The AI industry spent about a trillion dollars this year and took in about $50 billion, on Cory Doctorow's figures, and he says the gap cannot be closed by operating the assets better.
Most bubble arguments end at valuation. Doctorow's ends at the balance sheet: the cost of answering a query is not falling, the hardware has to be replaced every two to three years, and customers switch models for free, so the spending never converts into an asset that simply runs.
"They spent a trillion dollars this year. They made $50 billion"
Doctorow is the author of The Reverse Centaur's Guide to Life After AI, published by Farrar, Straus and Giroux in June 2026, and holds an honorary computer-science doctorate from the Open University; he studied cellular automata at the University of Waterloo before dropping out.
I listened to the full interview so you can skip it. 57 minutes of audio, 23 minutes of reading.
Here are the 11 arguments that matter.
👤 Guest: Cory Doctorow, author of The Reverse Centaur's Guide to Life After AI, who holds an honorary computer-science doctorate from the Open University
🎙️ Host: Isaac Pound, who presents The Tech Report
📰 Published: 10 September 2026 on YouTube (The Tech Report)
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 57 min | ✅ Time saved: 34 min
Key Takeaways
The AI industry spent about $1T this year and made about $50B
His counter-offer: hand him the trillion, and he will give $60B back and keep the rest under a mattress
Six companies are losing money on AI, a seventh is lending them the money they lose, and together they are 35% of the S&P 500
OpenAI's marketing budget is the size of Coca-Cola's, with none of the billboards
He credits Ed Zitron's reading of leaked financials: free inference is being booked as a marketing expense
The spending never becomes a finished asset, because the models have to be rebuilt continuously
Hardware is depreciated over five years and replaced every two to three, and switching to a better model costs a customer nothing
A language model does not understand anything, and that only shows up when someone says something it has not seen before
The Writers Guild is the only group of workers that has beaten AI, using a bargaining right Congress outlawed for almost everyone else in 1947
Firing skilled workers destroys process knowledge that never appears on a balance sheet, and rehiring them does not restore it
Gig nurses in the US are offered a lower wage the more credit-card debt they carry
His advice to governments building an AI strategy is to wait and buy the chips at 10 cents on the dollar after the bankruptcies
1. A World Without People
Isaac Pound opened by separating the AI narrative from the technology, and asked why the narrative has been so successful. Doctorow answered by describing who is paying for it.
He divides the early money into two groups of billionaires, and says the first group has stopped treating other people as real. His account is that extreme wealth surrounds a person with agreeable subordinates and turns everyone else into a statistical abstraction, so any task a non-billionaire performs starts to look easy. He cited Elon Musk calling people who disagree with him NPCs, the term for a video-game character with a fixed script
His example is Amazon's delivery and warehouse work. Doctorow, who is 55, said Jeff Bezos is about ten years older and presides over a workforce that is not allowed to take bathroom breaks. "And if you felt that when a driver or a warehouse worker needed to pee, it was like when you needed to pee, you would have to let them pee."
He read Mark Zuckerberg's career as a series of attempts to remove the people from a social network. The rating site Zuckerberg built at Harvard became, in Doctorow's telling, a machine that held a user's friends hostage — you stay because leaving costs you them. Friends then declined to organize their social lives for maximum engagement, so the feed filled with paid creators; the creators wanted moral consideration too, and the next step is replacing them with chatbots that behave like friends and ask for nothing
The second group of billionaires, he said, may believe none of it and are investing in the sales pitch rather than the product. The pitch to a business owner is the removal of the employee who tells them an idea is illegal, dangerous or impossible. His comparison was that an investor need not believe peptides turn teenagers into love gods to believe Andrew Tate can sell teenagers a lot of peptides
The two promises he says AI dangles at employers are a world without people and a world in which knowing how to do something confers no power. He described the same impulse running through several products at once: job sites without workers, social media without socializing, dating without lovers, programming shops without programmers
2. No Understanding Inside
Asked whether AI will deliver what its builders promise, Doctorow said no, and was careful to separate that from a claim about what machines can ever do.
He is a materialist and accepts that a conscious machine is possible in principle. His objection is to the method: he does not think you get there from what is being built now. He also noted that the industry has called something different "AI" every five to ten years since the mid-1950s
The current approach is a bet made about a decade ago, and he calls it a shrewd one. The older dominant method, symbolic AI, required building a model of the system first — describing everything known about a pancreas in software before asking about tumors — which is hard partly because there are many things people can observe but not explain. The alternative was to take all the observable data, have the computer find statistical correlations, assume a correlation is a cause, and predict from there
He said the fidelity of the results genuinely surprised everyone, himself included, and still does not amount to understanding. The test he offers is how the system fails
A booster once put it to him that his phone's autocorrect predicts his wife's words as well as he does, so it must understand her as he does. Doctorow's answer was that two people can predict each other partly because they repeat themselves and partly because they understand each other, and that a chatbot manages the same trick only because people repeat themselves
The case that matters is the sentence nobody has said before. "And at that point it really helps if you have a model of the other person that you can draw upon to find a response rather than finding the median response to I want a divorce and hoping that it works this time."
On the trajectory, he said returns are now steeply diminishing: each new model costs far more and adds far less than the jumps between the early versions. His analogy for expecting consciousness to emerge from a larger word-prediction system was breeding faster horses and waiting for a mare to give birth to a locomotive
3. Building God, Fighting Rules
Pound asked how Doctorow squares the labs' claim to be building something more powerful than nuclear weapons with their resistance to safeguards and testing environments.
His answer is that the contradiction gives away the position. He contrasted it with J. Robert Oppenheimer, who at least had a stated reason to build the weapon and no idea how to defend against it; here, he said, the builders claim to know how to make a version that does not kill everyone and then decline to build that one because it is harder
He does not think the leaders are lying all the time, and describes the culture inside as self-inflicted dread. His description of it was 98% of a corporate culture consisting of going into a dark bathroom with a flashlight under your chin and saying "AI" into the mirror
He takes the researchers who quit in a panic at their word. "I think they're totally sincere. I just think they're high on their own supply."
4. $1T Spent, $50B Earned
On the claim that AI is inevitable because too many companies and governments need it to work, Doctorow went to the arithmetic and to what data centers consume.
He said states are fiscally constrained rather than monetarily constrained, and that the constraint here is real resources rather than money. His point is that government debt is not household debt, but the fiscal space occupied by data centers is large: they use energy, water and memory chips that other parts of the economy also need, and the people who want those inputs for their own projects have political weight too
On memory in particular, he said anger about RAM shortages preceded anger about data centers, and that there is a limit to how long an economy can be told there is no RAM
The return does not cover it. "They spent a trillion dollars this year. They made $50 billion" — and his counter-offer was that for a trillion dollars in cash he would spend about $600,000 of it on mattresses to hide the money under, hand back $60 billion, and leave everyone less angry
A bailout would not be a one-time repair, which is what he says separates this from the railways. When too many railroads were built, the state could restructure the debt, find someone who could afford to run rail cars, and step back while the track carried traffic. His view is that the AI build would have to be bailed out every year
The reason is that the asset never finishes. Hardware is being depreciated over five years and in practice replaced every two to three, and because switching to a better model costs a customer almost nothing — he said users left in a body when a new version of Claude beat a new version of ChatGPT — the models have to be retrained continuously. "You have to keep building the same railroad over and over again on the same patch of ground."
What he does expect to survive is the open-source side. Open-weight models keep working as long as people want to use them, even if the company that made them fails. Asked whether Sam Altman will still matter, he was blunter: "I think he's going to be a punchline."
5. Inference Called Marketing
The cost of answering a query is the number Doctorow says the sector's profitability claims rest on, and he does not believe it is falling.
The industry calls that cost inference — the work of generating an answer once a model exists — and he treats it as the operating expense against which the capital spending has to be judged. Every independent look at the claim that it is coming down has, in his account, found the claim flimsy
His evidence is a set of leaked OpenAI financials obtained by the tech journalist Ed Zitron. The line item for inference in those documents is modest enough to suggest a path to profitability; further down, the marketing budget is the size of Coca-Cola's
Coca-Cola's spending is visible — three global agencies, billboards on street corners — and OpenAI's is not. "Ed's theory, which I think is totally true, is that OpenAI is giving away shedloads of inference and calling it marketing."
Relabeling the expense does not change it. In his words, it does not matter whether an inference expense is called a marketing expense; it is still an inference expense
6. Stop Growing, Get Repriced
Pound put his own comparison of the capital spending on the table and asked what creates an appetite for it. Doctorow answered with a corporate-finance mechanism.
The host's figures, offered with an invitation to correct them, were railway mania at roughly $45 billion to $50 billion inflation-adjusted, about 7% of UK GDP; the dot-com bubble at $1 trillion to $1.5 trillion, some estimates $2 trillion, about 1% of US GDP at the time; and Morgan Stanley's estimate that AI spending reaches $3 trillion by 2028
Doctorow's reply was that the same bank once forecast everyone spending "six hours in the metaverse" by now
His explanation for the decade of invented markets starts with why growth is worth so much. A share is a claim on future earnings, so a growing firm is worth more than a mature one. Highly valued liquid stock can then be spent instead of cash — to buy a company, to hire a person — which means an acquirer can type zeros into a spreadsheet rather than ask a banker for money
That advantage has a trap attached. The day a company stops growing it becomes, in his words, "grossly overvalued" — and, as he put it through Stein's law, anything that cannot go on forever eventually stops. Google holds about 90% of search; it cannot add more searchers
So the growth has to come from somewhere else, and the options run out in order. New products people like come first; then squeezing a captive audience, which raises profit without adding customers; then narrative. He noted the ceiling on the squeeze — a platform can run ten ads before a video but not fifty
Google Classroom, he said, is an attempt to raise a billion future customers to maturity, and the market wants returns sooner than fifteen or twenty years
Keeping the story going is defensive, not optional. If investors sell the shares, employees hired with stock leave for somewhere the stock is worth something, and those are the people needed to find the next business
He ran through the invented markets that came before this one: the metaverse, NFTs, web3 and DAOs. Blockchains, he said, are real — a way to host a database across computers whose operators do not trust each other and still detect interference — and interesting as computer science, but they turned out to have few applications, leaving money laundering, gambling and remittances to Russia. He added that a use case that holds only while Vladimir Putin is under sanctions for attacking his neighbors is a limited one
AI, he allowed, has more real science behind it than any of those. The last of the four undergraduate programs he dropped out of was at the University of Waterloo, where he worked on cellular automata — unsupervised machine learning, or what he calls theory-free inference — a subject he has followed since he was a teenager and was genuinely impressed by. He said that dispensing with explanation altogether would let researchers find and operationalize many more levers that move the world than they can currently explain
He closed the section on a contradiction he thinks is worth pulling on: the same wealthy investors are preoccupied with falling birth rates and with building a world that has no people in it
7. The Only Union to Beat AI
The reverse centaur of Doctorow's book title is a person doing the grunt work underneath an algorithm that directs them. Pound asked what a worker whose targets have been raised because of AI, and who now spends the day cleaning up after it, can actually do.
The first answer is about language, and comes from the screenwriters' strike. His friend Adam Conover was part of the Writers Guild bargaining unit during the strike over AI. "And one of the things he spent a lot of time talking to people on the picket line about was the importance of being precise in your language and never saying that your fear was that the AI was going to take your job, but rather that your boss was going to give your job to an AI that couldn't do it."
The reason the wording matters is that the audience does not know who you are. Most filmgoers cannot name a screenwriter — which is why a writer's next film is advertised as being by the writer of the last one. If the audience believes it can have as many good films without screenwriters, the screenwriter arguing for their mortgage has made an enemy of the person who enjoys their work
The same structure applies to a radiologist, who exists to find tumors rather than to pay a mortgage. The argument that recruits an ally is not the mortgage but the missed tumor
The durable answer is labor power, and he is specific about the kind. "Because they're the only group of workers in the world who've beaten AI." They did it with a union that can do a weak form of sectoral bargaining, negotiating as freelancers with the employers across a sector. That was formally outlawed by the Taft-Hartley Act in 1947 and the Hollywood guilds' version was more or less grandfathered in
He was dismissive of the individual approach. A shop where expertise is respected and the boss cannot insist on a tool that degrades the work is not won by walking into an office and having words; it is won by having a union
8. The Luddites Were Skilled
Asked how to tell a genuinely useful automation from one designed to devalue labor, Doctorow said there is no bright-line test, and went to the industrial revolution for the reason.
His position is that the test is contingent and the target moves. Automation does eventually get past work that no longer needs doing, but that comes long after the same automation was introduced to discipline labor
He objects to the standard story about the Luddites: they were not afraid of technology, they were the most skilled technologists of the day. They served seven-year apprenticeships, which he compared to a master's degree in mechanical engineering. "Calling a textile worker scared of technology is absurd."
What the mill workers were also angry about, he said, is one thing the story leaves out: the quality of the cloth coming off the steam looms was bad
His point is about the interval, not the endpoint. Machine weaving did eventually beat hand weaving — for the finest weaves, the most complex patterns, and synthetics such as fireproof or wicking fabrics, nobody works by hand. It took hundreds of years, and people lived with poor material for all of them
9. Six Losers and Nvidia
Pound asked what pops the bubble and what it looks like while it happens. Doctorow's answer was that bubbles are brittle and the trigger is only obvious afterward.
He started with how the last one gets retold. The account of Lehman Brothers' margin call setting off the housing collapse was revised, he said, when the Epstein files showed a margin call against Jeffrey Epstein triggering Lehman's — a fact nobody had, and an illustration that the cause gets rewritten as new facts arrive. He called the effect casualty bias
The first fragile structure is circular. "you have six companies that are losing money on AI, and a seventh one they're losing money to called Nvidia, and together they are 35% of the S&P 500, and Nvidia is loaning them the money that they're losing to Nvidia"
This is not consumer hire-purchase, in his framing — not a sofa retailer lending a customer a few hundred pounds, but hundreds of billions lent to firms with no route to profitability, spent at the lender's loading dock on more of the lender's goods. A margin call or an interest-rate shock would be enough
The second is the people. He said OpenAI nearly blew up once already, when a board that he said had grown angry at a chief executive of whom "every word that comes out of his mouth is a lie" fired him and then watched the company fall into chaos and reverse itself within three or four days. Companies staffed that way, he said, can destroy themselves from the inside
The third is the debt behind the buildings. Data-center financing is an inverted pyramid: a small first tranche from risk-tolerant money, then larger tranches from investors who are progressively less tolerant and who demand more break clauses. Late delivery, a move in energy prices, a delayed grid connection or missing diesel turbines lets a financier walk
The danger is correlation. One project losing a financier can replace them; all of them losing financiers at once creates a buyer's market and not enough buyers, and the finished sites "might never get turned on"
The fourth is where some of the money comes from. A lot of it, he said, is Gulf money, and those states now have liquid-gas terminals, pipelines and refineries to rebuild because they are being hit by missiles. Set a contract with an American data-center developer against a national export terminal, and he said the choice is obvious: "Like of course I'm going to rebuild the oil terminal."
10. Process Knowledge Dies
On what is left after the dust settles, Doctorow separated the technique, which he thinks survives, from the damage done on the way, which he thinks does not repair itself.
He says a great deal of what is being done with statistical inference now would be done better and cheaper by conventional software. His example is chess: an LLM has no representation of a board or a piece, only tokens. "And so it will move a chess piece onto a square that one of its pieces is already on." The earliest chess computers, built from valves and electromechanical relays, never made that mistake and were a fraction as powerful
The sensible division of labor, he said, is the reverse: the best chess programs were written with help from language models, so use the model for the grunt work and let a programmer build the chess engine
He was similarly skeptical of scale for its own sake. Coordinating thousands of language models is a genuinely interesting result and also expensive, because every extra model added to a swarm is another bill. A task that needs a thousand of them is probably not being done the cheapest way
The open-weight models, he said, keep getting better because they arrive unoptimized. They are released as demonstrators for the frontier models, so the companies have no incentive to make them efficient. A Chinese hedge fund's internal team, given a reported $6 million to clean up a Meta open-source model, produced DeepSeek, and it was good enough that the market "lopped $600 billion off Nvidia share price" within 24 hours, as investors asked what high-end chips were for
What he worries about is the interval in which workers are fired and replaced by tools that are then switched off. Rehiring does not restore the position: the people have retired, emigrated, retrained or given up looking
The asset being destroyed has a name and no line on the balance sheet. Intellectual property is the intangible that can be written down, filed and sold with the firm, which is why it can be valued. Process knowledge is the part nobody can write down. "But process knowledge is all the stuff that like even if someone put a gun to your head and offered you 20 million pounds, you couldn't write down all of your process knowledge"
His examples of the ropes run from the retired employee who will come in for £50 to unjam a machine nobody else can fix, to the warning women pass to each other about which colleague not to be alone with — which he said is disgusting as a fact and still something the firm carries forward
A chip foundry produced a batch that would only run reliably at about half its clock speed, and knew a customer had a faction internally arguing for a cheap laptop nobody would fund. It called them and offered the duds cheap. The product was the MacBook Neo, which he called the most successful new Apple product of the last ten years and said made Apple billions; the foundry now manufactures low-end chips deliberately, having used up the defective ones
That is the argument for wanting the bubble over sooner. "It's one of the reasons we need to pop the bubble sooner rather than later because every time they fire a skilled worker and replace them with an unsustainable chatbot that's bad at their job, we annihilate more process knowledge." He put the productivity cost at a generation, across several sectors and several economies
11. Wages Set by Your Debt
Asked whether demand will absorb the hardware the way the dot-com fiber was absorbed, Doctorow said this build leaves less behind, and then answered a closing question about whether AI is the endpoint of the platform decay he writes about.
He expects a gross oversupply of data centers, because nobody needs several redundant sites doing the same computation. On what a government should do about it, his standing advice is to wait: "You know, one of the things that I say when I talk to political leaders is even if you think you need a British or a Canadian or an Australian AI strategy, that strategy should start with do nothing until you can buy the chips at 10 cents on the dollar once the company is bankrupt"
Demand for the product does not disappear; the subsidy does. The same people will want a chatbot to generate them a picture afterward. The question is whether anyone will pay the unsubsidized price, and because data centers carry fixed costs, every user who refuses to go from $20 a month to $200, or $200 to $2,000, raises the price for those who remain. He called that a potential death spiral
The hardware is the weakest part of the salvage case. The chips are highly specialized and increasingly not backwards compatible, which he said is deliberate: gamers replace motherboards every few years only after a socket's performance is exhausted, whereas Nvidia will ship an incompatible generation for a single good idea. The retrofit for a data center can then cost more than scraping the site to the slab and building a new one
On the closing question he said AI is the squeezing stage, and also a powerful new instrument for squeezing. The most efficient way to extract the maximum from a customer is surveillance data plus inference, priced per session and per person
His example is algorithmic wage discrimination against gig workers such as nurses and Uber drivers, offered a different rate for every shift on the basis of behavioral data. For nurses in the US, he said the wage is calculated in real time from how much credit-card debt the worker carries, and the more indebted the nurse, "the lower the wage you're offered"
The second mechanism is experimentation without supervision. He contrasted it with the advertising of the Mad Men era, where a Don Draper invented consumer categories by common sense and tested them in focus groups. A system given behavioral data can construct its own segments, run a price or wage change on 5% of a million users, and infer the maximum the rest will bear
The third is that unreliability is useful. A shopping assistant that recommends the more expensive product carrying the higher commission has a ready excuse — "Oopsy dupsy, I guess that was a hallucination, you know, those crazy AIs, what'll they think of next?" — and the same analysis that finds price sensitivity can find which users are least likely to notice
He drew one distinction at the end. Google made YouTube worse from a position of extraordinary profit. "AI is the money losingest thing anyone has ever done." Making a profitable service worse for more money and making an unprofitable one worse are not, in his view, the same act
Doctorow's bottom line is that the AI build cannot be made to pay at its current cost, so the argument is no longer about whether it ends but about how much skilled work is destroyed before it does.
Bonus Insights
The host's analogy for the applications that will not survive the subsidy was a subsidized Concorde running London to Manchester every fifteen minutes for 10p, which would make anyone who kept driving look irrational until the day it stopped
Doctorow's analogy for using a language model where conventional software would do was an advertisement he remembers for a blowtorch sold to melt snow off a driveway. Speaking as a Canadian who has shoveled a lot of driveways, he said it did not last long, for obvious reasons
On why the current excitement is disproportionate rather than misplaced, he used the smartphone. He heard about pocket computers for years from friends at Nokia before they arrived; the industry did not drop everything to perfect them, and he expects statistical inference to settle into that pattern — one line of inquiry among many, improving quietly
He mentioned a chronic pain condition and a course of therapeutic ketamine, and described the drug's main effect as the sense of waking from a dream in which nothing that happened was real. He said he does not think it is a coincidence that several of the people building this are reportedly heavy users
On AI safety specifically, he is not dismissive of risk itself — his objection is to probabilities presented as though they were calculated, when in his view they are asserted
Products, Companies & Tools Mentioned
OpenAI (The leaked financials at the center of his inference-cost argument, and the company whose board fired and rehired its chief executive inside four days)
Nvidia (The seventh company in his circle: the six AI spenders' supplier, their lender, and part of the 35% of the S&P 500 he says the group accounts for)
DeepSeek (Built by a Chinese hedge fund's internal team for a reported $6 million out of a Meta open-source model; he says it took $600 billion off Nvidia's share price in a day)
Claude and ChatGPT (His evidence that switching costs are near zero — users moved when one version beat the other, which is why the models have to be retrained continuously)
YouTube and Google Classroom (Two Google examples: the ad load that cannot keep rising, and an attempt to raise a generation of future customers)
Facebook (His account of a network that held users' friendships hostage, then paid creators, and now looks to chatbots)
Grok (His stand-in for undemanding consumer demand that exists at a subsidized price and may not survive an unsubsidized one)
Amazon (The warehouse and delivery work he uses to argue that extreme wealth stops registering other people as real)
Uber (Named alongside nursing platforms as gig work where the rate is set per shift from behavioral data)
Apple (The MacBook Neo, which he says came out of a foundry's defective batch and a customer's unfunded internal proposal)
Morgan Stanley (Source of the $3 trillion by 2028 estimate the host cited, and of the metaverse forecast Doctorow used to discount it)
Coca-Cola (The visible-marketing benchmark: same budget size as OpenAI's, with agencies and billboards to show for it)
Nokia (Where his friends were confident about pocket computers years before smartphones arrived)
DFS (The consumer hire-purchase contrast to hundreds of billions of vendor financing)
Writers Guild of America (The union he says is the only group of workers to have beaten AI, using multi-employer bargaining grandfathered past the Taft-Hartley Act)
University of Waterloo and The Open University (Where he studied cellular automata before dropping out, and where his honorary computer-science doctorate comes from)
Books & Resources Mentioned
The Reverse Centaur's Guide to Life After AI – Cory Doctorow (His new book, and the source of the reverse-centaur framing: a person doing the grunt work under an algorithm's direction)
Where's Your Ed At – Ed Zitron (The journalist whose reading of leaked OpenAI financials he credits for the free-inference-as-marketing argument)
Adam Conover (His friend and a Writers Guild bargaining-unit member during the AI strike, source of the argument that the precise fear is a boss handing your job to an AI that cannot do it)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: