The AI industry's insatiable demand for computing power is straining global energy systems. Data centers that train and run large language models now consume as much electricity as small countries, forcing utility providers and tech giants to confront hard limits on power generation and grid capacity. This surge threatens corporate net-zero pledges and raises urgent questions about how to fuel the AI boom without accelerating climate change.
Energy consumption outpaces efficiency gains
Each new generation of AI accelerators delivers more performance per watt, but the total number of chips deployed is growing faster than efficiency improvements. A single query to a generative AI model can require ten times the electricity of a standard web search. The International Energy Agency projects that data center electricity consumption could double by 2026, driven largely by AI workloads. Tech companies that once highlighted their renewable energy purchases now face scrutiny over whether those purchases actually displace fossil fuel generation on the same grids.
Training frontier models demands staggering amounts of power. A large-scale training run can consume tens of gigawatt-hours - equivalent to the annual electricity use of thousands of households. As models grow, so does the energy cost of keeping them updated and serving millions of users.
Grid infrastructure under pressure
Data center construction is concentrating in regions like Northern Virginia, which already handles over a third of global online traffic. Local utilities are struggling to build transmission lines fast enough. Some have delayed planned retirements of coal and gas plants to maintain reliability. Dominion Energy, for example, has warned that demand from data centers could exceed its ability to add new generation in its Virginia service territory.
This bottleneck is reshaping the priorities within AI for IT & Development, as energy efficiency becomes a design constraint. Hyperscalers are responding with direct investments in nuclear power, including small modular reactors, and long-term power purchase agreements for geothermal and advanced nuclear. Microsoft recently signed a deal to restart a reactor at Three Mile Island. These moves signal that the AI industry can no longer treat electricity as an abundant, cheap resource.
The hidden water cost
Cooling the servers that run AI models requires enormous volumes of water. A mid-sized data center can consume hundreds of thousands of gallons per day. In arid regions like Arizona and Chile, where several mega-projects are planned, this demand competes directly with agriculture and drinking water supplies. Researchers have calculated that training GPT-3 in Microsoft's U.S. data centers may have evaporated 700,000 liters of freshwater.
The environmental toll is becoming a central topic in AI for Science & Research, where scientists are developing methods to reduce the carbon and water footprint of training runs. Techniques like sparse training, model distillation, and better data center siting are gaining urgency.
Why this matters for IT, development, and research professionals
For those building and deploying AI systems, energy constraints are no longer someone else's problem. Data center capacity may limit how quickly teams can scale models. Cost per query will increasingly reflect real-time energy pricing. IT architects and developers who prioritize energy-aware system design-selecting efficient models, optimizing inference pipelines, and choosing regions with clean power-will hold a clear advantage. Researchers have a direct role in creating the next generation of algorithms that maintain accuracy while slashing resource use. The limits of growth are real, and they will define the next phase of AI deployment.
Your membership also unlocks:
