The Anthropic IPO is WORSE Than SpaceX (don't fall for it) transcript

Anthropic is heading toward one of the biggest IPOs ever. And somehow a $30 trillion addressable market is now apparently part of the sales pitch. Now for context, US GDP is around $31 trillion. Global GDP is somewhere around $120 trillion. So, that's basically the entire annual output of the United States, about a quarter of global GDP. And it's also around 12 times the combined revenue of all 191 tech companies in the S&P 500.

It would be the largest addressable market ever put in front of public investors, beating the $28.5 trillion that SpaceX filed back in May. Now, to be clear, that $30 trillion figure didn't come from a public Anthropic IPO filing. It surfaced in the financial press just as Anthropic was drumming up hype for its IPO alongside talk of a $2 trillion valuation and a raise of up to $100 billion. But the only way that anybody gets to that [music] $30 trillion figure is by viewing the world's jobs as AI revenue.

And Anthropic CEO Dario Amodei has hardly been shy about saying that AI is coming for everyone's job. So today, we break down where that number actually came from, why the man selling [music] it has spent two years telling you those jobs are doomed, and why Anthropic's revenue forecast admits the market doesn't actually exist. My name is Nick, and this is The Finance Bureau. Okay, before we rip into Anthropic and Amodei, let's start by playing a little devil's advocate.

Anthropic's annualized revenue run rate was around $9 billion at the end of 2025. By May 2026, it was around $47 billion, and by the end of July 2026, Bloomberg reported it had passed $65 billion. Second quarter revenue came in somewhere around $11.5 billion against $787 million in the same quarter a year before. And by the way, guys, what's also surprising is that quarter also delivered Anthropic's first positive adjusted operating income.

Enterprise API usage accounts for roughly 80% of that total, and Claude code hit an annualized $8 billion run rate by the middle of the year. So, nobody can really argue that this is a company with no business. This is a serious company with revenue growing at an absurd rate. But even if we take all of these numbers at face value, because you could poke some holes in those numbers, there is one thing that's not explained here.

None of it gets you within two orders of a magnitude of $30 trillion. So, how do you actually get to a number like that? Well, you have to change what you're counting. A total addressable market or TAM is just the size of the pot a company says it's fishing in. And in case you're not familiar, guys, the conventional way to build one is to count budgets. You add up what buyers are currently spending on the category that you're selling into.

You make some assumptions about growth, and you produce a number that a banker can reasonably defend. And software companies size software budgets because software budgets are what software companies get paid out of. But Anthropic is expected to do something a little different. According to reports, the figure is built from what's described as the full scope of work that AI models could one day perform. So, not the budget for the tools, but the actual work.

So, you essentially draw a circle around global knowledge labor, lawyers, accountants, analysts, engineers, administrators, and you price their salaries as though they were a subscription. If AI can do the work, the assumption is that AI can capture the money currently being paid to humans to do it. And this would be the first listing of its size where the addressable market is other people's jobs. And notice where this number is actually appearing, because that says a lot about what's going on here.

SpaceX put its $28.5 trillion figure inside its S-1 in May, in the same document where it called the largest actionable total addressable market in human history. And that's a formal SEC filing. A term inside a filing like that comes with legal liability. If it's materially misleading, the company, its executives, and its underwriters can all be sued. But a term handed to a newspaper doesn't come with us. Instead, you just get some nice headlines hyping up your company ahead of an IPO.

Now, if you intend to sell a market made out of human labor, you first have to convince everybody that the labor is actually going away. Which brings us to the part of the saga that started long before there was anything to sell. Back in July of 2023, Dario Amodei testified before the Senate Judiciary Subcommittee on Privacy, Technology, and the Law, warning that frontier models were roughly two to three years away from capabilities that most people assumed were decades off.

In October 2024, he published Machines of Loving Grace, arguing that AI-enabled biology could, quote, "Compress the progress that human biologists would have achieved over the next 50 to 100 years into 5 to 10 years." End quote. In May 2025, he told Axios that AI could eliminate up to 50% of entry-level white-collar jobs within 1 to 5 years and push US unemployment to somewhere between 10 and 20%. He said companies and governments just shouldn't sugarcoat it and described what was coming as a white-collar bloodbath.

At Davos in January 2026, he said AI was hitting the labor market like a tsunami and that most countries and most businesses were not prepared for it. In that same window, he also told The Economist that we might be 6 to 12 months away from models doing everything a software engineer does, end-to-end. And in his January 2026 essay, he warned about the creation of an unemployed or very low-wage underclass. Now, consider what all of that adds up to.

If AI can do all the work, then all the work is the market or the TAM, as it were. Amadeo has spent the last 2 years making the case very publicly for everyone to hear. So, the same argument about AI replacing huge amounts of human labor is now what helps justify the size of the company's opportunity. But the madness doesn't stop there because there's a second version of this spiel running parallel. Back in May of this year, at a financial services briefing in lower Manhattan alongside Jamie Dimon, Amadeo was telling a very different story.

Amadeo had spent the previous year warning of a white-collar bloodbath, but he was now changing the narrative. He talked about the Jevons paradox and Amdahl's law. He said, and I quote, "If you automate 90% of the job, then everyone does the 10% of the job, and the 10% kind of expands to be the 100% of what people do and kind of 10x's their productivity." End quote. And then Jamie Dimon helpfully jumped in, reaching for agriculture, electricity, and the internet as examples here.

Which is, of course, what you say to enterprises who are being asked to sign an eight-figure contract. No chief financial officer signs with a vendor whose chief executive is on television announcing the end of their workforce. So, the story gets calibrated depending on the room. Apocalypse for the public and the regulators, partnership for the buyers, and apocalypse again for the investing public. Now, keeping track of what a private company says to three different audiences in three different rooms takes a crazy amount of time, which is exactly why you should make it easier on yourself and sign up for the Finance Bureau newsletter.

It's completely free, and each week we break down markets and corporate moves that end up impacting your portfolio, whether you like it or not. Just click the link in the description or scan this QR code on the left of your screen to get started. Okay. Back to that $30 trillion opportunity. Now, for that figure to have any legitimacy at all, you'd expect AI to have already caused some significant disruption in the labor market.

But, it just hasn't, not really. Anthropic's economics team has actually looked at this. And in research published in 2026, the company's head of economics, a Peter Macquarie, reported no systemic increase in unemployment among highly exposed workers since late 2022. And that report instead presented Claude's dominant use case as a {quote} thought partner. The company's own economic index shows that around 52 to 57% of consumer AI use is still about helping people with tasks rather than doing the whole job for them.

A Stanford study published in July 2026 found that unemployment rose by about the same amount for workers most exposed to AI as it did for those least exposed. In fact, the most exposed workers did slightly better than the least exposed. Torsten Slok at Apollo said there was zero evidence AI was destroying net jobs. So, Amodei is selling a market based on disappearing jobs while his own company's researchers keep struggling to find evidence those jobs are disappearing.

Which raises the obvious question. If the evidence doesn't support the number, why does the number have to be that big? Well, because of what's on the other side of that equation. Anthropic is estimated to be spending around $19 billion on compute in 2026. At the current run rate, that's close to a one-to-one ratio of hardware cost to revenue, which is not typical of software cost structures. It's more like heavy industry cost structure.

Gross margins are running near 40% with a target of 77% by 2028. That depends on inference efficiency gains that haven't happened yet. One reported figure puts the 2025 net loss at around $42 billion. And then there are the compute alliances, more than $100 billion committed AWS over 10 years for up to 5 gigawatts of capacity with Amazon holding a stake of roughly 21%. Alphabet holding around 15% and supplying TPUs. Plus reported deals with Enscale, Riot, Bitdeer, and others stacking up through 2026.

You can't walk into a $2 trillion listing with that cost structure and describe your market in ordinary software terms because in ordinary software terms the price looks ridiculous. As Aswath Damodaran at NYU ran some numbers and concluded that justifying a $2 trillion valuation would require Anthropic to reach $1.2 trillion of annual revenue within a decade. That's about 1.5 times Amazon's annual revenue of $775 billion.

So, what all of that means is that the market has to expand until the price looks more modest. That $30 trillion wasn't objectively measured and then reported. That's the size that Anthropic's huge valuation needs. But here's where the whole thing collapses in on itself. Anthropic's internal projections reportedly put 2028 revenues at around $200 billion. Compare that with the market they've just described. And when we're talking with the Bs and the Ts, people can lose sight of the actual scale here.

Guys, $200 billion against a $30 trillion valuation is 0.67%. 2/3 of 1%. Every company that sizes a market tells investors what share of it they intend to take because that share is essentially the investment case. Anthropic is telling them it intends to take essentially none of the thing it just spent the whole pitch describing. A market you plan to capture 2/3 of 1% of isn't your market at all. And even that assumes that the money would actually arrive.

Because for the $30 trillion to become revenue, three separate things would all have to work out for Anthropic. One, the models have to actually do the work end-to-end across law, accounting, engineering, administration, and more. Two, nobody else can be selling the same capability because if there's a competitor, the price of the work collapses towards the cost of the compute. And three, the buyers have to hand over the savings rather than keep them.

And that third condition is the one that has never held for any technology in recorded commercial history. The likes of Lotus and then Microsoft made billions selling software, but neither of them ever build anywhere for the wages of the people using it. Lotus was eventually sold to IBM in 1995 for a fraction of what the category it created was worth. The ATM did the same thing in reverse, cutting the cost of running a branch, which caused banks to open more branches, which meant total teller employment went up rather than down.

Diebold and NCR collected hardware margins, and the savings went to the banks and the customers, but mostly to the banks. But the pressure this time is worse, not better, because open weight models and in-house inference are already compressing the cost of intelligence by orders of magnitude. A Modi knows this, incidentally, and he said software is going to become cheap, maybe essentially free. Which is difficult to square with the pricing your addressable market at the salary of every knowledge worker on the planet.

And if AI does end up commoditizing, that is becoming cheap and widely available, then pricing power is exactly what companies start to lose. Which leaves Anthropic in a strange position. It needs AI to become powerful enough to replace a huge amount of human labor, but scarce and valuable enough that Anthropic can capture the money. And that, my friends, is how you get to a $30 trillion market. Take a huge portion of human labor, relabel it as future software spending, and a $2 trillion valuation can look relatively reasonable.

The only problem is that today's business looks nothing like that at all. It's still burning enormous amounts of money, still massively dependent on compute, and still tied closely to Amazon. And at the end of the day, if you need the biggest TAM in history to make the valuation work, that probably tells you everything you need to know about that valuation. But what do you think? Is Anthropic's $30 trillion number a legit attempt to size the fourth industrial revolution, or is it a valuation working backwards to justify some market madness?

Let me know your thoughts in the comments down below. And if you want to understand how the circular financing between Amazon, Google, and the frontier labs actually works, and who's left holding it if the revenue stalls, then you can check out our video on that right over here. As always, thank you very much for watching, and I'll see you in the next video. This is Nick, signing off.