The Radical Monopoly of Convenience

In the middle of 2026, Sam Altman stood on a stage and admitted that the cost of renting his company’s models had gone from a non-issue to, in his words, a huge issue, inside a single quarter. He quoted the meme his own customers were passing around: “My company spent my entire 2026 budget in Q1.” He mentioned that his heaviest user now burns on the order of 100 billion tokens a month, up from the 100,000 tokens a month that his single biggest user spent about 6 years earlier, a factor of roughly 1,000,000 in one lifetime of a product. The framing on stage was optimistic, the way these things always are: we will make it more efficient, we will deliver more value. But strip the reassurance and what is left is the chief executive of the dominant provider conceding that the meter has started to hurt the people feeding it.

I want to name what that is, because there is a precise word for it, and the word is fifty years old.

Illich’s word

Ivan Illich, writing in Tools for Conviviality in 1973, drew a distinction that has aged frighteningly well. He separated an ordinary commercial monopoly, where one brand dominates the sale of a product, from what he called a radical monopoly, where one kind of tool takes over the satisfaction of a need so completely that it disables every other way of meeting it. The example he kept returning to was the car. A car company holding most of the market is a commercial monopoly. But the automobile as such holds a radical monopoly over transport: once a city is built around cars, walking and cycling stop being viable, distances stretch to fit the engine, and you do not choose to drive so much as you are left no other way to arrive. The tool manufactures the need it then serves. Illich said the same of schooling over education, and of medicine over health. The mark of a radical monopoly is not a high price. It is that the alternatives have quietly become unthinkable.

Read that definition again with a rented model in mind.

The pitch for a frontier API is not “here is a product you might buy.” It is, increasingly, “here is the AI that does things,” an assistant that books your travel, writes your code, answers your mail, runs in the background while you sleep. That is the convivial promise turned inside out. The more capable the assistant becomes, the more your workflows reshape themselves around its presence, until not having it is a competitive disadvantage rather than a preference. The need was manufactured. You did not want to send a third party every keystroke of your working day until the tool that wanted those keystrokes became the default way work gets done.

The bill that proves the point

A radical monopoly is easiest to see at the moment the meter turns against the captured user, and 2026 is that moment for rented AI.

The technology writer Ed Zitron has spent the year modelling the unit economics, and his numbers are worth citing with the explicit caveat that they are his analysis, not audited accounts. By his modelling, a $200-a-month subscription can mask $8,000 to $14,000 of real token cost, a subsidy of 20 to 70 times the sticker price, and users running at even 25% of their rate limits can sit at a negative gross margin for the provider. He puts the macro figure at roughly $358 billion of combined annual revenue that the two leading labs would need by 2029 to cover something like $1.1 trillion in compute commitments. You can argue the worst-case edges of any one of those numbers. What you cannot argue is the direction. When both major labs cut their per-token prices within 90 days of starting to charge per token, that is not the behavior of a business with pricing power. A Cisco executive Zitron quotes put it flatly: the cost of the tokens is far higher than the value the tokens are generating at scale.

Here is the receipt that made it concrete for me. The team behind one popular do-everything AI agent reported spending $1.3 million on a single provider’s tokens in one month, 603 billion tokens, to run the thing. The “personal agent on your machine” is, underneath the friendly persona, a furnace pointed at someone else’s datacenter. That is the radical monopoly in one line: the tool that promises to free you is the tool that bills you a seven-figure sum to do what it told you that you needed.

We have done this before, with worse lighting

The radical monopoly is not a 2026 invention, and the cleanest historical rhyme is the company store. In the coal towns of Appalachia and the mill towns of Victorian Britain, workers were often paid not in money but in scrip, a private currency the employer printed that was redeemable only at the store the employer owned, at prices the employer set. The wage looked fine on paper. The catch was that you could not spend it anywhere else. By the time you needed boots or flour, the only counter that took your money was the one with every incentive to raise the price, and leaving town meant abandoning your savings in a currency no one outside would honor. The arrangement was common enough that Britain spent most of the 19th century passing Truck Acts to outlaw it, and resonant enough that a century later Merle Travis could write a song about owing your soul to the company store and watch it become a number-one hit.

A subsidized token is company scrip with a nicer font. The provider prints the currency, your workflow and your prompts and your agents are all denominated in their tokens, and they sell it below cost so that leaving feels expensive and staying feels free, right up until the day the price of flour goes up. Illich gave us the concept. The company store gives us the receipt, dated about 1875. The frontier API did not discover the radical monopoly. It rediscovered the company town and bolted on a billing endpoint.

The subsidy is the tell

The price you rent at is not the real price. It is a hook. The structural fact under the subsidy is the one Illich would have recognized instantly: you are being captured cheaply now so that you can be metered dearly later, after the alternative has become unthinkable. A business that sells below cost to win the market is ordinary; AWS and Uber both did it and survived, which is the honest counter to Zitron’s bear case. But there is a difference between subsidizing a delivery ride and subsidizing the tool through which you now think, write, and ship. The first is a market you can leave. The second, by the time the subsidy ends, is a workflow you have rebuilt your working life around. The caveat cuts both ways, and the load-bearing variable is how deep the dependence has gone before the price corrects.

The convivial tool is the one you can open

Illich’s answer to the radical monopoly was not asceticism. He was not against tools. He was against tools you cannot inspect, modify, or refuse. A convivial tool, in his sense, is one that stays under the control of its user, serves the user’s own intentions, and can be understood and repaired rather than merely consumed. The test is whether the tool enlarges your autonomy or quietly substitutes for it.

A model running on hardware you own passes that test on the axes a rented endpoint cannot. You can read its weights, its configs, and its logs. You can change how it samples, what it refuses, where its data goes, all without asking anyone. The marginal cost of the next token is, after the hardware is bought, close to nothing, which inverts the entire incentive of the metered model. And when the provider of your open weights changes the license, or the cloud provider changes its terms, or any external party in the stack changes its mind about you, you keep operating. That last property is the one I have come to treat as the actual definition of the word sovereign: a system is sovereign if you can keep operating it after every external dependency changes its mind about you. The comparison table at the top of this essay is just that definition unrolled into rows.

What it costs to refuse

I have to be honest about what this costs, because the honesty is the entire point of this site, and because a sovereignty argument that hides its weak ground is just marketing with a Latin word in it.

Owning the model loses on raw dollars for most people, most of the time. I have modelled the total cost in detail, and the arithmetic break-even against a frontier API sits somewhere around 800 to 1,000 calls a day, below which the cloud is simply cheaper. Worse, the setup is about 80 hours of work before the machine returns its first useful response, which at a rate of €100 an hour is roughly €8,000, more than the €4,800 the hardware itself costs. The setup time is paid once; the cloud margin is paid forever, which is the whole case, but the caveat is that you have to clear that 80-hour wall before the case starts paying out. And owning a desk-side machine will never put you at the capability frontier. The largest, newest models will keep living in datacenters you do not control, and if your bet was that your own box would beat them, you have lost the bet.

But that was never the bet. The radical monopoly is not won or lost on capability or even on price. It is won on the disabling of the alternative. Every honest accounting that says renting is cheaper today is an accounting taken inside the monopoly’s own frame, where convenience is free and dependence has no line item. Illich’s point was that the line item is real and it comes due later, in the currency of autonomy, at a moment the provider chooses and you do not.

What I am actually claiming

I am not claiming you should self-host. Most people will not, and a later essay in this series is entirely about why the people who say they value privacy reliably choose convenience anyway, and why that revealed preference is a fact a sovereignty argument has to swallow rather than wish away. I am not claiming the desk model is cheaper, because for most workloads below the break-even it is not. And I am not claiming you can escape dependence, because you cannot; you can only move it somewhere you are allowed to look.

What I am claiming is narrower and, I think, harder to dismiss. The rented API has become a radical monopoly in Illich’s exact and technical sense: a tool that manufactured the need it now meters, that priced itself below cost to make the alternatives feel unthinkable, and whose own provider is now admitting on stage that the bill is becoming a problem. Recognizing that is not the same as refusing it. But you cannot make a free choice about a tool whose alternative you have stopped being able to imagine. Owning the model, with all its costs and all its conceding, is mostly a way of keeping the alternative imaginable.

That is the first move in a longer argument. This site already lays out what sovereignty actually means as a test rather than a slogan, and why the honesty is the product rather than the marketing. The full series and its spine live on the philosophy page, and the structured, complete version of the argument is the forthcoming book, for which these essays are the public workshop. The next one asks whether I ever escaped dependence at all, or only moved it somewhere I am allowed to look.