The Wrong Scandal

Someone posted on X that Fable burned 1.3 million tokens in seven minutes. "That's $160 per hour," they noted. "Equivalent to a $333k/year salary."

The replies were scandalized.

I keep thinking about a different number.

There are attorneys who bill $1,000 an hour to write letters and file litigation against home health agencies — specifically to pierce the layers of corporate shells erected to insulate ownership from liability for a mistake made by an underpaid, undertrained employee. A child died. Twenty million dollars moves. The vast majority of it flows to legal fees and financial engineering rather than to the family, or to any remediation of the conditions that produced the mistake in the first place. Nothing in that entire process — not the letters, not the motions, not the shell-piercing strategy, not the settlement structure — could not have been produced by a large language model.

$160 an hour is not the scandal. The scandal is what we decided $1,000 an hour was worth.

The Comparison Problem

The people doing the $160-per-hour math are comparing against the wrong baseline. The implicit reference point is free ChatGPT from six months ago, or perhaps a junior developer billing $80 an hour, or their own salary amortized into hours. What they are not comparing against is the human labor the model is actually displacing — and when you make that comparison, the outrage inverts.

The expertise economy runs on credentialing, guild membership, and regulatory capture. Medicine, law, finance, architecture: these fields are not expensive primarily because the work is difficult. They are expensive because access to the work has been restricted, and restriction creates scarcity, and scarcity creates pricing power that has nothing to do with the underlying value delivered. A significant fraction of what highly credentialed professionals do is what economists call rent-seeking: extracting value from a position rather than creating it. The corporate-shell-piercing litigation is a clean example. The complexity isn't in the thinking. It's in the paper. And the paper is exactly what language models are extraordinarily good at.

The $160-per-hour number lands as a scandal because it is legible. You can do the multiplication, get to $333k per year, and feel the wrongness of it. The $1,000-per-hour number doesn't land the same way because it has been normalized — it is what experts cost, and experts costing that much is simply how things are. The frame around the first number is "AI is expensive." The frame around the second is "expertise is valuable." These are not equivalent propositions, and we have been treating them as if they were for long enough that the asymmetry has become invisible.

Hype Vocabulary as a Lagging Indicator

Notice that "superintelligence" and "AGI" have quietly disappeared from the conversation. Two years ago, every serious technology publication was running the word; now you encounter it mainly in academic papers and in the rhetoric of people trying to pass AI regulation. The discourse moved on without announcing that it had.

This has happened before. Wired magazine's word of the year, in the mid-1990s, was "disintermediation." The elimination of the middleman was the defining story of the early internet — retailers, travel agents, record stores, classified advertising, all of it collapsing as buyers and sellers connected directly. It was accurate. The analysis was sound. And then the word became anachronistic not because it was wrong but because it described something that had already happened. The bandwidth that was once the ceiling became the floor. "Information superhighway" followed the same arc — a phrase minted in a moment of transition, worn smooth by ubiquity, retired without ceremony when the infrastructure it described became invisible.

The hype vocabulary always overshoots the reality and then falls silent when the reality surpasses the hype. Wired was publishing disintermediation think-pieces in the same issues as innovative-stock lists that included Enron. The frame is always ahead of the facts in one direction and behind them in another simultaneously.

"AGI" is receding the same way. Not because the technology stopped advancing, but because the thing it named has been absorbed into ordinary infrastructure. When every productivity tool, every legal platform, every financial model, every scheduling system has a language model in it, the word for that is not "artificial general intelligence." It is Tuesday.

The Wright Flyer Problem

The first powered flight lasted twelve seconds and covered 120 feet. The Wright Brothers ran four flights that day, December 17, 1903, the longest covering 852 feet in 59 seconds.

Nobody looked at those numbers and said: this is not cost-effective. The fuel consumption per passenger-mile would have been effectively infinite, since there were no passengers and the aircraft couldn't have carried one. The cost per useful output was, by any rational metric, absurd. The press largely ignored it for five years.

What mattered was not the fuel budget. It was that a heavier-than-air craft had left the ground under its own power and come back down under control. Every aircraft that came after — the 747, the F-22, the drone delivering medication to a village in Rwanda — descended from that twelve-second proof of concept. The Wright Flyer was not a product. It was a existence theorem.

We are somewhere in the middle of the equivalent period for machine reasoning. The token cost of an agentic run is the fuel consumption of the Flyer: a real number, measurable, and almost entirely the wrong thing to be paying attention to. The question is not what it costs per hour. The question is what becomes possible once the thing can do the work at all.

Builders, Rent-Seekers, and the Inevitable Succession

Here is the part that requires some ambivalence to say clearly.

Elon Musk and Jeff Bezos have, over the past two decades, genuinely moved humanity forward through technology. SpaceX made orbital launch economically viable when every institution with the resources to attempt it had stopped trying. Amazon built logistics infrastructure that did, in fact, approximate what "disintermediation" promised — removing friction from commerce at a scale that reshaped retail globally. Whatever their current positions in the public narrative — and the transmogrification of both men into cartoon villains of capitalism has been swift and not entirely unearned — they were builders. The things they built are real and they matter.

Builders, eventually, become rent-seekers. It is not a moral failing so much as a structural one. Once the infrastructure exists, the highest-return move shifts from creating new value to extracting value from what already exists. The builder who competed against incumbents becomes the incumbent. The company that disrupted the distribution chain builds a new distribution chain and charges for access to it. This is not unique to technology — it is the standard lifecycle of capital and infrastructure, from the railroads to the broadcast networks to the cloud providers.

What changes at the generational boundary is the loss of the builder's instinct. The founders, for all their faults, remembered what it was to compete against something bigger than themselves; that memory shaped how they built. Their successors in private equity have no such memory. Private equity is not a building operation. It is an extraction operation — acquiring the residual value of what builders created, optimizing the extraction, and leaving the shell. The difference between a builder-turned-rent-seeker and a private equity heir is the difference between someone who knows where the value came from and someone who only knows that it is there.

This matters for AI because we are in the builder phase. The companies deploying at the frontier are genuinely creating new capability, at real cost, with uncertain returns. The token price is partly real infrastructure cost and partly the premium for being at the frontier of something that hasn't commoditized yet. That premium will compress. It always does. Within five years, the compute that powers an agentic run will cost a small fraction of what it costs today, and the $160-per-hour number will look like what $5-per-gigabyte of storage looks like now: a historical artifact of an early market.

The question worth asking is not whether $160 an hour is too much. It is what gets built before the price drops, and who ends up holding the infrastructure when the extraction phase begins.

The Frame We're Using Is Wrong

The token-cost conversation is happening in exactly the wrong register. It is the fuel-consumption-of-the-Wright-Flyer conversation. It is measuring a transitional cost against a free tool that existed six months ago rather than against the human expertise it is actually making redundant. It is treating the disappearance of "superintelligence" from the discourse as a sign of disappointment rather than as the signal it actually is — that the thing has been absorbed, the way fiber was absorbed, the way overnight shipping was absorbed, the way the information superhighway became the thing you stare at during breakfast without particularly noticing.

There are attorneys billing $1,000 an hour to produce documents that a language model can produce.

That is the scandal.

The $160-per-hour number is, by comparison, a bargain.

---

Published by Geordie