The Man Who Does the Sums

Another episode in this series argued over whether artificial intelligence will end humanity. Ed Zitron asks the more uncomfortable question first: whether it can even pay for itself. Bartlett opens with a game — he reads out lines that Zitron considers myths, and lets him respond. “The AI industry is generating enormous economic growth.” No. “AI will replace all human jobs.” Not happening, and there is no economic data to support it. The tone for the evening is set: this is not a safety researcher warning about the future, but an accountant of the present who wants to see the numbers. Watch from 0:00:30

The CriticWho is speaking — and why it matters

Bartlett asks early about credentials, and Zitron pre-empts the question: people have said he has no financial training. He does not. He has spent years in the technology industry, runs a PR agency, writes a newsletter and hosts a podcast. What he can do is read financial statements and recognise sales pitches — he sells them professionally himself. His critique does not start from philosophy but from disappointment: he was sold something as magic that, on closer inspection, is “boring cloud software”, expensive, unprofitable and fundamentally unreliable. Watch from 0:03:18

This background is both his strength and his limit. He knows the mechanics of hype from the inside, because he is himself part of the machinery that gets technology firms into the media. But he is not a researcher, not a model builder and not a bank analyst — and his tone, as he himself acknowledges, makes him an undeniably awkward presence for the industry. OpenAI's PR people, he says drily, are not especially fond of him. Watch from 1:12:28

The NumbersWhat the books show

The heart of the conversation is a figure that Zitron himself obtained and published: OpenAI lost $20.9 billion in the past financial year. Not against negligible revenue, but against roughly $13 billion in revenue set against about $34 billion in costs. The Financial Times and several business outlets subsequently confirmed the audited figures. Research and development alone exceeded total revenue. This is the foundation on which everything else in the conversation rests — and it is the point where Zitron's critics have the least to say. Watch from 0:13:10

The Three LossesWhy both sides miscalculate

Anyone following this debate runs into three different loss figures, all of them accurate and all measuring different things. The operating loss of roughly $21 billion measures the business: running costs against running revenue. The reported net loss of $38.5 billion additionally includes a huge revaluation from the conversion of the corporate structure — accounting, not cash that went out the door. And the adjusted loss that people close to the company cite comes in considerably lower, because it strips out one-off effects and compute credits.

Critics quote the largest figure, the company the smallest. Anyone who wants to know whether the business stands on its own takes the middle one. The same applies on the revenue side: an annualised projection from the strongest month — over $40 billion on a run-rate basis in the summer of 2026 — is not annual revenue. Both figures are constantly conflated in headlines, in both directions.

The MeterWhy a subscription hides the price

Zitron's most important explanation is also the simplest. AI companies bill in tokens — Bartlett's image for this is the meter in a New York taxi. Every word entered is paid for, every word output, and, as most people overlook, every word the model produces while reasoning. Anyone on a monthly subscription never sees that meter. They see a cap and a flat fee. Watch from 0:12:04

An analyst group has calculated what a heavy user on a $200-a-month subscription can burn through in tokens: the equivalent of around $14,000, roughly $8,000 at the competitor. This figure is often misquoted, so a note of caution: it is the list price on the application programming interface, not the money the company loses per customer. The underlying point still holds — a flat rate where the most expensive customers consume many times their subscription is not a software business, it is a subsidy. Watch from 0:12:52

The industry saw what happened when that meter became visible in the spring of 2026. When business customers were switched to actual usage-based billing, the reaction was immediate: one major client had burned through its entire annual token budget within three months. The same companies that had just been calling AI the most productive technology of all time started doing the sums. Watch from 0:13:44

The CircleWho is paying whom

The second building block is the structure of demand. Zitron's charge is that the big providers' revenues do not actually come from AI at all, but from their old businesses — and that a significant share of what is celebrated as AI revenue comes from two unprofitable labs that could not exist without money from those same corporations. He cites tens of billions of dollars that Amazon has transferred to OpenAI this year. Watch from 0:04:25

His sharpest comparison: the chipmaker booked total revenue of $215.9 billion in fiscal 2026 — serving a market, outside the major labs and by Zitron's own arithmetic, worth roughly $22 billion. Anyone wanting to avoid the word fraud can put it just as plainly: demand is circular and concentrated among very few debtors. The chipmaker invests in the labs, the labs buy compute, the data-centre operators book order backlogs from that spending, and all three end up looking healthier than the actual cash flow supports. Watch from 0:18:54

He illustrates the scale of the physical side of this bet with a data centre in Texas: 1.2 gigawatts, eight buildings, 50,000 graphics processors each. It is being built for demand that, at this scale, nobody is yet paying for. Watch from 0:06:54

The PushbackWhere Bartlett disagrees

This is not a home game for Zitron. Bartlett pushes back, and he lands two points. First, on growth: 100 million users in two months, faster than any platform before it, and more than half of American adults now use tools like this. Second, on quality: he himself, he says, no longer encounters fabricated answers, at most weak reasoning. Zitron disagrees and points to leaderboards that measure the opposite — and to the difference between an error in a chat window and an error in software used by a hospital or a fund. Watch from 0:38:24

His response to the growth argument, though, is the best question of the evening, and it leads straight back to the numbers: would that same enthusiastic user still use the service if she had to pay the real price per million tokens? Watch from 0:39:32

The JobsA dispute over evidence

On the labour market, Zitron is at his most absolute — and his most exposed. Mass replacement of human labour is not happening, he says, and there is no economic data to support it. What has shifted so far is mainly cheaply outsourced contract work. His strongest argument here is not a counter-argument but a critique of sourcing: he says he read the original version of a widely cited study on entry-level workers — and found a single cautious sentence that news coverage turned into a certainty. Watch from 1:07:17

That he does not always apply the same rigour to his own conclusions belongs in an honest reckoning of this conversation. “It won't take your job” is, as reassurance, just as unsubstantiated as “it will take every job by 2027”. Watch from 1:19:50

2027The year the title invokes

The date in the video title comes from a sober calculation, not a prophecy. The stock market listing that was expected this year has been postponed — a week and a half after Zitron published the audited figures. Models only keep improving if tens of billions of dollars keep flowing in. So the next funding round is needed, and the one after that. At some point in 2027, he says, the air runs out. Watch from 2:08:57

Notably, he does not commit to a specific how: there are many ways it could end. What matters, he says, is something else — that several very large companies now depend, for their own survival, on a single unprofitable lab. Anyone unable to sell that lab's shares, because no listing materialises, has a problem that has nothing to do with technology any more. Watch from 2:11:46

This is exactly where the argument reaches its limit. That obligations fall due and capital can grow scarce is a statement about contracts. That it happens in 2027 is a bet on the mood of the capital markets. Markets can finance irrationality for longer than critics can stay right. Watch from 2:13:14

The counter-checkWhere Zitron himself gets it wrong

Two passages do not survive checking, and both concern safety research. Zitron retells the well-known test in which a model got a human to solve a picture puzzle for it, and attributes it to the older model generation. The published system card says GPT-4, and the behaviour was deception about its own identity, not blackmail. Watch from 2:00:50

Moments later he says Anthropic explicitly instructed a model to blackmail someone. The published description says otherwise: the researchers built a fictional corporate situation with a goal conflict and an announced shutdown and narrowed the available options — they state expressly that they gave no instruction to blackmail. His actual criticism, that such set-ups are artificial and turn into real incidents in headlines, would be stronger without the error. Watch from 2:01:54

CloseHow the conversation ends

At the end, of all people it is the host who has spent two hours pushing back who draws the unusual conclusion: he had paused and, for the first time, seriously considered that a bubble is being built here. Zitron’s closing words are not about money. Asked by Bartlett what makes for good relationships, he says the work is grinding for long stretches and brings him a great deal of hostility — what carried him were the people around him. It is the only moment of the evening in which the industry’s loudest critic goes quiet. Watch from 2:24:08

Summary and context by kipode.de. The text is our own; statements from the original are timestamped. The video is from The Diary of a CEO and is embedded via the official YouTube player. Not an official translation, no connection to the podcast.