He Did Tech PR. Now He Rails Against AI for a Living.
vanityfair.com · by Jack Holmes
On the floor of the New York Stock Exchange, AI’s foremost critic is flouting the dress code. This building is the Vatican of big business, and he’s there to appear on CNBC, where the suits deliver their sermons. But Ed Zitron has arrived at the broadcast desk in a charcoal gray T-shirt.
Right off the top, host Leslie Picker tests him. For years now, Zitron has been pointing out that while AI industry leaders OpenAI and Anthropic might have huge valuations—and might soon stage giant initial public offerings (IPOs)—they don’t actually make much money.
“They wouldn’t be the first with bad financial profiles to go public,” Picker says, alluding to past young tech companies, like Uber, believed to have world-changing potential.
“They’d be the first to be this bad, other than WeWork,” Zitron swiftly replies over the din of the trading floor, citing the super start-up of a decade ago that raised ungodly sums from venture capitalists before imploding after a failed attempt to go public. “And even then, this is so much worse than that. OpenAI burned $20.9 billion in 2025.”
In recent months, this 40-year-old former tech publicist has been popping up on Bloomberg, CNBC, MS NOW, and all over financial YouTube in his increasingly trademark Zelenskyycore. When he arrives, the locals generally seem amazed at his command of the numbers and the logic beneath the biggest economic story of our time.
Read A Billion to One: From the mad scientists to the megadonors, a dispatch on the money, power, and influence of technology.
There’s been a broad consensus for three years now that large language models (LLMs), like Claude and ChatGPT, represent a technological revolution that will transform nearly every aspect of our lives—and, naturally, make a lot of money for the companies involved. As a result, investors have poured money into virtually any firm associated with these technologies. In stock-speak, it’s called “the AI trade.” There are dissidents, like Zitron or Michael Burry of Big Short fame, but they are few and far between in financial and tech media.
At the top of the AI pyramid, there’s Nvidia, making GPUs (graphics processing units, a kind of advanced microchip) that are installed in giant data centers. Those facilities are built and serviced by the businesses one level down in the pyramid: big-tech “hyperscalers,” like Amazon, Meta, Microsoft, Alphabet, and Oracle, plus an array of “neoclouds,” such as CoreWeave and Nebius.
The data centers offer computing power (“compute”) that those firms rent to companies at the bottom of the pyramid—the ones closest to the consumer, like OpenAI and Anthropic. They use that compute to handle your prompt when you ask, “What color was George Washington’s white horse?”
That process of delivering your answer is called “inference,” and it’s measured in “tokens.” Each prompt uses a certain number of tokens. Complex requests—like coding—require a lot of tokens. The more tokens needed, the more money your prompt costs the company that runs your LLM. But most people using Claude or ChatGPT are paying a flat monthly fee, not a fee per token. For a company like Anthropic, that’s created a scenario allowing people to spend $200 for a month of Claude premium but use thousands of dollars worth of compute.
So there are some issues on the revenue side. What about the total costs of this revolution? How much has been spent on “the AI buildout” so far?
“Over $1 trillion now, easily,” Zitron tells me. “There are estimates that hyperscalers will spend another trillion next year. The reason I think these numbers get repeated like they’re normal is, they’re so unreal, we may as well be talking about unicorns or the legendary beast, the griffin.” Now many of these firms are taking on extraordinary levels of debt to cover their sprawling spending plans. Plus, there’s that whole thing with Chinese labs making much cheaper “open-source” models.
Six weeks after Zitron’s trip to the stock exchange, we’re sitting in a recording studio at the iHeartMedia offices in midtown Manhattan, where he often tapes his podcast, If Wall Street is a foreign land, then this is his home away from home. Born in west London, he’s based in Las Vegas after leaving the Bay Area, where he ran his own public relations firm. Now he’s launched a newsletter (Where’s Your Ed At) and the podcast, which Zitron says sees between 750,000 and 1 million downloads a month.
Today he’s hosting the writer Ed Ongweso Jr. and comedian Chloe Radcliffe for a discussion of the more existential stakes of our AI revolution. Zitron is freestyling, gesticulating significantly as he serves up fine-grained statistics with no notes and launches all manner of impish barbs at OpenAI chief Sam Altman and other AI leaders. There are some Briticisms too, plus the occasional apoplectic monologue as all the things he’s read about the business of AI come flowing into the mic like a river. His dark coif is unmoved even as he gets animated, the London accent crackling as he rails against what we’re told is the future.
Zitron was doing PR until earlier this year, but a while back he “started writing on the side because I was not doing well emotionally,” he says. “I was quite depressed for several years. I got way better. Tons of therapy. It’s awesome.” That’s how his new media business was born. Throughout 2023 and 2024, Zitron says, he watched the tech and financial media bear-hug LLMs and got curious. “Why is no one saying how much money they make from this?” he thought to himself. “They won’t stop talking about it, other than saying how much they’re making on it.” He’d never had much interest in AI before, but suddenly he was covering it all the time.
Then last year he realized his limitations. “I’m like, I can’t read this shit. So I just taught myself to read balance sheets and financials. And I’m really lucky that I have quite a few financial types who are very generous with their time—private equity guys, hedge fund guys who read my work and were like, ‘I’d love to get you the rest of the way.’” He grew as a writer and a broadcaster, “and it was quite an emotional journey for me,” he says. “It made me develop as a person. I realized this is what I wanted to do.”
"We may as well be talking about unicorns or the legendary beast, the griffin.”
By now, he’s brash, absolute, self-assured, and occasionally self-aggrandizing, but his expanding audience and busy interview schedule suggest a growing number of people seem to think he’s one of the few who understands what’s really going on.
“Right now Silicon Valley has such profound AI psychosis,” he tells me at a restaurant bar across the street from the studio, the postwork happy hour crowds beginning to build around us. “I think this might break venture capital and destroy large parts of it.”
Whatever you might think regarding the revolutionary potential of this technology, the companies involved are spending huge amounts of money but bringing relatively little in. This is historically a problem in business. In the Q&A below, edited for length and clarity, Zitron makes the case that this will soon be a problem for all of us.
Vanity Fair: The LLM users who cost $20 for every $1 they spend—they’re the worst cases, right? Do we have any data on the average user? Like, is the median Claude user a net negative for Anthropic?
Ed Zitron: We truly don’t.
So it’s theoretically possible that a lot of their users are profitable for them. They just haven’t released that, which they probably would.
They would say it with their whole chest. [Anthropic CEO] Dario Amodei said earlier this year, “Oh, a stylized fact would say that someone who had 50% gross margins on inference would be profitable.” And then he went, “I’m of course not talking about Anthropic.”
Anytime you ask these companies, “Hey, how do you become profitable?” or “Are you profitable?”, they immediately start talking like the Riddler. It’s like, “I have two legs at night and four legs in the day…”
How can OpenAI and Anthropic fix this and become profitable? They’ve tried, in some cases, going from a monthly subscription to charging per token. How has that gone?
Uber burned through their entire annual token budget in [four] months. There are a bunch of companies that now set token limits for their users. And Sam Altman, when asked about it, said, “Yeah, it’s become a really huge issue for our customers.”
When faced with paying the real cost of AI, a lot of businesses that were extremely bullish immediately started talking about austerity measures. Suddenly, everyone went from being like, “I’m gaga for AI, it’s the best thing ever,” to being like, “I’ve got to make some tough decisions, got to start cutting costs.”
And Ramp, which is a credit card company used by start-ups, put out a report saying that people are not paying for Fable, [Anthropic’s] big new model—the one that got banned briefly [by the US government for being too powerful]—because it’s too expensive. That is a cataclysmic problem because it means that they can’t charge more money. They can’t up the cost.
There’s no Ferrari product.
Yes, exactly.
Zitron speaking during Web Summit 2024 at the MEO Arena in Lisbon, Portugal.
Carlos Rodrigues/Getty Images
On the other side, they could lower the cost of inference, so it’s cheaper to process prompts.
If that’s happening, I’m not seeing it. And it’s yet another thing where, if it was happening, they would just say it with their whole chest.
Anthropic just produced a profitable quarter. How did they do it?
If you look in the SpaceX S-1, which is the document SpaceX had to submit to the government before they went public, they put in there that Anthropic had become a customer of SpaceX for compute and Elon Musk had given them a discount for May and June. So for two months of that profitable quarter, they had discounted compute.
In other words, yeah, if you don’t count all the costs, they were profitable.
When they were building Amazon Web Services—now a hugely successful business involving data centers—Amazon ran at a loss for years. Why is this different?
So Amazon Web Services was founded in 2003. Between 2003 and 2015, when it became profitable, Amazon’s total capital expenditures for that entire period were $29.7 billion. And that’s adjusted for inflation. Anthropic raised $30 billion in February and $65 billion in [May]. And Amazon sent $50 billion to OpenAI this year. Like, it’s just not comparable.
Meta and Amazon and Microsoft are giving money to OpenAI and Anthropic via investments. Then OpenAI and Anthropic are turning around and paying them back to rent compute?
Amazon funds OpenAI so that OpenAI can pay Amazon so that Amazon can buy GPUs.
I’ve seen some people naysaying the concerns about “circular financing” and saying vendor financing has been a thing in the past.
Yeah, vendor financing has been really popular in the past. It’s gone really badly in the dot-com bubble. But the big difference here is, Nvidia has been smart in that they are not lending money.
There’s a thing called a neocloud. Sounds fancy. It just means a company that buys GPUs from Nvidia, builds data centers, puts the GPUs in there, rents them out to people.
CoreWeave’s a great example. One of their first investors: Nvidia. One of their first large customers: Nvidia. The company that anchored their IPO: Nvidia. And how does CoreWeave raise the debt to buy Nvidia GPUs? Well, they take the contract from Nvidia, they take it to a bank, and they say, “Look, we have a customer. Can I borrow some money?” And because bankers, I guess, are simpletons, they went, “Oh, very good.”
“Anytime you ask these companies, ‘Hey, how do you become profitable?’ or ‘Are you profitable?’, they immediately start talking like the Riddler.”
I would sell more magazines if I gave people money to buy my magazines. But it also sounds like I’m then going to the bank and raising money off the idea that I have a lot of customers for my magazine.
Yes, exactly. That is actually exactly it. It’s so insane, because you are right now dealing with the situation with more nuance than most of the tech media.
These companies had great businesses before. If this is such a mistake, why are they doing this?
I think large language models are just a dead-end technology that they all jumped on, and I do have a theory. It’s what I call the “rot-com bubble.” They ran out of hypergrowth ideas. They haven’t had a new iPhone, a new Google Search, a new Microsoft 365, or a new Amazon Web Services in decades. In 2022 and 2023, every tech stock was tanking. Everyone was looking real sad then.
Then in November 2022, ChatGPT comes out, and everyone goes, “How did Microsoft do that? They bought a bunch of GPUs. We should all do that,” because tech companies don't have ideas anymore. They just copy each other. They went, “Oh, if we just buy GPUs, stock can go up forever.”
We’re in a very manic phase, and most people don’t realize how manic and ridiculous it is because (a) people don't disclose their AI revenues, and (b) it’s very hard to reconcile with the reality that these ultrarich, ultrapowerful people can be so wrong in such an obvious way.
And now they’re raising debt.
These companies went from being these cash-rich, asset-light, beautiful vehicles to becoming these bloated asset hogs. [They’ve added] $746 billion worth of PP&E—which is just properties, plants, and equipment—in the last four years. Insane. Insane. So much.
You say these things out loud, and a regular person goes, “Damn, that’s insane.” You say it to a financial journalist, they go, “Sounds good to me.” The media fails to hold the powerful accountable. And the people that will suffer will be regular people, people who are dumping into the stock market thinking the number will go up forever. The brutality that will face retail investors, that will face our economy, it’s going to be horrifying.
Maggie Shannon
Do you see this as a successor to the crypto and metaverse and NFT stuff that we saw over recent years?
There is more to the AI bubble than there was to crypto or the metaverse. There is a product. The problem is that large language models, when you remove all of the insane financialization, it’s probably a $30 billion-a-year industry. LLMs have an interesting underlying tech and can do some stuff, but every useful function that you see with an LLM comes from tens of billions of dollars of training expense and over $1 trillion worth of [capital expenditures] at this point.
And the financial media doesn’t just ask, “Hey, how does this pay off?” Not even “when.” How?
In this letter on Twitter that Jensen Huang of Nvidia posted, when it said at the end, “What’s the return on investment?”, he said, “The return on investment is the usefulness of AI.” We are four years in, mate, and that’s the best you’ve got?
What are you doing with your money right now?
Cash. Pure cash. I don’t want to touch the market. I think it’s irrational. And people say, “Oh, the markets can stay irrational longer than you can stay solvent.” The market is a casino now. It is a PR-driven casino.
Last year there were four separate announcements in the space of a month with OpenAI connected to them. When Oracle announced a five-year-long, $300 billion contract with OpenAI that OpenAI cannot afford and Oracle has not built and cannot afford to build, analysts were on the call being like, “We can’t believe this. This is incredible. We’re really impressed.” These are the people we’re meant to trust. These are analysts, and they’re hot and heavy like a teenager seeing their first boob. It was insane. And the stock ran [up] 30%.
“These are the people we’re meant to trust. These are analysts, and they’re hot and heavy like a teenager seeing their first boob.”
You’ve said the music is about to stop at this AI dance quite a few times since 2024. Your critics point that out and suggest you’re the boy who cried wolf. Why should people believe you now when you’ve said it before?
[The year] 2024 is a long time ago, and I think I was naive. I didn’t think that the system would accept such egregious distances between the amount of money spent and the very small amounts of money made and the massive losses. I genuinely thought the world was a better place, and learning that has been illuminating, to say the least. I’ve stopped giving exact time frames. That was a silly thing to do.
The reason to listen to me is, I publish all my numbers. I am extremely specific in the reasons I believe things. And indeed, since 2024, I’ve had to learn a great deal about economics and hardware and software. I’ve had to expand my knowledge, and I’ve published all of that knowledge with firm links and details so that my arguments are clear.
It will happen. It’s just a question of how much money they annihilate.
So you can’t predict when the bubble will pop, but what’s the needle that pops it?
It’s going to be someone not being able to get money or someone not being able to pay someone. Because this whole thing requires perpetual momentum. It requires endless resources, constant new faces and money into the system without fail.
What comes after the AI bubble is actually a little scarier.