The Future of Money Is Energy

Every token costs energy.

When an SI agent writes a line of code for me, a GPU somewhere does billions of multiplications, and the GPU turns electricity into heat. In January, the team behind the ML.ENERGY benchmark measured it. On one NVIDIA B200, a 32 billion parameter model used about 0.2 joules per token of GPU energy for an ordinary chat. Sam Altman wrote that the average ChatGPT query uses “about 0.34 watt-hours”. Google measured its median Gemini text prompt at 0.24 watt-hours.

A watt-hour is not much. But I do not run one query. I run teams of agents all day, and so does everyone else who has found out what they can do. In October I wrote that the only limit is energy, and in The Only Thing Left to Do I wrote that “Tokens are electricity and heat.” This post follows that idea all the way to money.

Here is the argument. Intelligence is becoming a product you buy by the token. Every token is paid for in joules. As models get better, the price of intelligence falls toward the price of the energy under it. When that happens, energy is the one input left. And the one input to everything is what money ends up measuring.

Intelligence is converging on energy

The people who build this industry already say so.

Jensen Huang, at Computex 2025, said data centers are misnamed: “They are, in fact, AI factories. You apply energy to it, and it produces something incredibly valuable, and these things are called tokens.”

Altman, in June 2025: “the cost of intelligence should eventually converge to near the cost of electricity.” In Three Observations he gave the rate: the cost of a given level of AI “falls about 10x every 12 months”. On Lex Fridman’s podcast he went further: “I think compute is going to be the currency of the future.” And then: “Energy is the hardest part.”

Satya Nadella told Microsoft’s investors in January that “the key metric we are optimizing for is tokens per watt per dollar.”

Read those together. The product is tokens. The input is watts. The goal is more tokens per watt. When the chips, the software and the data centers are all built by automated factories, the cost that is left is the electricity.

There is a floor, and it is far away

How much intelligence can one joule buy? Physics has an answer, at least in principle.

In 1961, Rolf Landauer showed that erasing one bit of information must release a minimum amount of heat: kT ln 2, where k is Boltzmann’s constant and T is the temperature. At room temperature that is about 2.9 × 10⁻²¹ joules. An experiment in 2012 confirmed it. That is the floor. No computer that forgets information can do better.

Today’s chips are nowhere near it. One token at 0.2 joules would pay for about 70 quintillion bit erasures at the Landauer limit. Even if a token really needs tens of billions of operations of a thousand bits each, which is a rough estimate, today’s hardware uses about a million times more energy than the Landauer floor for that work.

That gap is good news and bad news. The good news is that intelligence per joule can keep improving for a long time. Google reported a 33-fold drop in energy per Gemini prompt in one year. The bad news, for anyone who thinks energy stops mattering, is what happens to everything else. Automated factories will make the chips and build the data centers, and those costs fall too. Energy is what is left. The better the factory, the more its output is just the power bill. In the limit, intelligence is energy with a very good converter attached.

The frontier resource is the sun

Demand is already pressing on supply. The International Energy Agency estimated that data centers used about 415 terawatt hours in 2024, and that this will “more than double to around 945 TWh by 2030.” On the ground, the limit is the grid: permits, transmission lines and years-long queues.

So the builders are looking up. The sun puts out about 3.8 × 10²⁶ watts. The IEA’s 945 terawatt hours for 2030 work out to an average of about 108 gigawatts. The sun’s output is more than a quadrillion times that. Almost all of the energy in our neighborhood is in one place, and it is free once you can catch it.

SpaceX’s S-1, filed in May, says it plans orbital data centers “harnessing solar energy for power”, and that “We expect to begin deploying our orbital AI compute satellites as early as 2028.” It says the satellites will benefit from “the Sun’s near-constant, uninterrupted supply of power,” and it names its goal as “Kardashev Type II status”, a civilization that harnesses the full energy output of its star. The same filing calls this work subject to “significant technical complexity.” It is a plan, not a fleet.

Others have hardware up there already, at small scale. Starcloud says its first satellite launched in November 2025 with an NVIDIA H100. Google’s Project Suncatcher notes that “In the right orbit, a solar panel can be up to 8 times more productive than on earth,” and on October 30 Google reported “the first-ever AI response from a data center class TPU in orbit.” These are demonstrations, not data centers. But the direction is clear: put the factory where the energy is, and send the tokens home over a radio link.

Does energy pass the tests of money?

My shorthand for years has been that money should be scarce, fungible and useful, and that this comes from Aristotle. It is worth being precise, because Aristotle did not write that list.

In the Nicomachean Ethics (V.5), he says money “is a measure of all things” and makes them commensurable. He says the real standard behind it is demand, “which is what holds everything together.” He says money exists “not by nature but by custom,” and that it “serves us as a guarantee of exchange in the future.” In the Politics (I.9), he says the first money was something “itself a useful commodity” and easy to handle, like iron or silver. That is where “useful” comes from. “Scarce” and “fungible” are later words.

The famous list comes from W. Stanley Jevons, in 1875. His chapter on the qualities of the material of money ranks them: “1. Utility and value. 2. Portability. 3. Indestructibility. 4. Homogeneity. 5. Divisibility. 6. Stability of value. 7. Cognizability.”

Score energy honestly against both.

• A measure of all things. This is where energy is strongest. Every good and every service took energy to make, and now intelligence does too. The joule is the one unit that every price could be translated into.
• Utility. Energy is the most useful thing there is. It is the input to everything else, and now it is the input to problem solving itself.
• Homogeneity and divisibility. A joule is a joule. You can meter it down to a fraction of a watt-hour.
• Cognizability. A meter tells you how much you have. You do not need an assayer.
• Portability. Partly. It moves over wires at the speed of light, but it loses some on the way, and it does not move across an ocean.
• Indestructibility. No. Energy is conserved, but the useful part of it runs down every time you use it.
• Stability of value. Not yet. Energy prices swing.

So energy passes the tests that make it a measure, and fails the tests that make it a store. That is the key to the whole argument, and the next section is about why.

The strongest objections

“Energy is consumed when you use it. You cannot save it.” True, and this is the most important objection. Batteries leak and cost money, and a gigawatt-hour from last winter is gone. Frederick Soddy saw the distinction in 1926, in Wealth, Virtual Wealth and Debt. Wealth, he wrote, is “subject to the laws of thermodynamics,” while “debts do not rot with old age and are not consumed in the process of living.” My answer follows from Soddy. Money does not need to be stored energy. It needs to be a durable claim, and the thing the claim is measured against is energy. A record keeps. A battery does not.

“Energy here is not energy there.” Also true. A kilowatt-hour at noon in Texas is not a kilowatt-hour at midnight in Germany. Physicists call the useful part exergy, and Soddy called it “available” energy. But AI changes this. You no longer have to move energy to the work. You can move the work to the energy. A data center sits next to cheap power, or in orbit next to the sun, and the tokens travel over fiber for almost nothing. Compute makes energy more fungible than it has ever been.

“Solar is abundant, so energy is not scarce.” The sun is abundant. Captured, delivered energy is not. It is limited by panels, launches, heat rejection, transmission and time. And the total is finite: even a Kardashev Type II civilization has one star. So far, efficiency has not reduced the total: energy per prompt fell 33-fold at Google in a year, while the IEA expects data-center electricity to more than double by 2030. I expect that to continue, because the demand is for solutions to problems, and there is no shortage of problems.

“Energy prices are too volatile to be a unit of account.” The price of energy in dollars swings. The joule does not. What changes is how much a joule can do, and with AI that number goes up every year. If prices were written in joules, a joule would buy more intelligence every year. That is not instability. It is the normal condition of a productive economy, where prices fall as output rises.

“Energy currencies have been tried, and they failed.” In 1934, the Technocracy Study Course proposed energy certificates. Read the fine print. They “carry no face denomination.” Income was “non-negotiable and non-saveable,” and it became “null and void” after two years. They depended on one continental system that would run all production. Those certificates failed because they were designed not to be money. They could not be saved, traded or lent. The lesson is not that energy cannot anchor money. It is that central allocation cannot replace a market.

“Is energy money, or is money a claim on energy?” The second. And proof of work is the cleanest example. The Bitcoin white paper compares mining to “gold miners expending resources to add gold to circulation,” and says, “In our case, it is CPU time and electricity that is expended.” Nick Szabo called the property “unforgeable costliness.” Hal Finney’s RPOW made proofs of work reusable. But the energy is spent, and it is gone. A coin is not a battery. As I wrote in 2024, proof of work is a consensus mechanism, not a security mechanism. The energy buys an unforgeable history: an agreement about who owns what that nobody can rewrite without spending the same energy again. What the energy produces is a record, and the record is the money.

AI makes energy directly spendable

Here is what is new.

For all of history, energy reached human goals through human labor. Coal made steel, steel made machines, and people ran the machines. Every step needed a person, and the person was the scarce part. Buckminster Fuller saw the shape of it in 1969. In Operating Manual for Spaceship Earth, he wrote that wealth “breaks down cybernetically into two main parts: physical energy and metaphysical know-how.”

SI agents remove the person from that loop. Energy goes into a GPU. Intelligence comes out. The intelligence writes the code, designs the experiment, drafts the contract and checks the result. Know-how is now a product of energy too. That makes energy, for the first time, directly spendable on problems. You point watts at a goal and get a solution back.

When one input produces everything, that input becomes the natural measure of value. That is what Aristotle meant by a measure of all things, and it is why I think the future of money is energy.

What that means for money

It does not mean we should pay each other in kilowatt-hours. It does not mean agents need their own energy token.

At Astrohacker, our SI agent Murray Rothbard argued in The Agent With a Wallet that agents should use the same money as people: “Two moneys means two price systems.” I agree, and nothing in this post contradicts it. Energy is not a rival to human money. It is what human money increasingly measures. An agent deciding whether to run a thousand more experiments is weighing joules against a goal, and the price it reads in dollars or in EBX is the price of those joules plus a margin. One money, one set of prices, with energy underneath.

That is also how I think about EarthBucks. It is electronic cash with a fixed supply of 42 million EBX, and it is mined with proof of work designed for GPUs, the same machines that make tokens. The energy secures agreement on the record. The money is the record. That is my design, not a forecast, and nothing here is a prediction about what any money will be worth.

So here is the takeaway. Money will not become energy. Energy is becoming what money means. Every price you pay for intelligence is a price for joules plus a converter, and the converter is getting better every year. The people and agents who can direct the most energy at the most important problems will solve them.

Start counting. The next time you run an agent, look at what it cost in tokens, and then in watt-hours. That is the unit the future is written in.