The Lights That Never Go Out
Somewhere in the Virginia suburbs, inside a building with no windows and more security cameras than a casino, thousands of servers are humming at this very moment. They are processing your email drafts, generating someone’s marketing copy, answering a teenager’s homework questions, and training the next generation of models that will do all of this faster and more thoroughly. The air-conditioning systems keeping those servers from melting are the size of small houses. The electricity meters are spinning so fast that utility engineers have started using language once reserved for aluminum smelters and steel mills.
This is the unglamorous reality behind the polished interfaces of ChatGPT, Google Gemini, and their growing list of competitors. Artificial intelligence, in its current industrial form, is not a weightless, ethereal technology living in the cloud. It is a physical thing—a thing that requires staggering quantities of electricity, water, land, and hardware. And as the industry races toward ever-larger models and ever-broader deployment, the question of how much energy AI actually consumes—and who bears the cost—has become one of the most consequential and under-examined stories in technology.
The Numbers That Should Make You Pause
Start with a single search query. A traditional Google search consumes roughly 0.3 watt-hours of electricity. A query processed through a large language model—the kind powering ChatGPT or Gemini—uses approximately 10 times that amount, according to estimates from Goldman Sachs Research published in 2024. That may sound trivial in isolation. Multiply it by hundreds of millions of queries per day, and it becomes a different conversation entirely.
OpenAI’s ChatGPT reportedly processes somewhere between 10 million and 100 million queries daily, depending on the source and time period. Google, which has integrated AI deeply into its search product, handles over 8 billion searches per day. Even if only a fraction of those involve generative AI processing, the cumulative energy draw is extraordinary.
Goldman Sachs estimated in mid-2024 that data center power demand could increase by 160 percent by 2030, driven primarily by AI workloads. The International Energy Agency (IEA), in its 2024 report on electricity, projected that global data center electricity consumption could reach between 650 and 1,000 terawatt-hours annually by 2026—roughly equivalent to the entire electricity consumption of Japan. That’s up from approximately 460 terawatt-hours in 2022.
Training a large AI model is itself an extraordinary event in energy terms. A 2019 study from the University of Massachusetts Amherst found that training a single large natural language processing model could emit as much carbon as five cars over their entire lifetimes. That study predates the current generation of models by several years; GPT-4, Claude 3, and Gemini Ultra are vastly larger. Google’s own environmental report acknowledged that training its Gemini Ultra model produced substantial emissions, though the company declined to provide precise figures—a pattern of opacity that critics find troubling.
Microsoft, in its 2024 sustainability report, disclosed that its carbon emissions had increased by nearly 30 percent since 2020, despite aggressive commitments to become carbon negative by 2030. The company attributed a significant portion of that increase to the energy demands of expanding its AI infrastructure. It was a rare moment of corporate candor in an industry that often prefers to lead with its renewable energy pledges rather than its raw consumption figures.
The Architecture of Appetite
To understand why AI is so energy-hungry, it helps to understand what actually happens inside a data center when a model runs.
Modern AI systems—particularly the transformer-based large language models that underpin ChatGPT, Claude, and their peers—require specialized chips called GPUs (graphics processing units) or purpose-built accelerators like Google’s TPUs (tensor processing units). These chips are extraordinarily powerful but also extraordinarily power-hungry. A single NVIDIA H100 GPU, the workhorse of contemporary AI training, draws between 300 and 700 watts of power. A typical AI training cluster might contain tens of thousands of these chips running continuously, sometimes for months, to train a single frontier model.
Then there is inference—the ongoing process of actually responding to user queries after training is complete. Inference is less energy-intensive per operation than training but far more continuous. Training happens once (or occasionally); inference happens billions of times per day, around the clock, across a global network of data centers that must remain operational without interruption.
These facilities also consume enormous quantities of water. Data centers use water evaporation as a primary cooling mechanism. Microsoft’s data centers consumed 6.4 million cubic meters of water in 2022—a 34 percent increase over the previous year. Google’s water consumption for data centers exceeded 5.6 billion gallons in 2022. In regions already experiencing water stress, including parts of the American Southwest and parts of Latin America where major data centers are being built, this creates a secondary resource conflict that rarely makes technology headlines.
The physical footprint is expanding rapidly. Amazon Web Services, Microsoft Azure, and Google Cloud are collectively spending hundreds of billions of dollars on new data center construction. Meta announced plans to build what it described as a “city-scale” data center complex in Louisiana. Microsoft has signed agreements with nuclear energy companies, including a controversial deal to restart a reactor at Three Mile Island, the site of the United States’ most serious commercial nuclear accident. The symbolism was not lost on critics.
The Dirty Secret Behind the Green Pledges
Every major technology company—Microsoft, Google, Amazon, Meta—has made sweeping commitments to power their operations with 100 percent renewable energy. These pledges are real in their ambition but complicated in their execution, and the gap between the marketing and the reality matters enormously for the climate calculus.
The mechanism most companies use is purchasing Renewable Energy Certificates (RECs) or Power Purchase Agreements (PPAs), which allow them to claim credit for renewable energy generated somewhere on the grid, even if the electrons actually flowing into their data centers come from gas plants or coal. Critics, including many energy policy experts, argue this accounting method creates a misleading picture. “Matching your consumption with certificates doesn’t mean you’re actually running on renewable energy in real time,” explains Dr. Sasha Luccioni, an AI researcher at Hugging Face who has published extensively on AI’s environmental footprint. “The grid physics don’t care about your certificate portfolio.”
This matters most in the context of temporal matching. Data centers need power 24 hours a day, 7 days a week. Solar panels generate power only during daylight hours. Wind turbines are variable. Without massive battery storage—which remains prohibitively expensive at scale—a data center claiming to run on 100 percent renewables is almost certainly drawing from fossil fuel generation during nighttime hours or periods of low wind.
Google has made more aggressive commitments than most, pledging to match its energy consumption with carbon-free energy on an hourly basis by 2030. This is a meaningfully more rigorous standard than annual matching, and energy analysts have credited the company for raising the bar. But even Google’s current operations fall well short of that goal, and the company’s own disclosure that Gemini Ultra’s training produced substantial emissions suggests the gap remains significant.
Meanwhile, the geographic concentration of data center construction is creating acute local problems. The Commonwealth of Virginia, home to more data center capacity than any jurisdiction on Earth, is facing grid reliability challenges directly linked to hyperscale data center growth. Dominion Energy, the primary utility serving the region, has warned that it may need to delay the retirement of coal plants to meet demand. In Ireland, data centers now consume approximately 21 percent of the country’s total electricity—a figure that has prompted the Irish government to impose a moratorium on new data center construction in Dublin.
Who Actually Pays
The cost of AI’s energy consumption is distributed in ways that are frequently invisible to the people generating the demand.
At the most direct level, technology companies pay enormous electricity bills. Amazon, Google, and Microsoft collectively spend billions of dollars annually on energy. These costs are passed on to enterprise customers through cloud computing fees, and to consumers through subscription prices for AI products. A ChatGPT Plus subscription costs $20 per month. Analysts estimate that OpenAI loses money on many AI interactions at current pricing—a situation that has led some observers to wonder whether AI services are effectively subsidized, with the true cost of energy deferred or absorbed by investors.
But the deeper costs are societal. When data centers drive up electricity demand faster than clean generation can be added, utilities respond by burning more fossil fuels. The resulting carbon emissions are a cost borne by the entire planet, with the most severe consequences falling on communities with the least connection to the technology creating the demand. Climate change does not respect the clean, seamless interfaces of AI products.
There are also hyperlocal costs. Communities where data centers are built often receive tax incentives in exchange for economic development promises—jobs, tax revenue, infrastructure investment. The jobs generated by large data centers are frequently fewer than advertised, as these facilities are highly automated. The infrastructure strain—on roads, on water systems, on the power grid—can be substantial. And the tax breaks, which sometimes run to tens or hundreds of millions of dollars, represent public subsidies for private infrastructure.
Patrick Toomey, a former U.S. Senator and now an advocate for competitive energy markets, has argued that the hidden subsidies and grid externalities of data center expansion represent a significant market distortion. “When a facility pays below-market rates for electricity and doesn’t bear the full cost of grid upgrades, those costs get socialized,” he noted in a 2024 policy forum. “Residential ratepayers end up subsidizing AI infrastructure they may never directly use.”
The workers in the communities surrounding these facilities often have a complicated relationship with them—grateful for the economic activity, concerned about water and power competition, and frequently left out of the conversations that led to their arrival.
The Case for Nuance—and the Limits of Efficiency
It would be incomplete to tell this story without accounting for the counterarguments, and some of them are genuinely compelling.
AI, its proponents argue, can be a powerful tool for reducing energy consumption across the broader economy. Google’s DeepMind has demonstrated that AI-driven cooling optimization reduced the energy used for cooling its data centers by 40 percent. AI is being applied to optimize electrical grids, improve weather forecasting for renewable energy integration, accelerate materials discovery for better batteries, and reduce waste in industrial processes. Microsoft’s AI for Earth program funds hundreds of projects applying machine learning to climate and environmental challenges.
There is also the question of substitution effects. If AI-powered tools reduce the need for business travel, commutes, or physical infrastructure, the net energy balance might look different. If AI accelerates the development of new clean energy technologies—better solar panels, cheaper fusion, more efficient electrolyzers—the return on investment in AI energy consumption could be positive.
These arguments are not without merit. But energy researchers caution against letting the promise of future efficiency obscure the reality of present consumption. “We’re making real emissions now in exchange for hypothetical savings later,” says Dr. Luccioni. “That tradeoff deserves honest scrutiny, not just optimistic framing.”
There is also a phenomenon called the Jevons Paradox—the historical pattern in which increased efficiency of resource use leads to increased total consumption, as the lower cost encourages greater adoption. Energy-efficient AI chips that allow more computation per watt may simply enable larger models and more queries, resulting in higher total energy use. The history of computing offers little comfort on this point: efficiency gains have consistently been outpaced by demand growth.
The Reckoning Ahead
The next several years will likely bring this tension to a head. The IEA, major utilities, and grid operators across the United States and Europe are already grappling with demand forecasts that strain existing infrastructure. PJM Interconnection, the grid operator managing the largest competitive electricity market in the United States, has warned of capacity shortfalls within the decade. These aren’t abstract concerns—they translate into decisions about which power plants get built and which get retired.
The technology industry has, to its credit, become one of the largest purchasers of renewable energy globally. Its capital has helped drive down the cost of solar and wind and financed significant new generation capacity. Microsoft’s nuclear deal, controversial as it is symbolically, reflects a genuine recognition that carbon-free baseload power is essential to meeting AI’s round-the-clock demand with clean electrons.
But the pace of renewable build-out is not matching the pace of AI-driven demand growth. And in the gap, fossil fuels remain the marginal supplier.
What seems increasingly clear is that the current model—in which AI companies build demand aggressively, make ambitious long-range commitments, and rely on market forces and favorable accounting to manage the reputational risk—is not adequate to the scale of the challenge. More rigorous disclosure requirements, so that companies must account for their actual real-time energy mix rather than annual certificate matching, would be a meaningful step. So would honest pricing that internalizes the true social cost of AI energy consumption, rather than passing it on to ratepayers, communities, and the atmosphere.
The intelligence being built inside those windowless buildings in Virginia is genuinely impressive. It is writing code, diagnosing diseases, tutoring students, and accelerating scientific research at a pace that would have seemed implausible a decade ago. It is also drawing power at a rate that demands—at the very least—the same honest accounting we would ask of any other major industrial enterprise.
The lights inside those data centers never go out. The question worth asking, urgently, is what we are burning to keep them on.