AI data centers need renewable energy because they are the fastest-growing new electricity load on the planet, and the grids they plug into still burn fossil fuels for most of that power. The IEA’s Energy and AI report puts data centres at about 415 TWh in 2024, around 1.5% of global electricity, and projects roughly 945 TWh by 2030, with AI-accelerated servers the main driver. Wind and solar are the cheapest new supply and the only way operators can honor their net-zero pledges. But they are intermittent, and AI load runs around the clock. So the real role of renewables is bulk energy, paired with storage, firm clean power such as nuclear and geothermal, and hourly matching to cover the gaps.

How much power does AI actually use?

Less than most headlines claim, but growing faster than any other load. The widely repeated “8% of global electricity by 2030” figure has no primary source behind it; the IEA’s 2025 projection is just under 3% globally. The United States is the concentration point. Lawrence Berkeley National Laboratory’s 2024 data center energy report measured 176 TWh in 2023, or 4.4% of US electricity, and projects 325 to 580 TWh by 2028, between 6.7% and 12%.

Density is what changes the engineering. An NVIDIA DGX B200 draws up to 14.3 kW for eight GPUs, and a GB200 NVL72 rack is rated around 120 kW. A conventional enterprise rack ran 5 to 10 kW. That is why new AI halls are liquid-cooled and why a single campus can request 500 MW to 1 GW of grid capacity.

Why intermittent renewables are hard to pair with AI

A training run cannot pause when the wind drops. Utility solar delivers roughly 20 to 30% of its nameplate capacity over a year; onshore wind 30 to 45%. Buying enough wind and solar to match annual consumption is cheap, and that is what most “100% renewable” claims mean under the GHG Protocol’s market-based Scope 2 accounting. It does not mean the servers ran on clean power at 3 a.m. in January.

Google publishes the harder metric: hourly carbon-free energy by region. Its global average sits around two-thirds, with individual regions ranging from roughly 30% to over 90% depending on the local grid. The EnergyTag standard is what makes hourly claims auditable. Batteries help, but a 4-hour lithium system covers an evening peak, not a week of winter cloud.

What operators are actually building

• Power purchase agreements for new wind and solar, typically 10 to 20 years. Utility-scale solar is among the lowest-cost new generation in NREL’s Annual Technology Baseline.
• Nuclear. Microsoft’s 20-year deal to restart Three Mile Island Unit 1 (now the Crane Clean Energy Center), Amazon’s co-location at Susquehanna, Google’s small modular reactor agreement with Kairos Power, Meta’s Clinton deal. Delivery runs from 2027 to the mid-2030s, so none of it powers today’s clusters.
• Geothermal as firm baseload where the resource exists, such as Google’s Fervo project in Nevada.
• Natural gas as the bridge. The IEA expects gas to supply a large share of new US data-center demand through 2030, and turbine order books now stretch years out. This is the part most “green data center” marketing omits.
• Efficiency. Direct-to-chip liquid cooling, hyperscaler PUE near 1.1 versus an industry average around 1.5, and shifting flexible workloads to hours and regions with surplus clean power.

Where it goes wrong

Interconnection. In several US markets a new 300 MW load waits 3 to 7 years for a grid connection, which is why operators are buying generation behind the meter. Despite record renewable purchases, reported emissions at both Google and Microsoft have risen since 2020; Microsoft’s sustainability report attributes most of its increase to data center construction and supply chains (Scope 3), not just electricity.

What this means if you buy AI tools for video and marketing

We run generative video, upscaling and transcription for client projects, and the energy question now shows up in RFPs from brands with Scope 3 targets. Three things we have learned:

• Cloud region is the biggest lever you control. The same render in a hydro- or nuclear-heavy region can carry a location-based footprint several times lower than in a coal-heavy one, usually at no price difference.
• Iteration, not inference, is the waste. One generated clip costs seconds of GPU time. Forty rejected takes to get one usable shot is where the energy goes. Tighter prompts and reference frames cut more than any vendor’s green badge.
• Ask vendors for hourly-matched or location-based figures. Annual renewable certificate claims tell you almost nothing.

If you are evaluating a data center or AI provider, ask for its hourly carbon-free energy percentage by region and its PUE. If you are a buyer of AI tools, pick the lowest-intensity region your vendor offers and set a rejected-take budget before you start generating.