
Some inference must respond immediately, while selected training or batch work may move in time or place. Flexibility can help stressed grids and clean supply, but idle expensive hardware has costs. Pricing and contracts must reward a measurable response rather than assume it.
Track available megawatts, response time, duration and delivery against promises. Models of flexible demand are not substitutes for actual performance. Also track generation and grid projects entering service, not only proposals. Efficiency may reduce electricity per task while increased usage raises total demand. A credible power plan keeps those uncertainties visible and proves committed load, dependable hourly supply and fair allocation of risk.