Skip to content

Blog

Energy Is the Currency of the Universe

AI turns energy into solutions, and new power can't be built fast enough to keep up. The sun is the end state, and space is where it has no limits. Energy has the qualities of money, and AI fixes the ones it lacked. Energy is the currency of the universe, and we are entering the energy economy.

Three rays join a segmented sun to six solar tiles beside a stack of three round tokens.Three rays join a segmented sun to six solar tiles beside a stack of three round tokens.
Sunlight buys the computation that gives an AI economy its value.

Every cell in your body spends the same currency.

It is called ATP. Biology textbooks call it the cell’s “primary energy currency.” You do not keep much of it. You make it and spend it, all day. Two structural biologists wrote in 2008: “On any given day you turn over your body weight equivalent in ATP.” Muscles, nerves and thoughts are all paid for in it.

The energy behind almost all of that ATP came from the sun. The sun is a fusion reactor. NASA says it “converts about 4 billion kilograms of mass to energy every second.” Plants catch the light, animals eat the plants, and cells turn the food into ATP.

Now a new kind of worker spends energy too. Every token an AI model writes is paid for in joules. Intelligence has joined the economy of life and physics, and it uses the same currency.

That is the claim of this post. Energy is the currency of the universe. AI lets the human economy use it directly. And that makes the energy economy inevitable.

What I said last week, and what changed

On November 9 I wrote The Future of Money Is Energy. I argued that energy is the one input left once intelligence is cheap, and that the sun is the frontier. But I ended carefully: “Money will not become energy.” Energy passed the tests that make something a measure and failed the tests that make it a store. A battery leaks. A record keeps.

The next day, in Money Won’t Survive Super Intelligence, I pushed harder: energy is money is tokens, three names for one thing.

This post makes the stronger claim. Energy is, in essence, money. Two things changed my mind about how careful to be.

First, I looked at the supply side. My earlier posts were about demand: what a token costs in joules. They understated how hard the limit already binds. New power can’t be built fast enough for AI, and the people who run the grid say so in their own reports.

Second, I was judging energy in the wrong form. Raw joules are hard to save and hard to move. But AI doesn’t spend raw joules. It turns them into computation, and computation into solutions. Computation travels at the speed of light, and solutions last. That fixes most of what energy lacked as money.

So this post is mostly research: the demand, the runway, the sun, space, and then the money. I’ll keep three things apart. Evidence comes with a source. Estimates are mine and show their arithmetic. And the argument is mine. The full notes, with every source, date and unit, are in the repository.

Energy is the binding input

To run AI you need data centers, and to run data centers you need energy. No energy, no data centers. No data centers, no AI.

The numbers have moved fast. The International Energy Agency estimated in April 2025 that data centers used about 415 terawatt-hours (TWh) of electricity in 2024, and would use about 945 TWh in 2030. In June, presenting its 2026 update, the IEA revised the picture. Data-center demand “grew by almost 20% in 2025.” Spending by just five technology companies “is now larger than global investment in oil and natural gas production.” And the outlook to 2030 stays about the same only because “bottlenecks … constrain the near-term outlook.” The bottlenecks it lists are gas turbines, transformers, grid connections, transmission lines, memory and packaging.

The United States is where it bites first. Lawrence Berkeley National Laboratory found that US data centers used 176 TWh in 2023, 4.4 percent of the country’s electricity, and would use 325 to 580 TWh by 2028, or 6.7 to 12 percent. Grid Strategies, which tracks utility forecasts, reported in November 2025 that the five-year forecast of US peak demand growth went “from 24 GW to 166 GW” in three years, with about 90 gigawatts (GW) of it linked to data centers. It warns that the data-center share is likely overstated by about 25 GW. Even so, a forecast that grew nearly sevenfold in three years is a new world. NERC, which oversees grid reliability, now projects 224 GW of summer peak growth over ten years, and says AI data centers “account for most of the projected load growth.”

The companies plan in gigawatts. Epoch AI, which tracks the largest sites, counts 96 AI data centers with about 15 GW of IT power as of this month. Its estimates put Amazon and Anthropic’s campus in New Carlisle, Indiana, at about 1 GW, and xAI’s Colossus 2 in Memphis at about 0.9 GW. OpenAI’s Stargate sites add up to over 9 GW of planned capacity. Meta has announced a Louisiana campus that scales to 5 GW.

And they are buying power plants, or at least their output. Microsoft is paying to restart 835 megawatts (MW) at Three Mile Island, now called Crane, which is on track for the second half of 2027. Amazon contracted 1,920 MW from the Susquehanna nuclear plant. Meta signed for 1,121 MW from Clinton in Illinois. Google agreed to restart Duane Arnold in Iowa, as soon as early 2029.

Why spend like this? Because AI can work on almost any problem, if you can power it. That is my argument, and it is the whole reason the demand curve looks the way it does. Intelligence is the most valuable thing there is, and energy is the input it can’t do without.

How long is the runway?

Here is the question the issue behind this post asked me to answer: at the current pace, how long before AI demand outruns what traditional power can add?

First, what “run out” means. It does not mean we run out of fuel. There is plenty of gas, uranium and sunlight. It means new capacity can’t be built fast enough to serve new demand. That is a race between two growth rates.

On firm capacity, the runway is already gone in the biggest US market. PJM runs the grid for all or parts of 13 states and Washington, D.C. Its last two capacity auctions cleared at the price cap, and both came up short of its reliability target: 6,623 MW short for 2027/28 and 6,831 MW short for 2028/29. PJM’s own words: “data centers’ demand for electricity continues to far outstrip new supply.” That is not an estimate. It is the result of an auction.

The pipes are full. The median project built in 2025 took “over 5 years” from its interconnection request to operation, according to Berkeley Lab. GE Vernova, the biggest maker of gas turbines, reported in October that its gas backlog and slot reservations “grew from 116 to 122 GW.” Wood Mackenzie estimates that power transformers are about 15 percent short of demand.

Nuclear is slow. The last new American reactors, Vogtle 3 and 4, began construction in 2009 and finished in 2024, at “more than $30 billion.” Restarts like Crane and Duane Arnold arrive in 2027 to 2029. Ontario’s first small modular reactor is due in 2030. TerraPower’s Natrium reactor in Wyoming got its construction permit this year and expects to finish in 2030.

On energy, here is my estimate. The United States added 53 GW of generating capacity in 2025, the most since 2002, and plans a record 86 GW in 2026. But capacity is not energy. About half the 2026 plan is solar, which runs a quarter of the time, and 28 percent is batteries, which store energy but don’t make it. Only 6.3 GW is gas. When I convert the plan to energy with typical capacity factors, it comes to about 145 to 167 TWh a year of new generation.

Berkeley Lab’s range for data centers means they add about 30 to 81 TWh a year through 2028. So today, new data centers need roughly 20 to 55 percent of all the new energy the country builds. The rest of the economy, which also grows, needs the other half.

Then I let both grow. New supply grows 5 to 10 percent a year. Data-center demand after 2028 grows at one of three rates:

Case Data centers in 2028 Growth after Need half of new US generation Need all of it
Slow 325 TWh 15% a year 2035 to 2048 2043 to 2064
Middle 450 TWh 20% a year 2029 to 2032 2034 to 2040
Fast 580 TWh 30% a year 2029 2029 to 2031

The slow case assumes 15 percent, the IEA’s April 2025 projection for data-center growth from 2024 to 2030. The fast case assumes 30 percent, its projection for AI servers. The middle case assumes 20 percent, close to the almost 20 percent growth the IEA reported for 2025. All three are assumptions, not observations.

Read that as a shape, not a date. On firm power, we are out of runway now in the biggest market. On energy, it depends on how fast AI demand compounds. If it keeps growing at 20 or 30 percent a year, data centers need half of all new US generation by about 2029 to 2032, and all of it by 2029 to 2040. If growth slows to 15 percent, the half mark arrives between 2035 and 2048, and the all mark between 2043 and 2064. So the runway is about three years in the fast case and a few decades in the slow one. The arithmetic ignores retirements of old plants, which make it worse, and efficiency gains, which make it better. The world as a whole has more slack: the IEA’s projected growth in data-center demand is about 15 percent of the new solar energy the world added in 2025 alone. The limit is local, physical and industrial: turbines, transformers, permits and wires.

Traditional power will struggle to keep up. Not because the fuel runs out, but because the factories that make power plants can’t run fast enough.

There is no bigger nuclear plant than the sun

Solar is the one source that is growing at the speed of the demand.

The world added 511 GW of solar in 2025, about three quarters of all new renewable capacity, according to IRENA. Ember reports that solar generation grew by a record 636 TWh, and that solar alone “met 75% of the net increase in electricity demand.” Solar now makes 8.7 percent of the world’s electricity, almost as much as nuclear.

It got there by getting cheap. IRENA’s cost report puts the average cost of utility solar at $0.417 per kilowatt-hour in 2010 and $0.043 in 2024, a 90 percent fall. Fraunhofer ISE finds that module prices fall “about 26.7% for every doubling of cumulative production.” Batteries follow: BloombergNEF reports stationary storage at $70 per kilowatt-hour of capacity in 2025, 45 percent lower than the year before.

Now compare the resource to what we use. Sunlight arrives at Earth at 1,361 watts per square meter. Multiply by the disk of the Earth and you get about 174,000 terawatts (TW). Humans use about 19 TW on average, counting all energy, not just electricity. That is my arithmetic from the Energy Institute’s 600 exajoules a year. The sun delivers about nine thousand times what we use, and the surface alone absorbs nearly half of it.

And Earth catches only a sliver of the sun. Its total output is 3.828 × 10²⁶ watts, about twenty trillion times human use.

What about nuclear? I want safe nuclear to help, and it is moving. Four new reactors reached first criticality in June and July under a Department of Energy program, though all were zero-power tests. The US has a stated goal of 400 GW of nuclear by 2050, up from about 100 GW.

Fusion is further away than its headlines. The National Ignition Facility got more energy out of a target than the lasers put in for the first time in 2022, and set a record of 8.6 megajoules from 2.08 in April 2025. That is a scientific milestone, not a power plant. ITER, the international fusion experiment, now plans to start deuterium-tritium operation in 2039, four years later than its previous target of 2035. Commonwealth Fusion Systems says its SPARC machine will run in 2027, and Google has agreed to buy 200 MW from its first power plant. Helion reported deuterium-tritium fusion in its machine this year. Nobody has yet put fusion electricity on a grid.

But we already have a fusion reactor. It is 150 million kilometers away, it needs no fuel deliveries, and it will run for billions of years. The practical path to fusion power is to catch the light from the one that works. Most energy will ultimately come from the sun.

Space removes the limits

On the ground, solar has three limits: night, weather and land. In space, the first two mostly go away, and the third becomes very large.

Google’s Project Suncatcher notes that “in the right orbit, a solar panel can be up to 8 times more productive than on earth.” In a dawn-dusk orbit the satellite rides the line between day and night and almost never sees the dark.

This is why SpaceX and Elon Musk are pursuing data centers in space. SpaceX’s S-1, filed in May, says: “We expect to begin deploying our orbital AI compute satellites as early as 2028.” And: “Our goal over time is to launch 100 gigawatts of compute to space each year.” It adds that this “will require thousands of launches per year.”

They are not the only ones. Here is what has actually happened, as opposed to what has been announced.

Demonstrated:

  • Starcloud launched a satellite with one NVIDIA H100 in November 2025. The company says it trained and ran a small model in orbit.
  • Google and Planet launched a Suncatcher prototype on October 1. On October 30, Google reported “the first-ever AI response from a data center class TPU in orbit.”
  • Axiom Space launched its first two orbital data center nodes in January.
  • Four days ago, on its fifteenth flight, Starship reached orbit, deployed 26 Starlink satellites, and was caught by the tower. The booster was lost at sea.

Announced: SpaceX’s orbital compute from 2028; Starcloud’s 5 GW orbital data center; Suncatcher’s linked satellites in 2027; Jeff Bezos’s gigawatt data centers in space, in “10 plus years.”

Every demonstration so far is one machine. Nobody has run a data center in orbit.

The reason to try is launch cost. In 1981 the Space Shuttle cost about $65,400 per kilogram to low orbit, in 2021 dollars. Falcon Heavy brought it to about $1,500. SpaceX’s filing says Starship “can eventually reduce the cost to reach orbit by 99% or more.” Google’s team estimates that prices “may fall to less than $200/kg by the mid-2030s.”

Now the hard problems, honestly.

Heat. In a vacuum there is no air to carry heat away. A chip’s heat must leave as infrared light from a radiator. Starcloud’s own white paper works out that a black plate at room temperature sheds about 633 watts per square meter, net of the sunlight it absorbs. That is about 1,600 square meters of radiator for every megawatt of chips, by my arithmetic. A gigawatt needs over a square kilometer and a half of radiators, on top of the solar panels.

Radiation. Google put its TPU in a proton beam. Memory showed errors only after about three times the dose expected over five years in shielded low orbit, and there were no hard failures. That is encouraging, and it is one test of one chip.

Cost. Independent analyses are skeptical. One in IEEE Spectrum this June concluded that, even with “a highly optimistic launch cost per kilogram of US $44,” orbital compute costs “at least an order of magnitude higher” than on the ground. Another estimated that a 1 GW orbital data center “would exceed US $50 billion.” A paper this April calculated that to match the cost of a ground data center, launch plus satellite construction must come to $250 to $1,000 per kilogram, several times below today’s Falcon 9 price.

Latency and repair. A satellite is far from the user and farther from a technician. Training a model tolerates delay better than answering a chat, and nobody has shown repair at data-center scale in orbit.

So here is the honest case. In space, the marginal joule is nearly free. The sun shines all the time and nobody owns it. What costs money is everything else: getting there, shedding heat and replacing what breaks. Those are capital costs, and capital costs fall with scale. The ground’s limits are permits, wires and turbine factories, and those don’t. That is why I think the energy will move to space even if the first orbital data centers lose money.

The currency of life and physics

Back to the cell.

Erwin Schrödinger asked what keeps a living thing alive in What Is Life?, in 1944. His answer: “What an organism feeds upon is negative entropy.” In a note, he admits that to physicists he would have said “free energy.” Life is a flow of usable energy. Stop the flow and the order falls apart.

ATP is how cells keep the books on that flow. In 1941 the biochemist Fritz Lipmann described the energy-rich phosphate bond, the idea behind calling ATP a currency. It is a good name. ATP is made in one place and spent in another. Every cell in every living thing uses it. It is fungible: the ATP from your breakfast pays for your heartbeat and your thoughts the same way.

The astrophysicist Eric Chaisson went further. He measured the energy rate density of systems across the universe: how much energy flows through each gram, each second. Our galaxy runs about 0.5 ergs per second per gram. The sun runs about 2. Plants run 900, animals 40,000, and human society 500,000. The more complex the system, the more energy it pushes through itself.

Galaxies, stars, cells and societies all run on the same thing. That is what I mean by the currency of the universe. Energy is what everything pays to stay organized and to do work.

Energy has the qualities of money

People have made energy theories of value for a century. Frederick Soddy, a Nobel chemist, wrote in Wealth, Virtual Wealth and Debt in 1926 that “the flow of energy should be the primary concern of economics.” The Technocracy movement of the 1930s proposed paying everyone in “Energy Certificates.” Buckminster Fuller split wealth into energy and know-how. Elon Musk said it in an interview last year: “energy is the true currency.”

The critics have a point too. Friedrich Hayek mocked “social ‘energetics’” from Ostwald to Soddy as scientism: the belief that physics can replace prices. Nicholas Georgescu-Roegen argued that matter matters too, not just energy. And the technocrats’ certificates failed, as I wrote in my earlier post, because they were designed not to be saved or traded.

So test it. Here are the usual properties of money, from Jevons and the textbooks. For each, I score raw energy, and then energy in the form AI spends it: compute.

Property Raw energy Energy as compute
Scarcity Yes. Captured, delivered energy is limited by panels, wires, turbines and launches. Yes. Compute is limited by the same energy.
Fungibility Partly. A joule is a joule, but a kilowatt-hour at noon in Texas isn’t one at midnight in Germany. Yes. Move the work to the energy, and send the result anywhere.
Divisibility Yes. It is metered to a fraction of a watt-hour. Yes. Tokens are counted one by one.
Portability Poor. Wires lose power, and no wire crosses an ocean cheaply. Excellent. Results cross the planet in a fraction of a second.
Durability and store of value Poor. Batteries leak and cost money. Mixed. Tokens get cheaper fast, but what they produce lasts.
Unit of account In principle. Every price has energy inside it. In practice. AI deals are written in gigawatts, and intelligence is priced per token.
Intrinsic value The highest of any good. It is the input to everything. High. It is the input to solving problems.

Raw energy fails on portability and storage. That has been the objection to energy money for a hundred years, and it was the objection in my own post last week.

AI changes it. Joules become computation, and computation becomes solutions. You don’t have to ship the energy. You ship the answer. A gigawatt in orbit or in a West Texas field can do work for anyone on Earth over a radio or fiber link. And while a battery leaks, a solution doesn’t. Code, designs, cures and knowledge keep. So does the record of who owns what, which is what Bitcoin’s proof of work buys: in the white paper’s words, “it is CPU time and electricity that is expended.” Nick Szabo called the property “unforgeably costly.”

So the portable form of energy is compute. The durable form is the solution and the ledger. Put those together and energy has every quality of money.

That is why I now go further than I did last week. I don’t mean we’ll pay for groceries in kilowatt-hours. I mean that in essence, what money is for is directing work, and in an AI economy work is energy pointed at a problem. Dollars, bitcoins and tokens are different ways of writing down a claim on that. Underneath, it is energy.

Why this is good for humanity

For all of human history, the human economy ran on a currency nature did not use: human labor and the time it took. Energy reached our goals only through people. That was the bottleneck, and it made our economy separate from the economy of life and physics.

AI removes it. Energy goes in, solutions come out. For the first time, the human economy shares the same fundamental currency as every cell and every star.

That is good news, because the supply of that currency is enormous. The sun delivers about nine thousand times what we use, and space has far more. An economy of labor is limited by people and their hours. An economy of energy is limited by how much sunlight we can catch, and we have barely started.

It also makes the important question clear. If energy is the currency, then building energy is building wealth. Every solar farm, transmission line, reactor and launch is money in the most literal sense I can think of. The countries and companies that understand this will build. The ones that treat energy as a cost to minimize will find they are short of the one thing everything else is priced in.

The takeaway

Energy is the binding input to AI, and AI can work on almost any problem. New power can’t be built fast enough to keep up: on firm capacity the runway is already gone in the biggest US grid. On energy it lasts from about three years, if AI demand keeps compounding at today’s pace, to a few decades, if it slows. The sun is the end state, and space is where it has no limit.

Energy has always been the currency of the universe. Cells spend it as ATP. Stars spend it as light. Now intelligence spends it as tokens, and that lets the human economy use the same money as everything else.

We are entering the energy economy. It pays to notice.