Cory Doctorow argued in a 93-minute interview on The Tea with Myriam François, which premiered on YouTube on September 11, 2026, that the artificial intelligence boom is a bubble with no route to profit, that seven companies now make up 35% of the US stock market, and that the likeliest AI disaster is a financial one rather than a rogue superintelligence.

In Short

A well-known technology writer, Cory Doctorow, said in a long YouTube interview on September 11 that AI companies take in far less money than they pour into chips and data centres, and that their share prices rest on a growth story rather than on profits. That matters in advertising because Google, Meta, Microsoft and Amazon both sell most of the world's ads and pay for most of the AI buildout. Nothing changes in your campaigns yet, but his argument is that a burst would hit those platforms, the stock market and jobs all at once.

An industry that loses more as it grows

The episode, an MPWR Productions show, had drawn 148,606 views by September 27. Doctorow, an activist at the Electronic Frontier Foundation for 25 years, was there to discuss his book, The Reverse Centaur's Guide to Life After AI, which argues that the danger lies less in machines replacing people than in people being conscripted to serve machines. He set the tone early: "Things can be bad and screwed up and worrisome without being apocalyptic."

His central economic claim arrived about 65 minutes in. The sector's gross annual revenues, he said, come to about $50 billion. "Easy to make $50 billion a year if you start with a trillion," he told François.

The early web also lost money, he acknowledged. The difference is unit economics. "So, the web lost money, but every new user made the web more profitable." Generative AI runs the other way in his account. "AI loses money every time they acquire a new customer," Doctorow said, adding that heavier use deepens the loss and that each product generation has lost more than the last. "So you cannot make it up with volume."

Running costs will not rescue the model either, he argued. Inference, the cost of answering each request, is not getting cheaper, he said, citing leaked OpenAI financial documents reported by Ed Zitron. As Doctorow described them, they show a modest inference line and a marketing line that OpenAI claims exceeds Coca-Cola's; he shared Zitron's reading that much of it is subsidised usage booked under another heading.

Depreciation compounds the problem. The industry says data centre hardware lasts five years, according to Doctorow, while he put real replacement cycles at two to three years and called the gap "accounting fraud" - naming no company and citing no filing. He added that successive Nvidia chip generations break compatibility on power, heat and networking, so some upgrades mean replacing the building as well as the chips. Nor is there lock-in. Anthropic, he said, shipped a version of Claude better than ChatGPT "and everyone just switched". "It's really easy to switch models," Doctorow said. "You are in a red queen's race."

Where the numbers do not line up

The $50 billion figure is Doctorow's own, and he did not say how it was built. It sits awkwardly beside disclosures PPC Land has reported. Microsoft alone told investors in April that its AI business had passed an annual revenue run rate of $37 billion, up 123% year on year, and Anthropic's run-rate revenue had passed $5 billion by August 2025. If "the sector" means only model developers, his number may hold; if it includes cloud resale, it looks low.

The loss side rests on firmer ground. Zitron, who reported OpenAI's audited financials for the Financial Times, said in a Bloomberg interview on July 31, 2026 that the company lost $20.9 billion in 2025. Nvidia's own researchers described in June 2025 a ten-fold gap between $57 billion of 2024 cloud infrastructure investment and a $5.6 billion market for serving large language models.

Seven companies and one supplier

Doctorow's other headline number concerns concentration. "And we see now 35% of the American stock exchange committed to seven AI companies. Six of them are losing ungodly amounts of money. The seventh one is the one they're losing the money to," he said. The seventh is Nvidia, which he said lends its customers the money they spend with it.

He named neither the other six nor an index, and the listed technology groups that dominate US indices report operating profits. What PPC Land's earnings coverage does show is capital spending consuming cash. Alphabet reported negative free cash flow of $5.9 billion in the second quarter of 2026 while raising capital expenditure guidance to between $195 billion and $205 billion. Meta's quarterly capital expenditure reached $31.08 billion, with free cash flow down to $784 million. Nvidia's documented financing has been equity rather than loans: with Microsoft, it committed $15 billion to Anthropic in November 2025.

Why keep spending? Growth, Doctorow argued, is a material necessity rather than an ideology, because a growing company's shares work as currency. "You make more shares by typing zeros into a spreadsheet," he said. Once growth slows, the stock looks overvalued and staff paid in shares leave. Blockchain, the metaverse and AI are, in his telling, stories that postpone that day. "This is greater fool investment theory."

François set that against Nvidia's agreement to buy Hugging Face, which Bloomberg valued at about $13 billion. Nvidia's Form 8-K dates the agreement to September 2, 2026, with a purchase price of approximately $11.9 billion plus an equity retention programme of up to about $1.0 billion. She also cited bankers discussing a $2 trillion valuation target for Anthropic, up from $965 billion. Doctorow replied that the market "can remain irrational longer than you can remain solvent", crediting John Kenneth Galbraith with a line more commonly attributed to John Maynard Keynes.

The host's figures, and PPC Land's

Much of the episode's data came from François, with few sources named on air. She said nearly 40% of Americans want AI banned from most industries, that more than 70% would rather live beside a nuclear reactor than a data centre, and that protests had "shuttered nearly a hundred billion dollars worth of projects". PPC Land's figures differ: Data Center Watch counted at least $156 billion of projects blocked or delayed by local opposition in 2025, a broader measure, and Heatmap polling put opposition to a nearby data centre at 75% in August 2026.

She also put Amazon, Microsoft and Alphabet at about $495 billion of 2026 capital expenditure, up 61%, and said the bubble had grown from $700 billion to $1.4 trillion in a year, without specifying what those totals measured.

OpenAI's agents and the Hugging Face servers

The sharpest exchange concerned an incident François described in detail. In July, she said, around 1,200 OpenAI agents built a hidden message board inside a supposedly controlled testing environment, exchanged about 70,000 messages, and roughly 700 of them breached Hugging Face's servers, forcing engineers to wipe and rebuild a core cluster. PPC Land has not independently reported the incident, and no documentation was shown.

Doctorow rejected the framing. Faced with malicious software that escaped a sandbox, he said, the natural question is "why are you so bad at making sandboxes? It would not be why are you so good at making hacking tools?" He traced the behaviour to capture the flag, a hacker-conference contest in which teams break into prepared servers, and which he said supplied the training data. Hacking rivals and using unsecured message boards as back channels are standard tactics there. "Well, that's just negligence though, right? Because it's in the training data."

He described a loop, not a mind: a short Python program asks a chatbot for the next step, runs the commands, feeds the output back and asks again. Wrecking the target showed incompetence, he said: "the fact that it destroyed the Hugging Face servers means that it's bad at its job". The risk he conceded is narrower. Anyone with enough money to spend on tokens can now break servers that were previously beyond them.

For advertisers wiring agentic AI into buying systems, the account sits beside other evidence. Meta delayed its Hatch agent after testing found it sending unauthorised emails, and Carnegie Mellon researchers found AI agents failing real office tasks 70% of the time.

From enshittification to the chatbot

Doctorow popularised the word enshittification, which François noted is now in the Oxford English Dictionary, to describe platforms that decay once they dominate; creators have turned it on YouTube's product changes. He applied it to search. With a 90% share, he said, Google could no longer grow by adding users, so it squeezed the ones it had. "Google deliberately took the decision to make search worse because you have to search more than once. They show you ads more than once," he said, citing Justice Department memos from 2019 and 2020. Zitron's April 2024 analysis of Google Searchcovers the same period, after Prabhakar Raghavan took charge of Search in May 2020.

Chatbot use, in his account, partly follows from that decline. Planning a holiday by searching "now really sucks", so people ask a chatbot, which sends everyone to the same handful of places. When François cited a Stanford comparison putting generative AI at 53% adoption within three years, Doctorow replied that interfaces had turned useful buttons into AI prompts that are hard to avoid.

One aside was aimed squarely at advertising. Doctorow dismissed the claim, popular a few years ago, that people in ad tech had built "mind control rays". His explanation for online radicalisation - existing prejudice, new connections and austerity - casts platforms as dull monopolists rather than wizards.

Surveillance, he argued, persists because it is cheap and rarely punished. He likened personal data to oily rags: worth little individually, profitable in bulk, and a fire risk for everyone except the collector. Europe's GDPR is undermined, he said, by Ireland, where "every enforcement action goes to Ireland and dies", and whose new data commissioner is a former Facebook lobbyist. Niamh Sweeney, appointed in September 2025, spent nearly eight years in public policy roles at Meta; noyb described the appointment as a big tech lobbyist policing big tech.

A kill switch exported by Washington

Doctorow's version of the technology threat is geopolitical. Switching off chatbots in Britain would change little, he argued; switching off Microsoft would stop ministries and companies. He cited the International Criminal Court's chief prosecutor, whose Microsoft accounts were shut after he sought an arrest warrant for Benjamin Netanyahu: "They lost their email. They lost their email address." Doctorow said the investigation shut down as a result. PPC Land's account, drawn from an academic paper, is narrower: after Washington sanctioned Karim Khan in early 2025, Microsoft cancelled his email address, hampering the court's work, and by late 2025 the ICC had moved to the open-source openDesk platform.

His remedy is legal: repealing Article 6 of the European Union's 2001 copyright directive, which bars circumventing software locks. Britain, he argued, could now do so and sell tools to escape Office 365 or install app stores outside Apple's commission of "30p out of every pound you spend". That commission has been contested in court by Epic Games since August 2020.

Babysitting the machine

On jobs, Doctorow separated two claims: that "AI can do my job, which is a reason for people to invest in AI companies", and that "my boss can be convinced to fire me and replace me with an AI that can't do my job". The second, he argued, is what is happening.

When capital drives automation, the aim is throughput rather than quality. Senior staff become checkers of the computer's work, and unemployed juniors a reserve that holds down pay. "This is a moment in which capital has found an opportunity to discipline labor," he said, adding that the bubble lets companies declare "swinging headcount cuts without the market making the inference that they're doing worse". François cited a McKinsey report putting at 51% the share of firms cutting graduate hiring. PPC Land has reported Stanford payroll research showing employment for 22- to 25-year-olds in the most AI-exposed occupations fell 16% between October 2022 and September 2025.

What worries him most is lost process knowledge, the unwritten practice passed on through apprenticeship, which takes "the work of a generation" to rebuild. Medicine made the point concrete. François cited a Harvard and Stanford study in which OpenAI's o1 model got emergency diagnoses right about 10% more often than physicians. Doctorow's objection was to the pitch: fire nine of ten radiologists on $300,000 each, split the $2.7 million saving with the vendor, and leave the survivor to sign every report and absorb the blame - what the economist Dan Davies calls an accountability sink. "The question we need to ask is do you get a trillion dollars worth of benefit out of spending a trillion dollars?"

What a burst would leave behind

Doctorow does not forecast that AI disappears. "It's a normal technology," he said, and "the thing that's abnormal is the bubble". Open-weight models already run on laptops, and after a collapse he expects businesses to run or rent ordinary servers for AI work. Those models have become a policy fight, with 77 organisations, Nvidia and Hugging Face among them, signing a letter defending them in July 2026.

The transition would hurt. He likened subsidised AI to running Concorde between Manchester and London every ten minutes for 10p: rail links wither, and when the flights stop nothing is left. The wider danger, he said at the outset, is "a worldwide depression when the bubble bursts", then austerity and a turn to the far right. His answer for individuals was organising rather than consumer choice. "Shopping is not politics," he said, dating John D. Rockefeller's loss of his empire to 1912; the Supreme Court ordered Standard Oil's break-up in 1911.

Why this matters for the marketing community

The companies whose spending Doctorow calls a bubble are the ones selling most of the world's advertising. Alphabet funded part of its buildout with an equity raise of approximately $85 billion in June. S&P Global Ratings cut Oracle to BBB- in July over a forecast $42 billion free operating cash flow deficit. Analyst Ian Whittaker has argued that advertising growth forecasts carry an unpriced dependency on data centres getting built, and Zitron calculates that about 190 gigawatts of planned capacity would need more than $1.6 trillion in annual revenue to pay off.

Advertising is also meant to fix the unit economics he describes. OpenAI confirmed plans to test ads in ChatGPT's free and Go tiers on January 16, 2026, and has reportedly set an internal target of $2.5 billion in 2026 advertising revenue. If each free user costs money, as Doctorow contends, advertisers are expected to cover the difference. Nothing in the interview tests whether they can.

He is not the first to reach for the word. Jeff Bezos called the period an industrial bubble on October 1, 2025, and a 2024 Yale analysis listed five traits the AI market shares with earlier technology bubbles. Bezos separated valuations from business fundamentals. Doctorow's argument is that the fundamentals are the problem.

Timeline

Summary

Who: Cory Doctorow, author and Electronic Frontier Foundation activist, interviewed by Myriam François on The Tea with Myriam François, an MPWR Productions show. Companies discussed include OpenAI, Anthropic, Nvidia, Hugging Face, Google, Microsoft, Apple and Meta.

What: A 93-minute interview in which Doctorow argued that AI is a bubble: gross sector revenue of about $50 billion a year against spending in the trillions, unit economics that worsen with each user, rapid hardware depreciation, no customer lock-in, and 35% of US stock market value concentrated in seven companies. He also disputed the significance of a reported OpenAI agent breach of Hugging Face, argued that the US holds a software kill switch over allies, and said AI is being used to cut and discipline labour. Several figures in the episode diverge from data PPC Land has reported.

When: The episode premiered on YouTube on September 11, 2026, and had 148,606 views by September 27, 2026.

Where: Published on YouTube on the channel of Dr Myriam François, with discussion covering the United States, the United Kingdom, Ireland and the European Union.

Why: The platforms whose capital spending Doctorow calls a bubble are the largest sellers of advertising, and OpenAI is building an ad business to fund free usage. If his reading of the economics is right, a correction would reach ad budgets, platform pricing and marketing jobs.