Every time a new AI data center goes up, the coverage reads like an indictment. Look how much water it drinks. Look how much power it pulls off the grid. The framing treats energy consumption as evidence something has gone wrong. I think the physics says the opposite, and once you follow the physics instead of the outrage, the whole story flips.
Energy is the one rule you can’t break
Start with the Kardashev scale, the old thought experiment for ranking a civilization by how much energy it can put to use. A Type I civilization harnesses the energy hitting its planet, Type II the output of its star, Type III the output of its galaxy. The scale was never really about spaceships or gadgets. It’s a statement about the one constraint that governs everything else: energy cannot be created or destroyed, only converted from one form to another. That’s the one rule the universe never lets anyone break, and every other capability is downstream of how much usable energy you can move around.
A material is strong because the energy required to tear its bonds apart isn’t available in whatever is hitting it. A rocket needs fuel because moving mass at speed is an energetics problem before it’s an engineering one. If you ever meet a civilization doing interstellar flight, or something out of Star Trek involving matter-antimatter annihilation, the first question isn’t how they built the ship. It’s where they got that much energy and how they contained it. No amount of cleverness substitutes for the energy budget. The Kardashev scale is a blueprint: you don’t get to be a technological civilization without solving energy production first.
Information is an energy state
That framing is why I don’t read AI’s power bill as a scandal. Computation, at bottom, is the movement of energy. A transistor switching on and off is electrons moving, and moving them is never perfectly efficient: some of that energy becomes heat, some becomes wear on the material as particles get knocked loose. Swap electrons for photons in fiber optics and nothing changes. You’re still spending energy to hold a configuration that means something. Information is the configuration of an energy state. When you query a system, what you’re actually asking for is the energy state it’s holding. Storing information means creating that state, and creating it is never lossless. Energy goes in, some fraction survives as the information you wanted, and the rest leaves as waste heat.
Even the water angle is an energy question wearing a costume. A data center doesn’t use water because computation needs water. It uses water because heat has to go somewhere, and moving heat away from a system that would otherwise cook itself is the actual job. Every refrigerator in every kitchen already does this without water, by pressurizing and depressurizing a gas and using the latent heat of evaporation to carry heat away, on nothing but electricity. Refrigeration isn’t new technology. Data centers could run closed-loop the same way. Most run on evaporative cooling instead because of cost and legacy design, not because physics demands it. Keep that distinction straight and a lot of the panic evaporates along with it.
So when a data center draws enormous power to run an AI workload, that isn’t AI breaking some unwritten rule. It’s a technology doing what all technology does: buying capability with energy. It helps to know the actual size of that bill. The IEA puts all the world’s data centers at about 1.5% of global electricity in 2024, on track to roughly double their share by 2030 as AI scales. That is real growth worth planning for, and it is nowhere near the civilization-ending number the coverage implies. The real debate isn’t about AI. It’s about how the electrons get generated in the first place, and we’ve been having that argument since long before anyone typed a prompt into ChatGPT.
We’ve had this fight before
The “not in my backyard” reaction to AI data centers isn’t a phenomenon AI invented. Communities were debating the power, water, and land demands of large data centers before generative AI arrived. Netflix alone runs about 15% of global internet traffic, and on-demand video is more than half of all downstream data. Streaming a show to your living room carries its own client-side and server-side energy bill. I don’t know if that’s bigger or smaller than AI’s draw. I know it isn’t nothing, and it arrived slowly enough for our infrastructure and culture to absorb it before most people noticed the scale.
AI didn’t invent the tension. It arrived fast enough and visibly enough that the tension shows up all at once instead of spread across a decade, a problem of scale colliding with infrastructure and culture that were never built for it. The same collision shows up behind the obesity epidemic.
We keep rejecting dense energy, then act surprised
Stay with that comparison, because it isn’t a throwaway line. Our bodies evolved to crave sugar, fat, and protein, not fiber and not vitamin C. Nobody’s monkey brain lights up at the sight of leafy greens. We crave those specific things because for most of human history, the sweetest thing in the forest was also the thing your body needed, wrapped in nutrition that was hard to extract in bulk any other way. Then food technology got good enough to isolate the sugar and the fat from everything else and hand it over in unlimited quantities, and our biology, which never evolved a mechanism for craving fiber or vitamin C, walked straight into the trap. The obesity epidemic isn’t a failure of willpower. It’s a mismatch between old biology and new technological capability.
We’re doing the same thing with energy, except the mismatch is cultural instead of biological. We’ve been happy digging fossil fuels out of the ground and burning coal and petroleum for the bulk of our energy needs, but the moment someone proposes a dense, scalable alternative, we find a reason to say no. Nuclear power could solve a meaningful share of these problems today, and most people are still against it on principle, cooling water aside. Solar and wind farms at real scale get vetoed over cost and over the noise and visual footprint on land people want to keep looking untouched. We want the energy without looking at where it comes from or carrying the risk of producing it densely, so we default back to the sources we’ve already normalized and act shocked when a new technology needs more power than the infrastructure we never upgraded can deliver.
Electrification is the fix, and it’s a decoupling trick
Here’s the part I think gets missed: the fix isn’t picking the “right” energy source and forcing everyone onto it. The fix is standardizing on electricity as the one universal currency and letting the source stay flexible.
Picture a Dyson sphere, a structure that captures a meaningful share of a star’s total output. Earth intercepts only a vanishingly small sliver of the sun’s total output. A structure capturing even a fraction of the whole would dwarf everything the planet now receives, and all of it converts straight to electricity. The catch is that electricity alone doesn’t get a plane off the ground. You’d still need to convert it into a fuel you can carry onboard, an extra engineering step and an extra point of failure. But if planes and cars and trains already ran on electricity directly, that conversion problem disappears.
That’s the case for electrification. Once your car runs on electricity, you don’t care whether that electricity came from coal, a nuclear plant, a satellite beaming down microwaves, or something nobody has invented yet. The consumer-facing engineering is already done. The only remaining question is how to produce electricity better, and a small number of centralized, regulated power producers can answer that instead of billions of individual people changing what they drive. Getting a corporation to change what it does, through tax policy and regulation, is far more tractable than getting every household on a street to adopt the same standard. Electrification takes an energy-source question that fragments into oil, gas, and a dozen competing industries, and narrows it to one target: cheaper, cleaner electricity. Cars, homes, and cooking all benefit automatically, with nobody having to change what they drive or cook with.
Carbon capture is the same story from the other direction. We could pull CO2 out of the air, compress it into synthetic diamond, and drop it to the bottom of the Mariana Trench, inert, for millions of years. Diamond is, as they say, forever. Take that as an illustration of the principle rather than a shovel-ready plan; even so, it would measurably reduce atmospheric carbon. It also takes real energy to capture, compress, and form that carbon before you bury it. Treat energy production as the priority instead of the afterthought, and climate change stops being an unsolvable problem and starts being an engineering budget line.
AI is often the more energy-efficient choice, not less
Which brings me back to where I started, because I think the framing has the sign backward. People treat AI’s energy consumption as proof of waste. In plenty of real cases it’s the opposite.
Start with the unit that gets quoted most. The scary version, that one AI query burns ten times what a web search does, came from a 2023 estimate that has not held up. More recent work from Epoch AI puts a typical ChatGPT query near 0.3 watt-hours, comparable to a web search and less than a laptop pulls in a couple of minutes. The per-query panic ran on worst-case assumptions that better models and hardware have already erased.
Then take the Human Genome Project: about $2.7 billion over thirteen years of data collection, curation, and interpretation, an enormous amount of skilled human labor, and just as much energy, to resolve one genome. If that project had today’s AI tools, I’d bet the same result comes in for less time and less total energy, not more, because so much of the original cost was brute-force human and instrument time standing in for a capability nobody had built yet.
We don’t have to guess, because something close to it already happened. Protein structure prediction was a fifty-year grand challenge, worked through x-ray crystallography and decades of accumulated PhD labor, one structure at a time. In 2020, AlphaFold reached the point where the field called the problem solved, and by 2022 its public database held more than 200 million predicted structures, roughly a thousand times what fifty years of experimental work had deposited. Structures that used to take a career now resolve in an afternoon. It’s worth asking what the old approach cost in energy, not just labor: laboratories running instruments, computing clusters processing crystallography output, years of institutional overhead, to resolve a fraction of what one model run now handles before lunch. I’d bet real money the AI path costs less total energy than the brute-force path it replaced, even counting the data center’s own power draw.
That’s the piece the current conversation keeps missing. AI doesn’t have an energy problem. It has an energy bill, the same as every technology that has ever bought capability with power. What we have is an energy investment problem: decades of underinvestment in dense, scalable production, patched over by fossil fuels we’ve grown comfortable with and dense alternatives we keep vetoing. Fix that, and the AI data center stops looking like a crisis. It starts looking like what it actually is: a technology doing what every technological leap before it has done, and in a lot of cases doing it with less waste than what it replaced.