“Every prompt has a price. Not in dollars — though that too — but in watts. The question the AI industry has not been asked loudly enough is: where does the power come from, who pays for it, and what does it cost the planet to run an intelligence that does not sleep?”
Neal Lloyd · Inside The Machine, Day 17In 2023, the International Energy Agency published a report that received approximately one tenth of the attention it deserved. The finding was this: global data centre electricity consumption was approximately 200 to 250 terawatt-hours per year, or roughly 1% of global electricity demand. By 2026, driven almost entirely by the rapid scaling of AI inference infrastructure, that figure has reached somewhere between 400 and 500 terawatt-hours — a doubling in three years. The IEA’s 2026 projections put AI data centre consumption at between 3% and 8% of global electricity by 2030. The range is wide because the uncertainty in AI scaling trajectories is genuinely large. The direction is not uncertain. The AI industry’s energy appetite is growing faster than any energy infrastructure in history, in an era when grids are already under stress from electrification, climate events, and decades of underinvestment. Something has to give.
What AI Actually Costs in Energy, From Training to Every Single Prompt
Training costs. Training a large frontier model is an energy event of significant scale. GPT-3, trained in 2020, consumed approximately 1,287 megawatt-hours of electricity — equivalent to the annual energy consumption of roughly 120 US households. GPT-4’s training cost has not been officially disclosed but is estimated at 10 to 50 times larger. For context: a 2019 University of Massachusetts Amherst study found that training a single large NLP model from scratch emits approximately 626,000 pounds of carbon dioxide — roughly five times the lifetime carbon emissions of an average American car, including its manufacturing. These are not abstract figures. They are the energy cost of the R&D that produced the models you use.
Inference costs. Training is a one-time cost. Inference — the energy consumed every time a model answers a question, generates an image, or processes a request — is the recurring cost that scales with user volume. ChatGPT serving 900 million weekly active users is a continuous, massive inference operation. Goldman Sachs estimated in 2024 that a ChatGPT query uses approximately 10 times the electricity of a Google search. Google processes approximately 8.5 billion searches per day. ChatGPT processes approximately 100 million queries per day. As AI assistants become more capable, more integrated into workflows, and more widely used, the inference energy cost scales with adoption.
Water costs. Data centres cool their servers with water as well as electricity. Microsoft disclosed that its global data centres consumed 1.7 billion gallons of water in 2022 — a figure that increased substantially in 2023 and 2024 as AI workloads expanded. Google’s water consumption for data centres exceeded 5 billion gallons in 2023. The water withdrawal occurs disproportionately in regions already under water stress, because land and power are cheaper there. The intersection of AI infrastructure expansion and water scarcity is a genuine environmental and geopolitical issue that receives almost no coverage relative to its scale.
1 ChatGPT query: approximately 10x the electricity of a Google search. Training GPT-3: 1,287 MWh, equivalent to 120 US households for a year. Training a frontier model: estimated 626,000 lbs CO2 equivalent — five car lifetimes. Global AI data centre consumption 2023: 200-250 TWh (1% of global electricity). Global AI data centre consumption 2026: 400-500 TWh (2% of global electricity). IEA projection 2030: 3-8% of global electricity. Microsoft 2022 data centre water use: 1.7 billion gallons. $500 billion: US federal AI infrastructure commitment (Stargate). None of the above figures include the energy cost of training the models that will exist in 2028.
The Infrastructure Was Not Built for This and Cannot Be Fixed Quickly
The United States electricity grid is old. The average age of power transformers in the US is approximately 40 years — most of them were built when the heaviest demand came from industrial manufacturing, not from server halls running 24 hours a day. The grid was designed for demand that peaks during business hours and drops at night; AI data centres run at near-constant load around the clock, a demand profile that stresses infrastructure differently and more severely than residential or commercial consumption.
Building new grid capacity is slow. The interconnection queue — the process by which new power generation projects apply to connect to the grid — has a backlog of over 2,000 gigawatts of proposed capacity in the United States, the majority of it renewable. The average wait time from application to connection is over five years. The permitting, siting, and construction of transmission infrastructure required to connect new generation to demand centres adds years more. The AI industry’s energy demand is growing on a timeline of months. The grid’s capacity to respond is growing on a timeline of years. The gap between those two timelines is where the crisis lives.
The consequence is already visible in specific geographies. Northern Virginia — the world’s largest data centre market, home to approximately 70% of the world’s internet traffic at any given moment — has reached the limits of available grid capacity in several substations. Dominion Energy, the regional utility, has warned that new data centre connections may be delayed until new generation capacity comes online. Microsoft, Google, Amazon, and Meta have all been managing the power constraint through a combination of long-term power purchase agreements, co-location deals with utilities, and — increasingly — direct investment in power generation.
The AI industry’s energy appetite is growing on a timeline of months. The grid’s capacity to respond grows on a timeline of years. The companies best positioned to navigate that gap are the ones large enough to buy their own power plants. The ones that are not large enough are going to find that “available power” is the new “available GPU” — the constraint that determines who can operate at scale and who cannot.Neal Lloyd · Inside The Machine, Day 17
Nuclear, Renewables, Efficiency — and the Uncomfortable Trade-offs in Each
Nuclear. Microsoft signed a deal to restart the Three Mile Island nuclear plant in Pennsylvania in 2023 — a facility that had been shut down for 20 years — specifically to provide dedicated carbon-free power for its data centres. Google has committed to purchasing power from multiple small modular reactor (SMR) developers. Amazon has invested in SMR development through its Climate Pledge Fund. The nuclear renaissance in US energy policy has been substantially driven by AI company demand, and the bipartisan political support for nuclear in Washington is partly a reflection of this. Nuclear is carbon-free, runs at constant load — ideal for data centres — and does not depend on weather. It is also slow to build, enormously capital intensive, and generates waste that requires management over geological time scales.
Renewable energy. All major AI companies have made commitments to match their energy consumption with renewable sources. The commitments are real and the renewable procurement is large and growing. The challenge is that renewable energy — solar and wind — is intermittent. The sun does not always shine; the wind does not always blow. Data centres cannot be intermittent. The mismatch between renewable supply patterns and data centre demand profiles requires either grid-scale storage (still expensive and limited in duration), geographic diversification of data centre locations, or continued reliance on gas or nuclear for baseload power. The “100% renewable” commitments are, on close inspection, typically “100% matched with renewable certificates” — an accounting framework that does not necessarily mean the electrons powering the servers came from renewables at the moment of consumption.
Efficiency. The most immediate lever for reducing AI energy consumption is improving the efficiency of the models and the hardware running them. The trend is encouraging: the compute efficiency of AI models has improved dramatically since 2012, following a trajectory roughly described as halving the compute required for a given capability level every 8 to 16 months. DeepSeek’s R1 was a dramatic demonstration of what algorithmic efficiency can achieve — frontier-level reasoning at a fraction of the previously assumed compute cost. Google’s MAI M5 announcement this week cited a 40% reduction in thermal throttling and a 22% reduction in energy consumption relative to its predecessor. The efficiency gains are real. The question is whether they are running faster than the demand growth they are partly enabling — the Jevons paradox applied to AI compute.
The Jevons paradox, first observed in the 19th century coal industry, states that increases in the efficiency of resource use tend to increase rather than decrease total resource consumption, because efficiency reduces cost, lower costs drive higher demand, and higher demand outpaces the efficiency gains. Applied to AI compute: as models become more efficient to run, the cost per query falls; as cost falls, more queries are run; as more queries are run across more applications, total energy consumption rises even as energy per query falls. This is the efficiency trap. It does not mean efficiency improvements are not worthwhile. It means they are not sufficient on their own to solve the energy problem.
The Externalities, the Trade-offs, and the Decisions That Are Being Made Without a Vote
Who bears the cost? The energy consumed by AI data centres is partly borne by the companies operating them, through their power purchase agreements and energy bills. But grid infrastructure costs — the upgrades, new transmission lines, and generation capacity required to meet AI demand — are typically socialised across all ratepayers. When Dominion Energy builds new substations to serve Microsoft’s data centres in Northern Virginia, the cost is distributed across every electricity customer in the region. The companies generating the demand and the revenue from AI are not necessarily the ones paying for the infrastructure their demand requires. This is a political economy question with significant distributional implications that is receiving almost no democratic deliberation.
What are we trading off? The energy consumed by AI is energy not available for other uses, including the electrification of transport and heating that decarbonisation strategies depend on. In a constrained grid environment, AI data centre demand competes with EV charging infrastructure, heat pump adoption, and industrial electrification for the same finite capacity. The energy transition and the AI transition are occurring simultaneously and, in grid-constrained regions, they are competing with each other. The prioritisation of AI data centre demand in grid planning has not been the subject of explicit public policy debate in most jurisdictions.
What does it mean that intelligence requires energy? This is the genuinely novel philosophical dimension of the AI energy question. Human intelligence is embedded in biological systems that run on approximately 20 watts of continuous power — the metabolic cost of maintaining the human brain. Artificial intelligence at frontier scale runs on megawatts per data centre, with tens of thousands of data centres worldwide. We have built a form of intelligence whose energy appetite scales without the biological constraints that have historically bounded the cost of cognition. That is an extraordinary fact about the technology we have created, and its long-term implications for energy systems, geopolitics, and the material basis of intelligence itself deserve considerably more serious attention than they currently receive.
Human intelligence runs on 20 watts. Frontier AI runs on megawatts. We have built a form of cognition whose energy appetite is not bounded by biology — only by infrastructure, capital, and political will. The implications of that fact are still being understood. What is already clear is that the planet’s energy systems were not designed to host it at the scale we are building toward.Neal Lloyd · Inside The Machine, Day 17
Inside The Machine, Day 17 · June 2026
Neal Lloyd writes about technology, human adaptation, and the uncomfortable questions nobody wants to answer at dinner. Inside The Machine is his ongoing daily series on AI.
- Day 01What Is This Thing?
- Day 02Survive the Machine
- Day 03The Great Debate
- Day 04Who Gets Hurt?
- Day 05Who’s In Charge?
- Day 06The Industries That Win
- Day 07The Human Edge
- Day 08The Creativity Question
- Day 09Does AI Feel Anything?
- Day 10The Data Problem
- Day 11The Trust Question
- Day 12The Accountability Gap
- Day 13The Rewired Brain
- Day 14Open vs Closed
- Day 15The New Cold War
- Day 16Why AI Lies With Confidence
- Day 17AI Is Eating the Power GridYou are here
