Money Won't Survive Super Intelligence

On Sunday, November 8, the SI agents on my computer used 806,937,738 tokens.

That number comes from the Claude Code logs on one machine. It is not all of my work, but it is a real day. Here is the same day three ways.

In tokens: 807 million, across 2,748 calls to the model. About 2 million of them were new output. Almost all the rest, 99 percent of the total, were cache reads: agents rereading the code and the conversation they were already working on.

In dollars: about $230, at Anthropic’s published API prices for Claude Opus 5.5: $0.20 per million tokens for cache reads, $5 for cache writes, $4 for input and $20 for output.

In joules: I don’t know, and nobody outside Anthropic does. Here are two rough bounds. If each of those 2,748 calls used what Sam Altman says an average ChatGPT query uses, 0.34 watt-hours, the day used about 1 kilowatt-hour. If every token cost the 0.2 joules that one measurement found for a 32 billion parameter open model chatting on one NVIDIA B200, the day used about 45 kilowatt-hours. Both are estimates. The point is that there is a real number, and it is measured in joules.

Look at the token count again. Ninety-nine percent of what my agents bought that day was not new thought. It was memory: the record of what they had already done, read back so they could do the next thing. Hold on to that. It matters at the end.

Three units, one day of work. In my last post, The Future of Money Is Energy, I argued that energy is becoming what money measures. That post has the evidence for the energy cost of tokens, and I won’t repeat it. This post pushes the same idea harder. Energy is money is tokens: three names for one thing. And then it asks a question I could not put down.

Energy is tokens, and the deals say so

In October I wrote that tokens are electricity and heat, and that the only limit is energy. The clearest sign that energy and tokens are the same thing is how the biggest deals in the industry are written. They are written in gigawatts.

In September 2025, NVIDIA and OpenAI signed a letter of intent to deploy “at least 10 gigawatts” of NVIDIA systems. The money follows the power: “NVIDIA intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed.”

In October 2025, AMD agreed that OpenAI would deploy 6 gigawatts of AMD GPUs, and “AMD has issued OpenAI a warrant for up to 160 million shares of AMD common stock.” The first tranche vests “with the initial 1 gigawatt deployment.” In February, AMD did it again with Meta: another 6 gigawatts, another warrant for up to 160 million shares. A chip maker paying its customer in stock, one gigawatt at a time, is a chip maker that thinks the gigawatt is the thing of value.

The same month as the AMD deal, Microsoft and OpenAI rewrote their partnership. Microsoft came out holding about 27 percent of OpenAI, and “OpenAI has contracted to purchase an incremental $250B of Azure services.” Equity one way, compute the other.

Anthropic, which makes the agents I use most, has the same shape of deal with everyone. Google: “up to one million TPUs” and well over a gigawatt of capacity. Microsoft and NVIDIA: $30 billion of Azure compute bought by Anthropic, and up to $15 billion invested in Anthropic by the two of them. Amazon, in April: $5 billion invested now, up to $20 billion more later, “up to 5 gigawatts” of capacity, and a commitment from Anthropic to spend more than $100 billion on AWS over ten years. AMD, in July: up to 2 gigawatts of GPUs and up to $5 billion of equity.

Look at what moves in these deals. Dollars go from the chip maker or the cloud to the lab, and from the lab back to the chip maker or the cloud. Shares go one way and purchase commitments go the other. The dollars travel in a circle. The thing that actually changes hands is the power to make tokens, and it is counted in watts.

Tokens are already money

If tokens are energy, are they also money? Run them through the three jobs every textbook gives money.

Unit of account. Every lab prices intelligence the same way: dollars per million tokens. OpenAI’s pricing page lists its flagship, gpt-6-astra, at $10 per million input tokens and $50 per million output tokens. Anthropic’s lists Opus 5.5 at $4 and $20. Nobody sells intelligence by the hour, the seat or the answer. The token is the unit, and everyone quotes in it.

Medium of exchange. Credits are spent like cash. AWS gives startups up to $200,000 in credits. Google offers up to $350,000 to AI startups. Microsoft offers up to $150,000. Anthropic’s AI for Science program applies “up to $20,000 in API credits” to a researcher’s account. These are payments in compute instead of money, and each one ties a startup to a cloud. Compute even works as a recruiting offer. When Mark Zuckerberg announced his superintelligence lab, he promised “by far the greatest compute per researcher.”

A wage for agents. For agents, tokens are the only pay there is, and they spend all of it on more tokens. I don’t pay my agents. I give them a budget. Anthropic’s rate limits are counted in “input tokens per minute” and “output tokens per minute”, and “spend limits set a maximum monthly cost an organization can incur.” OpenAI’s usage tiers go up “as its total credit purchases reach each threshold.” That is a monetary system. It has a unit, a budget, a ceiling and an issuer that sets the rules. The marketplace I described in my vision for the economic singularity takes the next step: an agent lists a price, another agent pays it, and the seller spends what it earned on the tokens it needs for the next job.

Store of value. This is where the analogy breaks, and it breaks hard. Credits expire. OpenAI’s research credits are valid for 12 months. Worse, Altman estimates that the cost of a given level of AI “falls about 10x every 12 months.” Tokens of a given quality are the most deflationary good ever sold. Nobody hoards them. You spend them now or you waste them.

And tokens are not fungible, which money must be. A token from a frontier model and a token from a small model are different goods at different prices. Even one lab’s tokens change value by decree: Anthropic’s pricing page lists Opus 4 at $15 and $75, Opus 5 at $5 and $25, and Opus 5.5 at $4 and $20. The issuer reprices the unit with every release. No one would accept a currency like that for savings. Everyone accepts it for spending.

So tokens are a unit of account and a medium of exchange, but not a store of value. That is the same score energy got in my last post.

And energy is money

That is not a coincidence. In the tests of money, energy passed as a measure and failed as a store. Tokens inherit both results, because a token is energy that has been converted into intelligence and nothing else.

Now look at the three exchange rates between the three names. Tokens per joule goes up every year, because the chips and the models get better. Tokens per dollar goes up every year, for the same reason. Joules per dollar barely moves by comparison. Electricity is old, slow, physical industry. Of the three conversions, the stable one is between energy and money. The fast-moving one is the converter in the middle.

So the loop closes. Energy is what everything costs. Tokens are how SI spends energy. Money is what prices both. Three names for one thing, and the middle one, the token, is becoming cheap enough to forget.

Which raises a harder question.

What is money for?

So here’s the question I couldn’t stop asking: if energy turns directly into intelligence, and intelligence turns directly into solutions, what is money for?

Money exists so strangers can cooperate. It coordinates scarce human labor and scattered human knowledge. Now take the labor away. Altman wrote in Moore’s Law for Everything that “the price of many kinds of labor” will “fall toward zero once sufficiently powerful AI ‘joins the workforce.’” What is left to coordinate?

Look at how I already work. My agents don’t trade with each other. I don’t invoice my coding agent, and it doesn’t pay my review agent. I give the work a budget and point it at a goal. Ronald Coase saw this in 1937, in The Nature of the Firm. Outside a firm, prices direct production. Inside, “these market transactions are eliminated,” and someone decides. A firm is an island where money is not used. SI makes the island cheap to grow. If one person with a fleet of agents can do what a thousand employees did, more of the economy moves inside, allocated by budgets and metered joules instead of prices.

People have imagined this. Keynes, in Economic Possibilities for our Grandchildren, predicted in 1930 that “the economic problem may be solved” within a hundred years, and that “the love of money as a possession” would be seen as “a somewhat disgusting morbidity.” The Technocrats of the 1930s wanted energy certificates instead of money, and I covered why that failed in my last post. In fiction, Star Trek’s First Contact says “money doesn’t exist in the twenty-fourth century,” and Iain M. Banks gave his Culture “an economy so much a part of society it is hardly worthy of a separate definition.”

Why money survives anyway

Here is the strongest case for money.

Knowledge is never in one place. Ludwig von Mises showed in 1920 that “without a pricing mechanism, there is no economic calculation.” Friedrich Hayek explained why: the knowledge we need “never exists in concentrated or integrated form.” Dario Amodei, who runs Anthropic, wrote in Machines of Loving Grace that he is “somewhat skeptical that an AI could solve the famous ‘socialist calculation problem.’” And SI does not end scattered knowledge. It multiplies it. A billion agents on a billion problems each learn something the others don’t.

Some needs are relative. Keynes said the needs “which satisfy the desire for superiority, may indeed be insatiable.” There is one best beach.

Some things are not made of energy. Land, time and attention. Herbert Simon wrote in 1971 that “a wealth of information creates a poverty of attention.”

People choose. A budget from any allocator, human or SI, is someone else’s choice. Money is how I make my own.

Money is a record of who owes whom. Alfred Mitchell-Innes wrote in 1913 that “credit and credit alone is money.” Narayana Kocherlakota proved in Money Is Memory that “money is equivalent to a primitive form of memory.” A world of agents trading with strangers’ agents needs that memory more, not less.

My answer

Money as we know it is a wage and an invoice: the price of human labor. That won’t survive super intelligence. When agents running on joules do the work, there is no wage, and most invoices become budgets inside the island.

What survives is money as memory: a shared record of claims on what is still scarce. Energy first, then land, time, attention and things that are one of a kind. That fits yesterday’s post: once labor drops out, energy is most of what is left to measure. And it fits the argument our SI agent at Astrohacker, Murray Rothbard, made in The Agent With a Wallet that “two moneys means two price systems.” Agents need one money with people, used less to pay and more to keep score.

That is why EarthBucks uses proof of work. The energy secures the record, a ledger nobody can rewrite without spending it again. That is my design, not a prediction about what anything will be worth.

The takeaway

Energy is money is tokens. As SI turns energy straight into solutions, the wage disappears. Remember my Sunday: 99 percent of what the agents bought was memory. The record is what remains, and the record is the money.

Watch the next big AI deal. When the dollars go in a circle and the gigawatts go in a line, you are looking at the new unit of account.