AI Energy Consumption: Scientists Race to Slash Data‑Center Power Use

While sipping coffee in Berlin, I asked Google’s Gemini a simple question and watched the answer appear in seconds. AI Energy Consumption is hidden behind that flash; each reply travels through fiber, hits a massive rack of processors, and consumes electricity before it returns to my screen.

AI Energy Consumption and the Data‑Center Surge

Data centres already power the web, but generative AI has turned them into power‑hungry behemoths. Google estimates a median Gemini prompt uses about ↓0.24 Wh, roughly the energy of a nine‑second TV show. Multiply that by billions of daily queries and the load explodes.

In the United States, data‑centre demand reached ↑224 TWh in 2025, according to the International Energy Agency—over 5 % of national electricity use, a jump from under 2 % in 2018.

“Every megawatt powered by fossil fuel pushes us back,” says Eric Masanet, UC‑Santa Barbara.

Must Read Intel Review our latest briefing on this sector

Related Intel: Claude Code Triples Engineer Output, Sparking a Surge in Product Thinkers

Tech giants are pouring tens of billions into AI‑focused facilities. Some “hyperscale” sites aim for gigawatt‑class capacity, comparable to a tenth of Los Angeles’ grid.

Software tricks to cut the draw

Researchers are pruning models, creating task‑specific versions that run on a fraction of the parameters. A 2025 UNESCO study found that compact models can slash energy use by more than 90 % for the same job.

Mixture‑of‑expert architectures activate only relevant sub‑networks, routing queries to the most efficient path.

Alternative neural designs, such as the revived xLSTM, keep a short summary of the conversation, avoiding the quadratic growth that plagues transformers.

Hardware innovations

Chip makers are scaling up. Wafer‑scale processors from Cerebras pack nearly 70× more transistors than a typical GPU and cut internal communication energy by a factor of 143.

Analog “gain cells” merge memory and compute, promising four orders of magnitude reduction in the most energy‑intensive step of transformer inference.

Photon‑based chips, still experimental, could slash power needs by up to tenfold, according to research at Friedrich Schiller University.

Beyond tech, siting matters. Most U.S. AI farms sit in northern Virginia, a region with limited renewable resources and water stress. Relocating to the Midwest could trim carbon and water footprints by 73 % and 86 %, respectively, says Cornell’s Fengqi You.

Companies tout offsets and renewable purchases, but unless the clean energy exceeds the data‑centre draw, net emissions stay flat.

Regulators are beginning to act. Over 200 state‑level proposals target data‑centre sustainability, and the federal government has introduced 18 measures, though many remain drafts.

The race to tame AI’s power appetite is now a race to keep the planet livable, echoing lessons learned during the pandemic about scaling infrastructure responsibly.

Intel provided by Julian Reed (Consumer Electronics Expert).