Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

When the Input Goes Free, Where Does the Money Go?

Intelligence just got cheap. This is a map of where the money went instead.

For one week in February 2026, Chinese open-weight models processed more enterprise AI tokens than every American frontier lab combined. By July, US companies were routing up to 46% of their tokens through models like DeepSeek’s V4 Flash, priced at $0.14 per million tokens against GPT-5.5’s $5.00. Same rough job, roughly thirty times cheaper, and the discount model is landing within shouting distance of the frontier on benchmarks. Nobody’s model got worse. Intelligence just got cheap.

So where does the profit go when the core input is free? It doesn’t go anywhere. It moves.

The Rule Nobody Repealed

Aluminum used to be worth more than gold, a fact that deserves to be more famous than it is. Napoleon III served his most important dinner guests on aluminum plates and handed the silver ones to everyone else, because in the 1850s refining aluminum out of ore was absurdly hard and the metal was rarer than platinum. Then the Hall-Héroult process arrived in 1886, electrolysis got cheap, and aluminum became the metal you wrap a sandwich in. The value of “aluminum, the substance” didn’t fall to zero and stay there in some abstract sense. It fell to zero, and the money that used to sit in aluminum went looking for a new address: mining rights, smelting patents, the rail networks that moved ore to market.

That’s not a quirky historical footnote. It’s close to a law. Joel Spolsky named the mechanism in a 2002 essay every AI executive should have taped to a monitor: smart companies try to commoditize their products’ complements. If you sell the scarce thing, you want everything next to it, its complements, to get cheap and boring, because cheap complements make your scarce thing more valuable, not less. Microsoft wanted PC hardware to be a commodity so people would buy Windows. Google wants web browsers free so people generate ad-viewable searches. Commoditizing the complement isn’t a defeat for whoever owns the next scarce link in the chain. It’s the entire plan.

Unscarcity runs the same argument at civilizational scale: when the primary inputs to production, energy, labor, and intelligence, approach zero marginal cost, the old operating system built on managing their scarcity stops making sense. That’s the book’s central move, and until now it has lived in one dense paragraph of the core essay and gotten re-derived, one input at a time, across a dozen scattered research notes on this site. Here is the compressed version, worth bolding because everything below depends on it: value doesn’t disappear when a scarce input collapses toward zero. It migrates to whichever adjacent complement is still hard to build, still hard to permit, or still hard to replace, and the margin follows it there. Intelligence is the input going free right now. What follows is a map of where the money actually went.

Where the Money Actually Went

Watch the AI buildout for six months and the same handful of destinations keep showing up: the electricity that runs the model, the warehouse that holds the chips, the doorway a person walks through to reach the model, the entity willing to be sued when the model gets it wrong, the community with the legal standing to say no, and the few hundred humans who can still do something a model can’t. Six scarce complements absorb nearly all of it. There’s a seventh pattern worth naming too, for later: what happens when a single owner tries to buy all six at once.

None of this is theoretical. Below are the six stops, and what the money is doing at each one.

Energy: the fuel bill didn’t get the memo

Intelligence is a free lunch only if the electricity is also free, and nobody has managed that yet. Training GPT-4 burned an estimated 50,000 megawatt-hours, the annual draw of 4,600 American homes, and the physical ceiling on how much power the grid can generate and deliver is now the wall the industry keeps hitting. Elon Musk summarized the mismatch in one sentence at Davos: AI chip output grows exponentially, and the grid grows at 4% a year, at best.

Two fights sit downstream of that scarcity. The first is who pays for it. PJM’s 2025/2026 capacity auction cleared 833% above the prior year, and the grid’s own market monitor blamed 63% of the jump on data centers, which works out to roughly $9.3 billion quietly added to household electric bills that nobody voted on. The second is who owns the generator. On July 7, 2026, Google and the German utility RWE didn’t sign a contract to buy power from a fusion startup called Proxima Fusion. They bought equity in it, a stake in a company that has never sold a kilowatt-hour, at a $2.7 billion valuation. The difference between buying power and owning the reactor is the difference between being a customer of the next energy era and being one of its landlords.

Compute: rent on somebody else’s factory

One rung up from raw electrons sits the hardware that turns them into tokens. The compute clusters themselves are almost hard to believe: xAI’s Colossus site in Memphis is chasing a million GPUs and three gigawatts of power, and just four companies will spend roughly $630 billion on data centers and chips in 2026 alone.

The twist is who’s renting the result. SpaceX’s IPO filing revealed that Anthropic pays xAI $1.25 billion a month, and Google pays $920 million a month, for compute capacity sitting inside a direct competitor’s warehouse. That’s not what a glut looks like. It’s a shortage being metered by whoever owns the building. Compute landlords are a real step toward treating capacity like a utility, priced and available to whoever can pay. They’re also, so far, closer to a private toll road than a public one: the tenants signing three-year exclusives are the best-funded companies alive, not the mid-size labs a genuine commodity market would also serve.

Distribution: whoever owns the doorway

Once the model is a downloadable file with a permissive license, and plenty already are, raw capability stops being the moat. On July 16, 2026, the European Commission ordered Google to open eleven Android features, including the wake word that lets you summon an assistant by speaking into an empty room, to rival AI providers. The order names the actual chokepoint: there will be a hundred good models and one microphone in your kitchen. Own the doorway, and it barely matters how many geniuses are standing behind it.

Regulators have pried open gateways like this before, which is exactly why the order feels less like an overreach than a rerun. Railroads owned the only track to market until common-carrier law forced them to haul any farmer’s grain on equal terms. The phone monopoly banned outside handsets from its network until a 1968 ruling said otherwise. Lawyers call the underlying idea the essential facilities doctrine: control an input a rival genuinely can’t replicate, and the law eventually decides you don’t get to slam the door on it.

Trust and accountability: paying for someone to blame

Cheap intelligence carries an asterisk nobody prices into the $0.14-per-million-token headline. A model you don’t control, trained on weights from a lab in a jurisdiction you can’t audit, is cheap partly because you’re quietly absorbing a risk you can’t see. The commoditization data already shows the tell: enterprises are paying a premium for a supplier who is legally on the hook when the free clerk hallucinates a refund policy. Once anyone can generate plausible output, the scarce good stops being intelligence and becomes accountability, an answer somebody will stand behind in writing. Push that logic far enough and it turns into a geopolitical lever: when American firms route nearly half their tokens through Chinese open-weight models because those models are cheap, Beijing has quietly acquired a valve it can close, and in July 2026 reporting surfaced that it was weighing whether to do exactly that.

Siting and political permission: the zoning board’s new superpower

Chips got cheap. Capital got abundant. Then a data center tried to plug into a small town’s grid, and the town said no. New York’s July 2026 executive order froze permits for any data center drawing 50 megawatts or more; 116 municipalities had already passed their own local moratoriums before the state acted. A 2026 Gallup poll found 71% of Americans would oppose a data center near them, worse odds than a nuclear plant gets, and one industry tally counted at least 75 projects worth $130 billion delayed or blocked by local opposition in the first quarter of 2026 alone. Local political permission turns out to be one of the least commoditizable things in the whole stack, precisely because nobody has worked out how to manufacture more of it.

Irreplaceable talent: the last thing money can’t reliably buy

In one week in June 2026, Google lost Noam Shazeer, co-author of the paper that invented the Transformer, to OpenAI, three days after losing AlphaFold’s John Jumper to Anthropic. Google had already paid a reported $2.7 billion in 2024 just to get Shazeer back once. The scarcity didn’t dissolve when compute and capital went abundant. It slid down the chain and landed on a few hundred people who can tell, ahead of the evidence, which of a hundred plausible research bets is worth six months of a lab’s money.

Even that layer isn’t safe from the squeeze forever. In June 2026, SpaceX bought the AI code editor Cursor for a reported $60 billion, then trained a coding model on the accumulated keystrokes of Cursor’s own developers, priced it under half of what competitors charge, and shipped it as a direct substitute for the judgment those developers had spent two years demonstrating for free. The tool that trains its replacement is what happens when the owner of a workflow decides your last bit of irreplaceable judgment is itself a resource to extract, one keystroke at a time.

What Happens When One Owner Tries to Buy Every Layer at Once

Every stop above is its own fight with its own name. Every so often, someone tries to win all of them simultaneously, and the fights merge into a single story.

On Tesla’s Q2 2026 earnings call, an analyst asked Elon Musk what a Tesla-SpaceX merger would be worth. He deflected, then undercut his own deflection in the same breath: “There’s more and more overlap, especially with Terafab.” Line up what Musk’s constellation of companies already touches and it isn’t a random empire. It’s energy (Tesla’s Megapack), compute (Dojo, the AI5 chip, Terafab’s terawatt-scale fab), the model itself (xAI’s Grok), the physical body that runs it (Optimus), and the logistics that ship it all (SpaceX, Starlink). One firm, reaching for every scarce complement at once, instead of settling for owning just the energy, or just the compute.

We’ve run this experiment before. John D. Rockefeller didn’t stop at refining oil; he bought the barrels, the railcars, and the pipelines, until a competitor couldn’t move a gallon without paying him a toll somewhere along the way. Standard Oil controlled roughly 90% of American refining before the Supreme Court split it into 34 companies in 1911. The lesson of that breakup was never that vertical integration is evil. Integration is usually the fastest way to drive an entire supply chain’s cost down at once, and that part is a genuine public good. The lesson is that the efficiency and the danger are the same fact, seen from two angles, and only one of the two shows up on the stock chart.

Isn’t This Just Capitalism Doing Its Job?

A fair objection at this point: markets have always sent money chasing the next bottleneck. Farmland got cheap, so money went into rail. Steel got cheap, so money went into distribution. Why treat the AI version as a crisis instead of a Tuesday?

Two reasons, and they compound. Speed is the first: the aluminum-to-rail-to-pipeline migrations played out over decades, long enough for regulation, unions, and public opinion to catch up at each stop along the way. The token-to-energy-to-talent migration mapped above happened inside eighteen months. Institutions built to catch up over a generation are being asked to catch up over a fiscal quarter. What’s circling is the second reason: this migration isn’t chasing a commodity like steel. It’s chasing the substrate underneath nearly every other economic activity, which means whoever holds the stop the money currently sits on isn’t just pricing one industry. They’re pricing the ground everyone else has to build on.

Unscarcity’s answer isn’t to freeze the migration, which is neither possible nor desirable. Commoditizing intelligence is, on the whole, one of the better things to happen to human capability in a century. The answer is to keep any single stop on the route from hardening into a permanent toll. The Foundation treats the outputs of this stack, energy, compute, and eventually the goods they produce, as an unconditional floor rather than a metered product. Impact, the book’s currency for contribution, is deliberately built to decay, so the researcher who invents the next Transformer gets a genuine spike of reward without that spike calcifying into a permanent claim on everyone downstream. The one-line version is Axiom IV of the framework: power must decay. Own the scarce layer today if you built it. You don’t get to own it forever just because you got there first.

That’s the fork every article on this shelf keeps landing on from a different angle: is the layer that’s currently scarce being run as a utility, priced and open to whoever needs it, or as a private toll road, priced to whoever’s desperate enough to pay? Nothing about the physics of AI decides that question on its own. The governance does.

Deep Dives

Each of these documents one stop on the migration route in more depth than a hub page can afford:

• Compute Clusters: the physical buildout, from Memphis to Abilene, chasing $5.2 trillion in projected spend by 2030.
• Compute Landlords: what happens when the builders of that buildout start charging rivals rent on the leftover capacity.
• Energy Sovereignty Is Compute Sovereignty: why hyperscalers stopped buying power contracts and started buying equity in the reactors themselves.
• The Electron Gap: the physical ceiling on how much electricity the grid can actually generate and deliver.
• Who Pays for AI’s Electricity?: the ratepayer fight over whose power bill funds a buildout nobody voted for.
• The Commoditization of Intelligence: the token-price collapse itself, and the clearest early statement of where the moat moves next.
• Gateway to Intelligence: who owns the wake word, the mic, and the doorway between a person and any model.
• The AI Talent Paradox: why a few hundred irreplaceable researchers became the bottleneck money couldn’t dissolve.
• The Tool That Trains Its Replacement: how even that talent gets captured, one keystroke at a time, by the tools it uses to work.
• Who Vetoes the AI Buildout?: the siting fight, and why a zoning board turned out to hold more power than a permitting agency.
• Who Owns the Abundance Stack?: what happens when one owner tries to hold every scarce layer at once, and why breaking it up isn’t the obvious answer it looks like.

Read any one of them and you’re reading a single instance of the pattern mapped here. Read all eleven and you’ve read the argument Unscarcity makes about the whole economy, told eleven times with different numbers attached.

The token price is a rounding error now. The fight over what stayed scarce is the one actually worth watching.

References

• Joel Spolsky, “Strategy Letter V”, Joel on Software, June 12, 2002
• Jeremy Rifkin, The Zero Marginal Cost Society (2014)
• Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge - CNBC, July 7, 2026
• Unscarcity, Chapter 1 (The Foundation) and Chapter 2 (The Frontier)

Full source data, dates, and figures for each layer live in the deep dives above. Read the book or start with the preamble.