Intelligence is becoming free, and the past five weeks made the argument for me. In June, Z.ai released GLM 5.2: the top-ranked open-weight model in the world, priced at roughly a sixth of GPT-5.6, free to download and run yourself. In July, Tencent shipped Hy3 and launched the API at a price of zero, literally free for its first two weeks. Then Moonshot announced Kimi K3, a 2.8 trillion parameter model that beats Claude Opus on coding benchmarks, the largest open model ever built, full weights promised on Hugging Face within days of this writing.
Read that list again. The frontier is not leaking anymore, it is being handed out. Three years ago a frontier-grade token cost roughly a hundred times what it costs today, and the labs doing the discounting are not startups burning venture money to buy share. One of them is Tencent. You can argue about the slope of the curve. You cannot argue about its direction. Whatever intelligence costs today, it will cost less next quarter, and less again after that, indefinitely. That is not a market. That is weather.
The standard reactions to this are panic or celebration, depending on which side of the model you sit on. Both are wrong, because both assume it is new. It is a rerun. We have already lived through an economy where the single most powerful force in the world was free, invisible, and available to everyone at once, and we know exactly who got rich off it and how.
The sail lesson
The East India Company was chartered in 1600 and went on to assemble one of the largest commercial fortunes in history. At its height it dominated the largest trade routes on earth and fielded a private army twice the size of Britain's. It was, for two centuries, the most valuable enterprise of its age, and its entire operation ran on wind.
Here is what the company never did: it never made a penny selling wind. Wind was free. It never made real money on sails or hulls either. Those were commodities you could buy in any port from Lisbon to Canton, and every competitor sailed with the same ones. The fortune came from exactly two things. Cargo, meaning what the ships carried: tea, spices, textiles, saltpeter, whatever the far end of the route valued and the near end could not produce. And routes, meaning control over how cargo moved: the royal charters, the ports, the monopolies on which water you were allowed to sail at all.
The power source was free. The vehicle was a commodity. The money was in what you carried and where you were permitted to carry it. Hold that shape in your head, because it is about to repeat twice.
The first cargo era: files
Right now, the cargo is bits. Nearly everything the knowledge economy produces is a file. The contract is a file. The ad campaign, the screenplay, the codebase, the financial model, the film, the model weights themselves: files. For seventy years the constraint on making them was skilled human hours, which is why a career meant becoming the kind of person who could produce a particular kind of file.
AI collapses the cost of making files toward zero, the same way wind made propulsion free for anyone with a mast. Value does not disappear when production goes free. It migrates to the routes: distribution, attention, trust, the channel between the file and the person who needs it. The companies compounding right now are not the ones with the smartest wind. They are the ones that own a port, whether that port is an app store, a feed, an enterprise contract, or a habit a hundred million people open every morning.
The uncomfortable part for people in the intelligence business is that a better model changes this less than they hope. A ship with slightly better sails still pays the same harbor fees.
The second cargo era: atoms and cells
The files era is the rehearsal. The same wind is about to start pushing much heavier cargo.
When robotics comes online at industrial scale, intelligence stops being limited to making and moving bits and starts making and moving atoms: manufacturing, construction, logistics, agriculture, everything that currently requires a human body standing somewhere specific. And behind atoms, cells. Drug design, synthetic biology, food, materials grown rather than assembled. These are not separate revolutions. They are the same revolution reaching cargo classes that could not fit through a screen.
The economics will rhyme with the first era, not with the model wars. The intelligence driving a robot will be as free as the intelligence writing a contract is becoming now. The robot itself will commoditize the way ships did, because hardware always does. What will not commoditize is the cargo and the route: which physical things get made, who has permission to make them, and who controls the channel between the factory and the need. Regulatory approval is a route. A supply chain is a route. A hospital system's trust is a route. The fortunes of the next thirty years belong to whoever picks a cargo class early and builds the route before the wind arrives, which is a sentence that would have made perfect sense to a merchant in 1610.
The East India Company's real moat was not seamanship, it was a royal charter and an army, a state-granted monopoly enforced at gunpoint. Routes are granted by governments and always have been, and the scramble for AI-era charters is already underway: export controls deciding who gets chips, licensing regimes deciding who gets to deploy, sovereign compute deals deciding which countries get ports at all. The winners of this era will hold approvals and allocations the way the Company held its charter, and pretending otherwise is how you end up with excellent cargo and no permission to sail.
The tell
If you think this framing is wrong, you should explain the behavior of the people best positioned to know.
Every lab that owns a frontier model is racing downstack as fast as capital allows: into consumer products, into browsers and devices, into enterprise distribution, into anything that looks like a port. These are the organizations with the most information on earth about where model capability is heading, and their revealed preference is unanimous. They are converting wind into routes at the maximum possible speed, because they can read their own pricing curves.
The generous reading of this is vertical integration: hold the frontier and own the port, capture both margins. Watch the capital allocation instead of the press releases. If these companies believed the frontier would hold pricing power, the products would be a sideline funding the real business, and instead the products are the center of gravity while the models become the cost center. Their own dashboards are saying what their pricing pages already say in public.
The Chinese labs reached the same conclusion from the other direction. Giving away GLM and Hy3 is not generosity. If you cannot own the wind, the next best move is making sure nobody else can either, and publishing frontier weights is the cheapest sabotage ever devised against anyone charging rent on intelligence.
Nobody who owns the wind is betting on owning the wind.
The weather machines
There is one objection worth taking at full strength: if intelligence is weather, why is everyone spending hundreds of billions of dollars on weather machines? Wind is free at the point of use. Tokens are not. Somebody buys the GPUs and pays the power bill, and a free-to-download model with 2.8 trillion parameters is not free to run.
The objection has the analogy pointing the wrong way. Compute is not the wind. Compute is a port. A datacenter is exactly what a harbor was: brutally capital-intensive, geographically concentrated, and permissioned at every layer, from power contracts and land to export controls on the chips themselves. Nobody who builds one believes they are buying intelligence. They are buying the place intelligence has to pass through on its way to a customer, and charging harbor fees on everything that docks.
So the buildout confirms the thesis rather than refuting it. You do not spend half a trillion dollars on an asset you expect to commoditize. You spend it on the thing that collects rent while the commodity flows through.
The voyage ahead
So the stack settles into the old shape, and the businesses still worth building come down to three.
You can build a port. That means compute and channels, the places everything has to pass through on its way to a customer, and it is a very good business if you can get it, which almost nobody can, because harbors have always belonged to whoever could pour the most stone.
You can hold a charter. That means approvals, licenses, and the slow trust of institutions, the things that decide who is allowed to sail at all. Charters take years to earn and compound for decades, and they are granted rather than built, which is why the people who hold them defend them with lawyers instead of prices.
Or you can carry cargo. That means deciding what is worth carrying, and to whom, before the market has repriced it. It is the only one of the three with no entry fee, and the only one anyone remembers. The Company's harbors are silted up and its charter is a museum piece, but the tea trade changed what half the world drinks in the morning.
Pick your cargo.
