Ed Zitron has one use of AI he will defend without hesitation. At his place in New York he runs a MacBook and a PC laptop off the same monitor, sharing one mouse and keyboard through a program called Synergy. When it stopped working, he dropped the troubleshooting log — the file a program writes recording what it did and where it failed — into a chatbot and asked what was wrong. It told him. "Super useful," he said. Then, without pausing: is that a trillion dollars? No. A two-trillion-dollar company? No.
Zitron, a critic of the AI industry, was speaking on The Diary of a CEO. He told the host he does not really use generative AI himself, beyond a research function in his Bloomberg Terminal for things like pulling analyst estimates for NVIDIA. So how do you know it's bad, the host asked. Zitron said he has put it through its paces: he tried building financial models with it, found an error, and was not impressed again. He also used Claude, Anthropic's assistant, on a broken Minecraft mod his son wanted working. It was awesome, he said — and it still took him half an hour, with the model getting things wrong along the way.
He borrows a line from the software engineer Carl Brown to describe the pattern: it makes the easier things easier and the harder things harder. A small, well-defined script it can produce is genuinely good. "Again, that is not what they're selling it as," Zitron said. "They're not selling it as a useful little tool."
The meter you never see
The rest of the argument runs on cost, and Zitron walked the host through the unit. AI services bill in tokens — roughly three quarters of a word each — and charge per million of them. The host reached for a comparison: the companies have a currency they charge you in, like a New York taxi with a meter. They call it tokens.
There are two kinds. Input tokens are what you feed in: a document, a code base, a question. Output tokens are what comes back — and, Zitron noted, also what the model produces while working out its answer. Ask for the best restaurants in a part of New York, and the model's own musing about how to find them is billed as output too.
One consequence follows directly from the billing unit. You pay when you use the model whether or not you get what you wanted. If it goes through a code base and breaks a pile of things, that work is billed the same as work that succeeds — unless you are on a subscription.
What the subscription hides
Most consumers are on a subscription, and there the meter disappears. There is a monthly price and a rate limit: you can use the service a certain amount, and the companies, Zitron said, obfuscate what that amount was.
He cited the analyst group SemiAnalysis for figures on the gap. On a $200-a-month ChatGPT subscription, he said, a user can burn $14,000 worth of tokens; on Anthropic's $200 plan, $8,000; on the $20 plan, $400. Those numbers price a subscriber's usage at the published per-token rates the same companies charge developers. They are not a measure of what it costs to serve that usage, and Zitron said as much when the host pressed him: he does not know the real ratio, it might be one dollar of revenue against thirty-four of cost or something else entirely, and the companies do not disclose it. "Even in their audited financials, they play funny games with how they categorize things." His own illustration of the position is that as a power user he might cost a provider $1,000 while paying $100, with the difference covered by the company.
Zitron said OpenAI lost $20.9 billion last year, and that the losses follow from letting subscribers burn as many tokens as they like. He also described what happened when the usage became visible to buyers: around March 2026, he said, OpenAI moved larger enterprise customers toward paying the actual cost of what they used, and companies reacted badly. He quoted Sam Altman saying people had a big problem with it. Uber, he said, burned through its entire annual token budget in three months. The moment people had to pay for it, he said, the enthusiasm acquired conditions: it's all great, but it costs too much, so we need to reduce the cost.
There is a supply side to the same problem. Inference — the work of producing outputs — runs on GPUs that have to be switched on and paid for by the hour before the demand arrives. Buy too many and the money is wasted; buy too few and customers can't use the service and go elsewhere. Zitron's reading is that nobody in the chain is making money on it: not the inference providers, not even the companies renting out the GPUs.
The chief of staff who still has a job
The host pushed back from his own experience. He uses agents for things he would previously have asked people to do. His chief of staff used to triage all of his inboxes, sort them and tell him what was in them; that job is now automated.
"You still have a chief of staff, though," Zitron said.
"This is what I'm saying," the host replied. "They're doing other things." He is still hiring, he added, like crazy.
Zitron's response was to shrink the achievement rather than dispute it: "Fairly basic automation." Told the tasks were fairly basic email, he answered that nobody spent a trillion dollars to triage email. Had the industry spent $10 billion and made smaller claims, he said, he would call it good software and move on.
He is similarly deflationary about the word "agentic." An agent, in his account, is a language model talking to another language model with a harness — the surrounding software that runs the loop — on top, sometimes passing screenshots between them. Anthropic's own engineering guidance draws a related line, distinguishing workflows, where code fixes the sequence of model and tool calls, from agents that choose their own steps and tools, and advises adding complexity only when the task needs it, because each extra step adds latency and cost. Impressive scripting, Zitron added, is often just a model writing Python — and Python, he said, is incredible on its own.
Spending ahead of the value, or subsidizing it forever
The host read out adoption figures: ChatGPT reaching 100 million active users within 60 days of launch, against nine months for TikTok, two and a half years for Instagram and roughly seven years for the World Wide Web; and a share of US adults using AI tools in daily routines within three years that took the internet five years and personal computers nearly twelve to reach. In its February 2026 financing announcement, OpenAI reported more than 900 million weekly active ChatGPT users and more than 50 million consumer subscribers — company-reported measures at that date, and two different things: weekly use is not payment.
That gap is where the host located the real question. Are they spending ahead of the value showing up, which is what the companies would argue, or subsidizing users in a way that will never be justified? Technology history is full of firms losing money to take market share, he said, and these firms are also working to bring costs down.
Zitron did not accept the second half. If they were bringing the cost down, they would have brought it down, he said, and it seems to be getting more expensive instead. His simplest formulation of the case: if the services were profitable, and the companies believed they were worth what they cost, they would charge it. Ordinary people would not get a monthly subscription; they would pay what it's worth.
"Would she pay the actual rate?"
The host's sharpest counterexample was personal. When he showed ChatGPT to his fiancée, it didn't really land at first. But she runs a business alone, English is not her first language, and she has to produce a lot of text and copy and a lot of images — images she had been paying a graphic designer to make because she lacked the skills. She would describe the tools, he said, as transformative for her business. What he was hearing from Zitron was that there is no value in them for people like her.
Zitron did not dispute that she means it. He changed the question. Would she pay the per-million-token rate — the actual API price the companies charge developers? "If this was sold at its honest cost," he said, and people were paying two, three, four dollars every time they did something and were genuinely happy about it, "that might be an argument."
Then the host asked what the honest cost would be.
"Oh, God," Zitron said. "It depends on the model."