Two men spent an hour arguing about prices, and by the end they were arguing about friction. The occasion was The Ezra Klein Show on 16 October 2026. Paul Kedrosky, a partner at SK Ventures and a fellow at MIT, came to say the AI buildout is a bubble. Azeem Azhar, who writes Exponential View and tracks the industry’s finances, came to say it mostly is not. I have a book coming about what frictionless systems do to the people inside them, so I listened like a builder hearing his subject described by pricers.
Let us start with the disagreement, because it is real. Kedrosky defines a bubble plainly: asset prices detached from economic reality. His evidence is history. Eight or nine times the West has raced to build a general-purpose technology, railroads, canals, electrification, fiber, and each race ended in overbuild and failures. The technology was real each time. Half the track laid in America turned out to be useless, and railroads still remade the world. Useful and bubble are not opposites. They often arrive together.
His second charge is structural. The hyperscalers sit in a prisoner’s dilemma with a balance sheet. Klein quoted Mark Zuckerberg saying that overspending by a couple hundred billion would be unfortunate, but underspending and losing the race would be worse. When Anthropic suggested the industry pace the frontier, the answer, as Kedrosky tells it, was: you slow down. Add China, and no one slows down. Capital plus race, he says, guarantees overcapacity. It always has.
Azhar answers that the demand is real and the machines are spoken for. The asset is the data center, most of its cost in chips, and chip owners today can charge surge prices and earn back their cost in 18 to 24 months. His dashboard reads as a productive boom starting, in his phrase, to get a little bit stinky. He splits a question the rest of us merge: whether the frontier labs have a durable business, and whether the infrastructure gets used. Anthropic could fail, he says, and NVIDIA would still be fine, because firms switching to cheaper open models still rent the same buildings full of chips.
I will not settle that argument here. Listen instead to what both men describe when they stop arguing about prices. Kedrosky notes that token volume is not value. Tokens times price is value, and price keeps falling. The early market was coders, by his count roughly 80 percent of frontier use, and coders are strange customers. They type seven words and get back a million lines. Normal people do the reverse. They hand the machine 25 emails and ask which ones matter. One use expands. The other compresses. Inside real firms, he says, the era of token maxing, with leaderboards for who burned the most, is already over, because budgets blew up in months.
Klein added the quiet number of the hour. After three years, about 2.5 percent of Americans pay for an AI subscription. A stat from Azhar’s own newsletter, played back to him, says more Americans pay for GLP-1 drugs than pay for AI. Hundreds of millions use the free versions. The paid revolution, so far, fits in a small country.
Then Klein said the word the hour had been avoiding: friction.
He raised John Herman’s reporting on the agent flood. Firms now absorb waves of machine inquiries. Agents call restaurants to haggle over tables. Managers drown in machine-written reports. Help arrives so fast that responsiveness collapses. One guest said he now pays a human operator to screen spoof and robo calls from his life, and noted we will count that salary as GDP. Kedrosky called the model launches, in effect, unlicensed drug trials on the public, and predicted FDA-scale regulation after some incident forces it. Then the sentence I wrote down: societies require friction, and removing every friction at once amounts to something like a denial-of-service attack on modernity.
I spent thirty years building the systems these two men were pricing. My book makes a simple claim: friction is the pause where judgment lives. It is the resistance that makes contact real, the place where a user becomes an author again. Neither guest needed my book to get there. The economist trying to pop a bubble and the analyst trying to calm one landed in the same place from opposite sides. Friction is not gone. It moved, and someone else pays it.
It moved to housing, where data centers bid for the same electricians, HVAC crews, and construction capacity as the homes a short country fails to build. It moved to the restaurant that cannot answer its phone. It moved to the surgeon receiving a ten-thousand-word machine-written email from a patient. It moved to your retirement account, which rides this trade whether you ever typed a prompt or not, as Kedrosky kept reminding the room. The top ten stocks sit near 40 percent of the index, and about half of household net worth sits in equities.
That is the receipt. In the book I walk through one ordinary day. I price each small efficiency against its quiet cost: the route never learned, the sentence no one wrote, the apology drained of the discomfort that gave it meaning. The debate hands the same ledger to a whole economy. Each removed friction lands somewhere as a cost with a name. We miss it because the dashboard counts the saving and not the spending.
Machines sell output by the token, and output is the one thing ordinary life does not lack. We lack judgment about what deserves to exist. A market that prices volume calls that a boom until the bill arrives. A culture that prices volume does the same to a person and calls it productivity.
So what would good look like?
The debate offered two exhibits by accident. Apple sat out the frontier race, spent a fraction of its rivals, let their spending build the world its devices inhabit, and outperformed them this year. Azhar reads this as culture: Apple prizes control and detail, and will not risk a model misbehaving in front of a billion people. Call it free riding, as Kedrosky does. Or call it a firm that, by temper or intent, kept its judgment in the loop. The second exhibit is Cloudflare. Kedrosky notes it tripled in 18 months because the flood created demand for the one thing the smooth machine now needs to buy back: resistance. When friction grows scarce, it grows dear. It always was dear. We stopped charging for it and started paying for its absence.
Here is where I land.
The bubble question will settle itself. Bubbles always do. The filings, as both men noted, will soon make customer concentration checkable instead of rumored. The friction question will not settle itself, because it is not a market question. It is an authorship question. Either we put resistance back in by design, in our tools, our firms, and our days. Or it returns after an incident, imposed in anger, at FDA scale. Designed friction is a pause that guards the human. Imposed friction guards only the record that we once had the chance to judge.
Both witnesses named what would change their minds. Azhar watches whether real firms keep spending real money. Kedrosky watches for productivity growth above trend, which has not arrived. Here is mine. Show me a year when these systems give people back time that becomes authored work, work with a name and a judgment attached. Not more throughput for more systems to screen. Do that, and I will revise the book in public.
Here is one I will keep while I wait. In my shop, no machine draft ships without a named author who can defend each sentence aloud. It costs us an hour a day. Some days more. That hour is the tax I pay to stay an author. Multiply that tax by a firm, a city, a country, and you have the design problem. We can pay it on purpose, or we can pay it later with interest.
We were promised ease, and got inboxes and days under siege. The men pricing the bubble can see it. The people building the systems should say it plain.
Friction is the part of the future that remains ours.
the recent cast with dalio about the bubble and wealth to money is a good one if you haven’t checked it out. it’s on the atlantic channel. 👏