In Abilene, Texas, a data center campus that did not exist two years ago now draws enough electricity to power a mid-sized city. In Virginia, the world highest concentration of data centers is straining a power grid that was not designed for this load. In Singapore, the government imposed a moratorium on new data center construction for three years because the city-state was running out of power capacity. In Ireland, data centers now consume more electricity than all of the country urban homes combined.
The AI boom has a physical dimension that the coverage of language models and benchmark scores tends to obscure. Behind every query to ChatGPT, every image generated by Midjourney, every hour of AI-assisted coding, there is a physical building full of specialized hardware consuming significant amounts of electricity and generating significant amounts of heat. The scale of those buildings, and the rate at which they are being constructed, has made AI infrastructure one of the largest drivers of energy demand in history. And the demand is accelerating faster than the power grid can accommodate it.
The Numbers
The International Energy Agency estimated in 2024 that global data center electricity consumption would double by 2026, with AI workloads accounting for the majority of that growth. Goldman Sachs research estimated that data centers would consume 8% of US electricity by 2030, up from approximately 3% in 2022. Microsoft, Google, Amazon, and Meta each committed to spending fifty billion dollars or more on data center infrastructure in 2024 alone. The capital expenditure figures for AI infrastructure have reached levels that would have seemed implausible three years ago.
Training a single large frontier AI model — the kind of training run that produces a GPT-4 or a Claude Opus — consumes roughly the same energy as five hundred US homes use in a year. A single training run for a large model requires ten thousand to fifty thousand specialized AI chips running continuously for weeks or months. The electricity cost of a single training run can reach tens of millions of dollars at commercial electricity rates. And the training runs are getting larger, not smaller, as researchers continue to find that scaling produces capability improvements.
Inference — running queries against deployed models rather than training new ones — is lower intensity per query but adds up to enormous totals at the scale of hundreds of millions of daily users. Every ChatGPT conversation, every Copilot suggestion, every Claude query consumes electricity. The aggregate inference load across the major AI platforms is already comparable to significant industrial facilities.
Where the Power Comes From
The electricity powering AI infrastructure comes from wherever data centers can get it, which means the carbon intensity varies enormously by location. Data centers in the Pacific Northwest take advantage of abundant hydroelectric power and have relatively low carbon footprints. Data centers in coal-heavy regions of the Midwest or Southeast have much higher carbon intensity for the same compute load. The geography of data center construction is partly driven by electricity availability and cost, but not always by carbon intensity.
All three major hyperscalers — Microsoft, Google, and Amazon — have committed to powering their data centers with 100% renewable energy, with target dates ranging from 2025 to 2030. The commitments are real, but the accounting is complicated. Renewable energy certificates — the financial instruments used to claim renewable energy consumption — can be purchased without any direct connection between the electrons consumed and the electrons generated by renewable sources. Whether these commitments represent genuine decarbonization of AI infrastructure or sophisticated accounting is a legitimate question that the companies involved and their critics answer differently.
Nuclear energy has emerged as a preferred solution for AI companies that want reliable, dense, carbon-free power. Microsoft signed an agreement to restart a unit of Three Mile Island. Google signed a deal with Kairos Power for small modular nuclear reactors. Amazon purchased a data center campus adjacent to a nuclear plant. The carbon-free, always-on characteristics of nuclear power align well with the needs of AI infrastructure, and the AI boom has given the stalled nuclear industry a potentially important new customer segment.
The Construction Race
Building a modern hyperscale data center is a construction project of significant complexity and cost. A single large facility might cost two to three billion dollars, take eighteen to twenty-four months to construct from groundbreaking to operation, and require custom high-voltage electrical infrastructure, massive cooling systems, and specialized network connectivity. The lead times on critical components — large power transformers, cooling equipment, electrical switchgear — have stretched significantly as demand has exceeded manufacturing capacity.
The construction race is creating geographic concentrations of data center infrastructure in areas with available power, land, fiber connectivity, and favorable regulatory environments. Northern Virginia, Phoenix, Dallas, Chicago, and Amsterdam have emerged as global hyperscale hubs. The concentration is creating local infrastructure stress — utility companies in these areas are investing billions in grid upgrades to accommodate demand that was not anticipated in their multi-decade planning horizons.
Water is an equally significant constraint in many regions. Data centers cool their hardware using water — a large facility can consume millions of gallons per day in water-intensive cooling configurations. In water-stressed regions like the American Southwest, this is creating genuine conflicts between data center development and agricultural and municipal water needs. Several major data center projects have faced community opposition and regulatory scrutiny specifically over water use.
What Efficiency Progress Looks Like
The energy intensity of AI computing is not fixed. Significant progress has been made in the efficiency of both the hardware and the software that runs on it. Nvidia Hopper architecture delivers dramatically more AI compute per watt than previous generations. Model distillation techniques produce smaller models that achieve comparable performance with a fraction of the compute. Quantization reduces the precision of model weights in ways that significantly reduce memory and compute requirements with minimal performance degradation.
The troubling dynamic is that these efficiency improvements have historically been outpaced by the growth in demand for AI compute. Each generation of hardware is more efficient, but the models being trained on it are larger and the inference workloads running on it are growing faster than the efficiency gains. The net result, historically, has been increasing total energy consumption despite improving efficiency — the same dynamic that has characterized compute consumption for decades.
Whether that dynamic will continue, or whether the efficiency frontier will eventually catch up with demand, is one of the most important and most genuinely uncertain questions about the long-term physical footprint of the AI economy. The answer will be determined partly by the pace of technical progress in AI efficiency, partly by the pace of deployment in energy-intensive applications, and partly by decisions about what applications are worth the energy they require — decisions that the industry, regulators, and the public have barely begun to have in earnest.