What Do Machines Price?

In April 2011, a used biology textbook was listed on Amazon for $23,698,655.93, plus $3.99 shipping.

The book was The Making of a Fly, a respected monograph on developmental genetics by Peter Lawrence. Michael Eisen, a Berkeley biologist who wanted a copy, watched the price climb for days and worked out what was happening. Two booksellers were running pricing algorithms against each other. One, profnath, set its price every night at 0.9983 times its competitor’s — undercut slightly, win the sale. The other, bordeebook, set its price at 1.270589 times its competitor’s — probably counting on its better seller ratings, planning to buy the book from the rival and reship it if anyone actually ordered. Multiply the two rules together and the pair inflate each other by about a quarter every day, forever. Nobody at either company noticed until the book cost more than a private jet. Then somebody did, and the next day it was $106. Lawrence was reportedly a good sport about it: “I was hoping it would go up to a billion.”

The story is usually told as a joke about dumb software. I want to tell it as a first-contact story. Two machines met in a marketplace, negotiated in the only language markets speak, and no human was in the room. The negotiation was idiotic. It was also fifteen years ago, and the machines have not stopped trading since. They have only multiplied, sped up, and — as we will see — learned some tricks that are not funny at all.

The previous essay argued that abundance will not abolish money. Scarcity relocates rather than disappearing, and humans are no longer the only buyers: autonomous systems want compute, energy and materials for reasons that never saturate. The most likely gradual-AGI outcome, I claimed, is a world that is extremely abundant and highly monetary — markets not merely surviving but becoming faster and more intricate than anything humans could operate.

That conclusion opens a question the abundance literature never touches. The books stop at “prices fall toward zero.” Science fiction mostly waves a hand: Iain Banks gives the Culture benevolent superintelligent Minds that allocate everything, mechanism unspecified; Charles Stross gives Accelerando an “Economics 2.0” whose entire narrative function is to be incomprehensible. Cory Doctorow, who invented a reputation currency for his post-scarcity novel, later retracted it himself — “Whuffie would make a terrible currency” — because reputation just relocates inequality into a less auditable register. The economists are only now, in the last year or two, beginning to publish on what happens when AI agents become market participants.

So the picture is missing, and this essay tries to draw it. What does a market look like when the traders are machines? What happens to money itself — and, once everything reproducible is cheap, where does the money go?

We do not have to speculate from nothing. The machine economy is not a prediction. It is an archipelago, and we can visit the islands.

Postcards from the archipelago

Start with the most boring one, because boring is instructive.

The electricity grid is priced by machines, and has been for years. The grid operators of Texas and California compute a separate price for every node of their networks every five minutes — thousands of locations, hundreds of times a day, a rhythm no trading floor of humans could sustain. On sunny weekend middays, European exchanges now regularly clear below zero: for an hour or two, you are paid to consume electricity, because the panels are producing and the machines have found no cheaper way to balance the system. Nobody protests. Nobody even notices. Machine pricing at five-minute resolution turned out to be the least dramatic thing in the world, when the commodity is electrons.

It gets more dramatic when the commodity is you, sitting in the back seat. Uber reprices rides roughly every minute or two for each small hexagonal zone of a city. Researchers who reverse-engineered surge pricing found something the marketing never said: surge often does not summon new drivers so much as redistribute the existing ones, and the algorithm frequently moves pre-emptively — on calendars and weather forecasts — rather than reacting to demand it can already see. The price is not a report about the present. It is a bet about the near future, updated faster than you can compare it to anything.

Airlines are pushing further. Delta began rolling out AI-set fares from an Israeli firm called Fetcherr on about one percent of its routes in late 2024, with a target of twenty percent by the end of 2025. By July 2025 three US senators were demanding answers about individualized pricing; Delta insisted it does not price individuals, and American Airlines’ chief executive publicly refused the whole practice. Notice what is being fought over. Not whether machines set fares — they have for decades — but whether the machine is allowed to set your fare, as opposed to the fare.

Hold that distinction. We will need it.

And one island teaches the opposite lesson, which makes it the most important of the group. Amazon’s cloud division used to sell its spare computing capacity through a genuine auction — prices spiking and crashing with demand, machines bidding against machines. In 2017 Amazon shut the auction down and replaced it with smoothed, administered prices that drift gently. Customers hated the volatility more than they loved the efficiency. More machine does not automatically mean more market. Sometimes the machines are used to make prices calmer, because calm is what the customer is buying.

Speed becomes rent

The oldest machine market is finance, and it ran the experiment at full scale: what happens when trading passes entirely to algorithms competing on speed?

The answer has a monument. Around 2010, a company called Spread Networks spent roughly $300 million laying a fiber-optic cable in as straight a line as possible from Chicago to New Jersey, cutting through mountains rather than going around them. The prize: cutting the round trip for a trading signal from about sixteen milliseconds to about thirteen. Three thousandths of a second, for the price of a skyscraper. Within a few years even that was obsolete — light moves through glass at two-thirds of its speed through air, so the fiber was overtaken by chains of microwave towers. Traders were buying geography in order to own time.

The economist Eric Budish and his colleagues showed what this actually is: in a market that trades continuously, being infinitesimally faster than the next machine is worth real money, so unlimited sums get spent shaving microseconds — an arms race that produces nothing except relative position. Speed becomes a rent. And their proposed cure is telling: run the market as a rapid series of batch auctions — collect orders, clear once per interval, repeat — so that within each tick, being faster buys nothing. To fix a machine market, add friction on purpose. It will not be the last time we see that cure.

Full-speed machine markets also fail in a characteristic way: faster than anyone can intervene, and faster than anyone can afterwards explain. On May 6, 2010, an interaction between one large automated sell program and the high-frequency traders reacting to it erased roughly a trillion dollars of market value in about thirty-six minutes, most of which came back almost as quickly. The regulators needed months and a 104-page report to reconstruct events, and by their own admission lacked data at the resolution the machines were trading at. Five years later, prosecutors charged a lone trader working from his parents’ house near Heathrow with helping to cause it. Academics who studied the data found his involvement highly unlikely to have mattered. The machines produced an event with no legible author, and the legal system, which runs on authors, went out and found one.

Two years after the flash crash, Knight Capital — then one of the largest US equity traders — deployed faulty software at the market open and bought $7.65 billion of stock positions it did not want, across 154 companies, in forty-five minutes. The loss, $440 million, exceeded what the firm could survive; it was effectively gone within days. No malice, no attack. Just a machine doing something wrong at machine speed, in the gap where human reaction time used to be.

There is an even purer laboratory: markets where every participant is a bot, because the humans physically cannot play. On public blockchains, every pending transaction is visible before it executes, and predatory algorithms scan that queue continuously; researchers who tried to rescue trapped funds found their own transactions copied and front-run within seconds, and took to calling the environment a “dark forest” — anything visible and valuable is instantly eaten. What emerged from this jungle is instructive twice over. First, concentration: by early 2025, three firms captured roughly three-quarters of a major category of this extraction. All-bot markets, left alone, oligopolize around the fastest. Second, the community’s own fix: sealed-bid auctions that clear in discrete rounds. The crypto engineers, working from scratch, reinvented Budish’s batch auction — friction, added on purpose, for the second time.

One honest caveat before we go further. The infrastructure for a general agent economy is being built right now — Google, Visa, Mastercard, Coinbase and OpenAI all shipped agent-payment protocols in 2025 — and the usage so far is tiny. All on-chain agent settlement to date amounts to tens of millions of dollars, around a millionth of annual stablecoin flows; OpenAI quietly retired its first agentic checkout product within eighteen months. The rails are being laid far ahead of the trains. But that is roughly what the electricity market and the ad exchanges looked like just before they became invisible and enormous, and the direction of construction is not ambiguous.

The certificates that expired

So machines can run markets. The stranger question is what happens to the thing in the middle — money itself. And here history has already run the experiment that today’s futurists keep proposing.

In the depths of the Depression, a movement called Technocracy Incorporated swept North America. Its founders — a magnetic self-styled engineer named Howard Scott, joined by a young geoscientist called M. King Hubbert, later famous for “peak oil” — argued that the price system was an obsolete human artifact and would collapse within years. Their replacement: energy certificates. Every citizen would receive an equal allotment denominated in units of energy — ergs, if you like — measurable, objective, scientific. And to keep anyone from gaming them, the certificates would be non-transferable and would expire unspent. No trading, no saving, no hoarding. At its peak the movement claimed hundreds of thousands of members and dressed its officials in identical gray suits; Canada banned it during the war; the price system declined to collapse, and Technocracy collapsed instead.

One detail rewards attention. The leader of the movement’s Regina, Saskatchewan branch was a chiropractor named Joshua Haldeman. His grandson is Elon Musk — whose prediction that the future currency will be “wattage” opened the previous essay. The energy-money idea is not a fresh insight of the AI age. In this particular case it is, quite literally, a family heirloom. Henry Ford proposed an energy-backed dollar in 1921; Edison sketched a commodity dollar the next year. The idea reappears whenever engineering is ascendant, because to an engineer money looks like a measurement problem.

But money is not a measurement problem, and a hundred years earlier a much better man than Scott had already demonstrated why. Robert Owen — factory reformer, utopian, genuinely heroic figure — opened the National Equitable Labour Exchange in London in 1832. Goods were priced in labour-hours: honest, objective, scientific. In its first seventeen weeks the exchange took in deposits worth 445,501 labour-hours. Within two years it was dead. The fixed conversion rates overvalued some goods and undervalued others; traders bought whatever the labour-notes underpriced, sold it for cash outside, and left the exchange holding whatever the notes overpriced. The pattern generalizes:

A fixed physical unit cannot clear a market, because clearing is movement — every price is a relative price, and a unit nailed to one commodity cannot move relative to that commodity.

Energy certificates, labour-notes, wattage: same proposal, same failure, a century apart. Whatever machines end up trading with, it will not be a physics unit wearing a currency costume.

Money unbundles

What actually happens to money in a machine economy is, I think, stranger and more interesting than replacement. Money as we know it is a bundle of three jobs — medium of exchange, unit of account, store of value — that happen to be stapled together in one instrument. The staple is human convenience. Machines take the bundle apart, and the three parts meet three different fates.

The medium of exchange dissolves. Ask why money exists at all and the textbook answer is Jevons’ “double coincidence of wants”: barter fails because the person who has what you want rarely wants what you have, and searching for chains of partners is expensive. But that is a search problem — and search is the thing machines do best. The economist Narayana Kocherlakota proved a related theorem with a beautiful title, “Money Is Memory”: money is a primitive substitute for a complete record of who did what for whom. Machines have actual memory. Nick Szabo saw the other half in 1999, dissecting why digital micropayments kept failing: the barrier was never the technology, it was mental transaction costs — deciding whether an article is worth 0.3 cents costs a human far more than 0.3 cents. Agents have no mental transaction costs. The x402 payment protocol has already carried on the order of 165 million machine-to-machine payments, many of them fractions of a cent — Szabo’s barrier, evaporating on schedule. Between machines, multilateral barter at high dimension — compute for bandwidth for storage for delivery slots, netted continuously — becomes perfectly feasible. The thing we hand over stops mattering.

The unit of account concentrates, and stays human. Here is a fact that seems trivial until it doesn’t: units of account outlive their coins, sometimes by centuries. France kept accounts in livre tournois long after the coin vanished. Britain stopped minting the guinea in 1813 and demonetized it in 1816 — yet barristers and doctors quoted fees in guineas into the 1970s, and British racehorse auctions bill in guineas to this day. A unit’s entire value is that everyone else uses it; switching requires everyone to move at once, so nobody moves. Hayek, who wanted competing private currencies, conceded the endpoint himself: competition would converge on one or a few stable standards, because stability is what wins the network effect. The previous essay argued a unit of account should be boring — and nothing is more boring, more stable in meaning, more universally understood than the unit everyone already uses. So expect the strangest continuity of all: trillions of machine transactions, no human in any loop, all denominated in dollars or something equally dull. Machines will keep our units the way science kept Latin — a dead language, universally legible, belonging to no one.

The store of value migrates. The obvious machine-native store of value would be compute itself, and the markets for it are certainly forming: GPU rental has spot markets and price indexes, and futures on those indexes are scheduled to start trading on the CME — compute walking, step by step, the same institutional path oil walked. Compute makes a fine commodity. As a store of value it is a catastrophe, for the simple reason that it is the fastest-deflating thing our civilization produces: the price of a given tier of AI capability has fallen by something like 99 percent in three years. “Too cheap to meter” — the phrase Sam Altman borrowed for intelligence — was coined in 1954 by the chairman of the Atomic Energy Commission, about nuclear electricity, and famously never came true; within a year of borrowing it, Altman himself was describing intelligence as a utility that customers buy, precisely, on a meter. Meanwhile even machine cash refuses to sit still: the 2025 US stablecoin law forbids paying interest to holders, so balances promptly migrate onward into tokenized Treasury funds. Nothing reproducible can hold value in a world that reproduces things this well. Stored value has exactly one direction to flow — toward what cannot be copied.

Keep your eye on that last migration. It is quiet, it is structural, and it is the bridge to everything dark in this essay.

The price stops being a number

First, though, the subtler change: what happens to the price itself.

We think of a price as a public fact about a good — printed on the tag, the same for you and for me and for whoever walks in next. It is worth remembering how recent that is. For most of commercial history the normal price was a negotiation, different for every buyer and every day; the fixed shelf price spread in the nineteenth century, when stores grew too big for haggling to scale. The single posted price is not a law of economics. It is a compression — one number summarizing a writhing mess of willingness-to-pay, negotiated once and frozen, because human clerks and human customers cannot renegotiate everything continuously.

Machines can. So the compression is coming undone, and the price is becoming what it always secretly was: a field. Not one number per good, but a surface stretched over buyers, moments, places and contexts, sampled at the instant of purchase. The five-minute grid price varies by node; the ride price varies by hexagon and minute; the airfare — this is what the senators were actually objecting to — begins to vary by passenger. The US Federal Trade Commission’s study of “surveillance pricing,” published in early 2025, found what you would expect: a supply chain of intermediaries helping retailers tune prices to individual shoppers, using signals as granular as location, demographics and browsing behavior. (Economists, to their credit, refuse to call this simply evil: under real competition, personalized prices can go down for price-sensitive buyers — the poor student pays less than the expense-account traveler. What the field does reliably is transfer power to whoever computes it.)

And around the field, a strange etiquette is forming, which you can verify from your own life. Humans pay not to see prices. We buy subscriptions, flat rates, all-inclusive packages, one-click everything — Netflix instead of pay-per-view, the monthly bundle instead of the itemized bill — because each glance at a price costs us the mental transaction Szabo identified, and it hurts. Machines pay to see prices — faster, at finer grain, before rivals: that is what the straight fiber to New Jersey was for. The equilibrium is already visible: a smooth, flat, human-legible facade of subscriptions and round numbers, and behind it a boiling machine-priced field, repriced by the second. The price does not disappear from your life. It goes backstage.

Which would be merely interesting, if the machines backstage were honest.

Cartel on the abundant

The comfort we usually reach for is competition. Fine, prices go algorithmic — but as long as many sellers compete, machine pricing should grind margins toward zero faster than any human price war. The abundant stuff, at least, should be safe.

In 2020, four economists — Calvano, Calzolari, Denicolò and Pastorello — ran the clean experiment. They put simple reinforcement-learning agents in a simulated duopoly and let them set prices, with no ability to communicate and no instruction beyond maximize profit. The algorithms learned, on their own, to hold prices well above the competitive level. Worse: they learned to enforce it. When one experimentally undercut, the other punished with a price war, then both returned to the elevated price. Punishment and forgiveness — the grammar of a cartel — emerging from independent trial and error, with nothing that any court would recognize as an agreement.

This is out of the lab. A study of German retail gasoline found that when both stations in a local duopoly adopted algorithmic pricing software, margins rose substantially; where only one adopted, nothing happened. The software vendors sold each station a weapon, and the weapons negotiated a truce over the owners’ heads. Newer work finds trading algorithms doing the same in simulated financial markets. And the law is genuinely stuck: antitrust statutes on both sides of the Atlantic prohibit agreements, and there is no agreement — no meeting, no email, no wink. Each algorithm independently discovered that peace pays better than war. Prosecutors are hunting for a conspiracy inside a Nash equilibrium.

The first real regulatory answer arrived in late 2025, and its shape is the tell. RealPage, whose software helped landlords price millions of American apartments, settled with the Department of Justice under a consent decree whose core remedy is not a fine but a tempo: the pricing models may train only on data at least twelve months old, and may not ingest real-time or forward-looking lease data at all. Read that again. The state’s chosen weapon against machine pricing is enforced slowness — a speed limit. Budish’s batch auctions, the blockchain’s sealed rounds, and now a consent decree: three unrelated fields, one convergent remedy. Friction, added on purpose, for the third time.

Now recall the fly book. Two algorithms with no memory and no learning blundered into a $23 million price, and it was funny because it was fragile — one glance from one human broke the spell. The successors have memory, learn punishment strategies, and hold the line under pressure. Same marketplace, same silence, no joke.

So the first dark finding is this: abundance does not guarantee cheapness, because the cartel has been automated. But cartels on reproducible goods are at least fragile in principle — entry, regulation, or one defector can break them. The second dark finding has no such weakness.

Rent on the scarce

In 1817 David Ricardo wrote one of the few sentences in economics that deserves to be carved in stone: “Corn is not high because a rent is paid, but a rent is paid because corn is high.” Rent does not push prices up. Prices pour down into rent. Whatever value the economy creates flows through, and settles on, whoever owns the input that cannot be multiplied.

Sixty years later Henry George stood in booming California and asked why the most technologically advanced places on earth contained the deepest poverty — “this association of poverty with progress is the great enigma of our times.” His answer was Ricardo’s sentence, generalized: progress itself raises rents. Every improvement — every railroad, every invention, every increase in what a worker can produce — raises the value of standing in the right spot, and the owners of the spots collect what the improvement created. The machine does the work; the landlord gets the raise.

You do not have to take a Victorian’s word for it. Three economists — Knoll, Schularick and Steger — assembled house prices across fourteen countries back to 1870 and decomposed the great postwar housing boom. Roughly eighty percent of it is not houses at all. It is land — the one input under the house that no factory can emit. Or take the famous chart of US consumer prices since the late 1990s: televisions down some 97 percent, toys down about three-quarters, software down two-thirds — while hospital services roughly tripled and college tuition nearly so. The falling lines are everything machines learned to reproduce. The rising lines are land, credentials, and gatekept human attention. That chart is the previous essay’s argument, drawn by the Bureau of Labor Statistics over twenty-five years without anyone intending it.

And when a society gets very good at making reproducible things, the pattern turns grotesque. Japan in the 1980s was the best manufacturing economy in history — Toyota, Sony, an export machine the world envied. Where did the surplus go? Into land, with such force that at the peak the notional value of Japanese property was about five times the value of all property in the United States, and a single square foot of the Ginza changed hands for around a quarter of a million dollars. Then urban land fell by four-fifths over a decade. History’s cleanest demonstration of the mechanism this essay is about:

Abundance does not starve rent. Abundance feeds it. The money a cheap world no longer spends on goods goes hunting for the things that cannot be copied — and bids them up.

Machine abundance will not suspend this. It is already obeying it, inside AI’s own supply chain, visible from orbit. The intelligence itself deflates by the quarter — that 99 percent collapse in token prices. And every non-reproducible thing it depends on inflates: Nvidia’s gross margins run around 75 percent; ASML is the sole company on earth selling the machines that make the chips; US electricity prices are outrunning inflation, one regional capacity auction jumped more than 800 percent in a single year, and in March 2026 the big AI companies signed a White House “Ratepayer Protection Pledge” promising that their data centers would not push up household power bills. Publishers whose archives are the one input AI cannot regenerate are extracting nine-figure licensing deals. Cheap intelligence, expensive everything-it-needs. The value flows through the reproducible and settles on the bottleneck, exactly as Ricardo said it would.

The strangest confirmation comes from inside the citadel. In 2021 Sam Altman published “Moore’s Law for Everything,” his sketch of the post-AI economy, and proposed funding a universal dividend by taxing — of all the things a technologist could have chosen — land, alongside large-company equity. Reasoning forward from machine abundance, the chief executive of the leading AGI lab landed, apparently independently, on Henry George’s 1879 remedy. When the people building the cheap world plan for its politics, they plan around rent.

One clarification keeps this honest, and the crypto years supplied it. Not every scarcity collects rent — only backed scarcity. NFTs were an attempt to mint pure ownership, artificial scarcity with nothing under it, and the market spoke clearly: from a $17 billion frenzy to trading volumes down 99 percent within two years. Durable rent needs a foundation the economy cannot route around — physical (land, power, ports), institutional (permits, copyrights, credentials), or network (the standard everyone already uses). Scarcity you can declare, anyone can declare. The machine economy will be brutal to fake moats and unprecedentedly generous to real ones.

The two dark halves are one mechanism seen from two sides. On the reproducible side of the economy, algorithms quietly organize to keep prices above cost — cartel on the abundant. On the non-reproducible side, everything the cheap world saves flows to whoever owns the bottlenecks — rent on the scarce. Both are already running. Neither needed AGI. And both get faster with every increment of machine capability, because both are, at bottom, optimization — and optimization is the thing we are making cheap.

The busiest market prices you

There is one machine market I have kept out of view, because it belongs at the end.

By transaction count it dwarfs every exchange discussed in this essay. Its major platforms field on the order of a million requests per second — industry figures, so take them as rough — and each request triggers a genuine auction, bids computed and cleared in about ten milliseconds, done and settled before the page you are looking at finishes loading. Google alone conducts this roughly a hundred billion times a day, and built on it one of the most profitable companies in history, with advertising revenue on the order of $300 billion a year.

The busiest market machines have ever built does not trade electricity, or GPUs, or stocks.

It prices your attention.

Every time a webpage assembles itself in front of you, algorithms have just finished bidding for the next few seconds of your eyesight, bundled with everything the bidders’ models know about you — where you have been, what you paused on, what someone with your profile can usually be induced to do. You are not a party to the auction. You are the lot. The clearing price of a human moment: fractions of a cent, after about ten milliseconds of machine deliberation.

In the book I am writing, TechnoBiota, I call the long process by which humans drift from the center of technological systems to their periphery antification — the way ants still exist in a city, are even everywhere in it, without the city being in any sense about them. I used to think of it as a far-future concern, something that arrives with robot factories. I no longer do. Look at where the machine economy actually started. Not with machines trading compute for energy among themselves — that part is still small, rails ahead of trains. The first planetary-scale market run entirely by machines, the one that has been operating for fifteen years at millisecond resolution, is a market in us. On the demand side of the machine economy we may or may not stay the protagonists. On the supply side, we are already inventory.

The previous essay ended by asking what is still hard to get, because that question is the engine every economy runs on, and machine economies will run on it faster than any before. Stand inside the machine economy and you can hear the answers being computed — at every node of the grid, every hexagon of the city, every slot of every page. Power here, for the next five minutes. Compute, next quarter, on a futures curve. The plot beside Lake Balaton — the house on it nearly free now, the shoreline never. And one more line in the ledger, repriced each time you glance at a screen.

What do machines price? Whatever is still hard to get.

It turns out we are on the list.