The Thermostat Will Eat the Grid Before the GPU Does
 |
| Copyright: Sanjay Basu |
Everyone in tech is worried about the wrong electricity crisis.
Open any industry publication today and you will find breathless coverage of AI’s energy appetite. Data centers consuming the equivalent of Japan’s electricity by 2030. GPU clusters guzzling megawatts. The grid buckling under the weight of large language models. It makes for compelling reading, and it is not entirely wrong. But it is misleading in a way that borders on dangerous.
The biggest emerging source of electricity demand on Earth is not artificial intelligence. It is not cryptocurrency mining. It is not electric vehicles. It is air conditioning.
That sentence should stop you cold. Or rather, it should stop you warm, because warmth is the whole point.
The Numbers No One Wants to Talk About
The International Energy Agency’s World Energy Outlook 2025 lays this out with painful clarity. Global electricity demand is projected to grow by roughly 40% by 2035. The headlines seized on data centers, but the IEA’s own breakdown tells a different story. Income-driven air conditioning use alone will add approximately 330 gigawatts to global peak demand by 2035. Rising temperatures from climate change tack on another 170 gigawatts. That is 500 gigawatts of peak demand from cooling alone, a figure that dwarfs anything the data center industry is planning.
Let me put that in context. Total investment in data centers is expected to reach $580 billion in 2025. That number grabbed headlines because it surpassed the $540 billion being spent on global oil exploration. But the IEA projects that by 2030, the increase in electricity demand from air conditioning (651 terawatt-hours) will exceed the increase from data centers (530 terawatt-hours). Electric vehicles add another 838 terawatt-hours. Industrial growth adds 1,936 terawatt-hours. Data centers, for all their visibility, represent roughly 8% of total projected demand growth through the end of the decade.
Eight percent. Not eighty. Eight.
The fixation on AI energy consumption is a peculiar kind of myopia. It is driven partly by the fact that tech companies issue press releases and hold earnings calls where these numbers get discussed. Air conditioning manufacturers in Shenzhen and São Paulo do not command the same media attention. But the physics of thermal comfort does not care about your LinkedIn feed.
The Thermodynamics of Inequality
Here is where the story gets uncomfortable, and where it diverges sharply from the tidy narrative of data center energy management.
Roughly 3.5 billion people live in regions with high cooling demand. Only about 15% of them currently own an air conditioner. India’s annual room AC sales have exploded from around 1 million units in the early 2010s to over 11 million by 2023. China already has 569 million air conditioning units in operation. Southeast Asia, where fewer than 15% of households own an AC unit, represents an enormous untapped market. Daikin reported a 40% surge in sales in India during the first quarter of 2024 alone.
This is not optional consumption. This is survival infrastructure.
The European summer of 2024 killed an estimated 62,775 people from heat-related causes across 32 countries, according to research published in Nature Medicine. Italy alone accounted for more than 19,000 of those deaths. In the United States, heat-related mortality increased by more than 50% between 2000 and 2025. Extreme heat is now the leading cause of weather-related death in America, averaging 238 fatalities per year over three decades. And those are the reported numbers in countries with functioning public health systems. The unreported toll in South Asia and sub-Saharan Africa is almost certainly far worse.
When the IEA says cooling demand will reshape global electricity systems, it is not making a prediction about consumer preferences. It is making a prediction about human biology colliding with atmospheric physics. As global temperatures continue to climb, as 2024 became the hottest year in recorded history and average temperatures breached the 1.5-degree Celsius threshold above pre-industrial levels for the first time, the demand for mechanical cooling becomes not a luxury but a medical necessity.
This creates a feedback loop that is almost elegant in its cruelty. Burning fossil fuels raises temperatures. Higher temperatures demand more air conditioning. More air conditioning demands more electricity. More electricity generation, in a grid still dominated by fossil fuels, produces more greenhouse gas emissions. Which raises temperatures further.
The loop has a name in engineering. We call it a positive feedback mechanism. In climate science, it is one of many. But this particular loop has a feature that makes it especially pernicious. It is self-reinforcing across socioeconomic boundaries. As developing economies grow wealthier, their populations do two things simultaneously. They consume more energy per capita, raising emissions, and they purchase air conditioning, adding peak load to grids often dominated by coal. Oxford Economics projects middle-class households in growth economies to double from 354 million in 2024 to 687 million by 2034. Every one of those households represents a future air conditioning purchase.
Consider the math. The global air conditioning market was valued at roughly $150 billion in 2024 and is projected to reach somewhere between $250 and $340 billion by the early 2030s, growing at nearly 7% annually. Total global AC shipments hit 140.6 million units in 2024. Asia-Pacific accounts for 70% of those shipments. China alone manufactured 244.9 million AC units in 2023, and its urban households now average over 120 air conditioners per 100 homes. These are not projections. These are installed realities humming along right now, drawing power from grids that in many regions still burn coal as their primary generation source.
The IEA found that in India, each 1-degree Celsius increase in outdoor temperature in 2024 was associated with a 7 gigawatt increase in peak electricity demand. That sensitivity could rise to 12 gigawatts per degree by 2030 without significant efficiency improvements. During the early summer heatwaves of 2025, New York’s evening electricity peak was 90% above the off-season average. France, where AC ownership remains low, saw peaks 25% above normal. The difference between those two numbers is a measure of what happens when a population adopts air conditioning at scale.
The Data Center Distraction
I work in GPU cloud infrastructure. I spend my days thinking about power density, thermal management, and the logistics of feeding tens of thousands of accelerators with clean, reliable electricity. I am not dismissing the energy challenge that AI infrastructure presents. It is real, it is growing, and it demands serious engineering attention.
But perspective matters.
Data centers currently account for approximately 1.5% of global electricity consumption, or about 415 terawatt-hours annually. Under the IEA’s base case, that doubles to around 945 terawatt-hours by 2030, representing just under 3% of global electricity consumption. Gartner projects data center electricity consumption rising from 448 terawatt-hours in 2025 to 980 terawatt-hours by 2030, with AI-optimized servers growing nearly fivefold from 93 to 432 terawatt-hours.
Those are significant numbers. They are not civilization-altering numbers, at least not relative to the total energy picture. The challenge with data centers is geographic concentration, not aggregate magnitude. Nearly half of global data center electricity consumption is in the United States, 25% in China, and 15% in Europe. Within those countries, the load is further concentrated. Virginia’s data centers already consume 26% of the state’s electricity. In Dublin, the figure is 79%. In Ireland nationally, data centers account for 21% of electricity use.
This clustering creates real grid stress. It is a genuine infrastructure problem that I deal with professionally. But it is a fundamentally different kind of problem from the global, distributed, relentless growth of cooling demand. You can plan around data center clusters. You can site them strategically, negotiate power purchase agreements, and design purpose-built generation capacity. Oracle is designing facilities powered by small modular nuclear reactors. Amazon, Google, and Microsoft have collectively committed over $10 billion to nuclear energy partnerships. TerraPower broke ground on its Natrium reactor in Wyoming in 2024. These are engineering problems with engineering solutions, and the companies involved have the capital and technical capacity to pursue them.
Air conditioning demand, by contrast, is distributed across billions of individual decision-makers spread across every climate zone on the planet. There is no centralized procurement process for a family in Chennai buying their first window unit. There is no power purchase agreement for a construction worker in Phoenix who needs a cooled home to survive nighttime temperatures that no longer drop below 90 degrees Fahrenheit.
The asymmetry runs even deeper than procurement logistics. When a hyperscaler builds a new data center campus, it undergoes years of planning, environmental review, grid impact assessment, and power contracting. The load is predictable, constant, and professionally managed. When a million families in Uttar Pradesh buy their first air conditioners in a single monsoon season, the load arrives unplanned, peaks violently during the hottest hours of the hottest days, and stresses grid infrastructure that was never designed for it. One challenge is an engineering problem amenable to engineering solutions. The other is an emergent property of civilization trying to adapt to a warming world.
The Efficiency Gap That Could Save Us
The IEA’s most striking finding on cooling is not about the scale of demand. It is about the staggering inefficiency of the equipment being deployed.
Globally, the average new air conditioner sold today is only about half as efficient as the best models available on the market. And here is the part that should enrage every engineer who reads it: those more efficient models often cost the same. The IEA’s analysis across Southeast Asia and Latin America found that for identical upfront costs, consumers can purchase air conditioners with efficiency levels ranging from 3 watts of cooling per watt of electricity to more than 6 watts per watt. That is a factor-of-two difference in energy consumption for the same price. The gap is not technological. It is informational and regulatory.
A study on thermal comfort in Singapore found that participants reported feeling equally or more comfortable when air conditioners were set to 26 degrees Celsius and supplemented with a fan, compared to 24 degrees without a fan. That two-degree setpoint change, in a well-insulated building, reduces energy consumption by approximately 30%. The fan adds negligible additional draw.
These are not speculative solutions requiring breakthrough technology. They are available now, at current price points, with current infrastructure. The problem is adoption, not invention.
Inverter technology has already transformed the high end of the market. Inverter-based systems accounted for over 65% of residential units sold in Japan in 2024, and India saw a 35% year-over-year increase in inverter AC sales in early 2025. But in price-sensitive emerging markets, where most of the growth is happening, non-inverter units with significantly lower efficiency still dominate. These are the units that will be running for 15 to 20 years, locking in energy consumption patterns for a generation.
The Energy System That Must Be Built
The good news is that the broader energy transition is moving faster than most people realize. In the first half of 2025, solar and wind power overtook coal as the leading source of electricity generation globally for the first time in recorded history. Solar generation alone surged 31%, covering 83% of the rise in global electricity demand. Renewables reached 34.3% of global electricity production while coal fell to 33.1%. Four countries now derive more than a quarter of their electricity from solar, and 29 countries have surpassed 10%.
This is not a marginal trend. This is a structural transformation of global energy infrastructure happening in real time.
But it is uneven. In the United States, electricity demand grew faster than renewable output in the first half of 2025, causing coal generation to rise by 17%. The European Union faced similar challenges when weather-related dips in wind and hydro output forced increased reliance on gas and coal. China, by contrast, deployed so much solar and wind that clean power covered essentially all of its demand growth, and Chinese power sector emissions actually fell year-over-year for the first time since at least 2015.
The nuclear renaissance adds another dimension. After two decades of controversy and stagnation, the global nuclear fleet could expand by a third over the next decade. The IEA projects nuclear energy providing an additional 190 terawatt-hours by 2035 specifically to meet data center demand. Small modular reactors are moving from paper to permitting, with NuScale holding the only NRC-certified design and TerraPower targeting a 2030 operational date for its Natrium reactor in Wyoming. Tech companies are signing power purchase agreements for nuclear output years before the reactors exist, essentially de-risking a technology that has historically struggled with private capital.
But nuclear timelines are measured in decades, not quarters. And the cooling demand crisis is measured in summers.
The Real Race
Global greenhouse gas emissions hit another record in 2025, reaching 60.63 billion tonnes of CO2 equivalent. Fossil fuel CO2 emissions set yet another high, driven by increases across coal, oil, and natural gas. The remaining carbon budget for limiting warming to 1.5 degrees Celsius stands at approximately 170 billion tonnes, equivalent to roughly four years at current emission rates.
That budget will be consumed whether we are powering GPUs or air conditioners. But the political and media attention devoted to the two categories is wildly asymmetric. AI energy use gets congressional hearings and newspaper investigations. Air conditioning energy use gets a paragraph in an IEA report that most policymakers will never read.
This asymmetry has consequences. When we frame the energy challenge as primarily an AI problem, we implicitly suggest that the solution lies with a handful of technology companies and their infrastructure decisions. That framing is comforting because it concentrates responsibility. It is also wrong because it ignores the vastly larger, more distributed, and more urgent demand growth coming from human beings trying not to die in the heat.
The engineers and policy makers who will navigate the next decade of energy transition need to hold two truths simultaneously. First, that AI infrastructure demands unprecedented power density and requires innovative generation and grid solutions. Second, that air conditioning represents a far larger aggregate electricity challenge, one with profound implications for equity, public health, and climate feedback loops.
I have spent my career building infrastructure for the digital economy. I believe in the transformative potential of AI computing. But I also believe in intellectual honesty, and intellectual honesty demands that we acknowledge the thermostat, not the tensor core, as the defining energy challenge of the next two decades.
The question is not whether we can power our data centers. We can, and we will. The big cloud providers have the resources, the engineering talent, and the financial incentive to solve their own energy problems.
The question is whether we can cool a planet of 8 billion people, billions of whom are only now gaining access to mechanical cooling for the first time, without accelerating the very warming that makes that cooling necessary.
There is a philosophical dimension to this that the tech industry would prefer to ignore. We have spent the last two years debating whether training a large language model constitutes an acceptable use of energy. We have not spent equivalent time debating whether a grandmother in Athens dying because she cannot afford to run her air conditioner constitutes an acceptable failure of energy policy. The first conversation happens at conferences and in boardrooms. The second happens in emergency rooms and morgues.
The energy transition is not a single story with a single protagonist. It is a thousand stories happening simultaneously across every latitude and income bracket on the planet. The AI energy story is important, but it is the one we tell because it is legible, concentrated, and populated by recognizable actors with investor relations departments. The cooling energy story is the one we avoid because it is diffuse, deeply entangled with inequality, and has no clean resolution.
But the grid does not care about narrative convenience. It cares about load. And the load is coming from the thermostat, not the tensor core.
That is the problem that should keep us up at night. And ironically, on the hottest nights, it already does.
Comments
Post a Comment