The AI Bubble Just Cracked — Anthropic's Stall Exposes the Trillion-Dollar Leverage Bet

01The Stall

Anthropic — the company behind Claude, valued at close to $1 trillion — has hit a wall. According to third-party revenue trackers that estimate spending in real time, the company's annualized revenue reportedly hit $47 billion in May and roughly $74 billion by July. That is still enormous. But the pace is the story: monthly growth reportedly collapsed from around 51% in May to roughly 8% by July.

You can watch the flattening in the weekly revision numbers: $54 billion in late May, $67 billion in early June, $70 billion mid-June, a wobble backward, then $71 billion, then $74 billion. That is no longer a rocket. That is a man on a ladder taking breaks.

A trillion-dollar valuation on a young, money-losing company is not a measurement — it is a promise. It prices in years of perfection in advance. At these prices, the company does not need bad news to be in trouble. It merely needs good news instead of miraculous news. A company priced for a miracle that delivers only excellence has, in the eyes of the market, failed. Anthropic just delivered its first merely excellent quarter.

02The Bet

Anthropic's curve matters far beyond one company, because its valuation is the reference price for the entire AI complex — chip makers, cloud giants, data-center lenders, and the index funds sitting inside ordinary pensions. When the poster child's curve bends, the repricing travels.

The reported economics of the boom: somewhere between $3 and $4 trillion has poured into the American AI industry — data centers, chips, power deals. The majority of that is reportedly debt, not equity. At a normal corporate bond rate of 3–4%, the interest bill on $2.5–3 trillion of debt runs to roughly $100 billion a year, every year, just to stand still.

The industry therefore needs around $100 billion of annual profit before a single promise is kept. Even at a generous 10% profit margin — which the industry currently does not have; most major labs reportedly lose money on every request served — that requires roughly $1 trillion in annual revenue.

Where is there a trillion dollars a year lying around? Not advertising — too small. Not chatbot subscriptions — far too small. There is exactly one pool big enough: the roughly $10 trillion American white-collar wage economy. To make the math work, AI must profitably capture about a tenth of it — the equivalent of replacing on the order of 10 million white-collar workers every single year. Not as a side effect. As the revenue target.

03The Reckoning

Three forces are now colliding with that business plan. The first is the death of the monopoly play. The American labs run closed models at prices that lose money deliberately — the classic tech playbook: subsidize below cost, kill the competition, then jack up prices and harvest. Except China never agreed to the script. Chinese labs, with a fraction of the compute, have produced open models comparable to flagship American systems — and open means downloadable, free, un-meterable. According to OpenRouter routing data, Chinese open models now reportedly account for more than 60% of all tokens used by American firms. You cannot drive a competitor out of business when the competitor's business model is not charging money. The jack-up-the-price ending is off the table.

The second force: the intelligence escaped the building. Through distillation and quantization — shrinking the brain on purpose — those giant open models compress into local models that run on a decent desktop. No subscription, no meter, total privacy. Compressed models like Qwen reportedly perform on par with the American flagships of barely a year ago. Last year's $100-a-month product is this year's free download. The everyday tasks that make up most of what anyone actually asks an AI — emails, summaries, boilerplate code — have fallen out of the paid tier without ceremony.

The third force is the biggest: replacing humans turned out to be really hard. Today's AI is a probability machine — astonishing at remixing what exists, incapable of true reasoning about what does not. Getting reliable work out of it takes constant human supervision, which is precisely the component the business plan needed to delete. The field results are in: a survey of thousands of companies reportedly found over 70% of deployed AI customer-service agents had to be rolled back or shut down. Some firms laid off human reps early, watched the AI faceplant in front of paying customers, and frantically hired the humans back. Across the whole economy, the most optimistic estimates put genuine AI job replacement at under 100,000 workers a year. The plan requires 10 million. That is off by a factor of 100.

04The Behavior

When the math stops working, watch what companies do, not what they say. The AI industry is now doing two things at once: sprinting toward stock-market listings at some of the most aggressive valuations ever attempted — dream-priced offerings aimed squarely at retail investors — while simultaneously reportedly asking Washington for loan guarantees and taxpayer backstops. "We are a historic investment opportunity" and "government, please catch us" do not belong in the same sentence. Companies confident in the future do not need catching.

This is the classic final phase of every bubble. The insiders head for the exit, and the exit is shaped like a first-time investor on a trading app. The dot-com era ended with taxi drivers and dentists holding stock the founders had exited a year earlier. The pattern is old enough to have a pension.

05What To Watch

Three signals from here. One: the next revenue reading. If the tracker pace holds around 8% a month or slips, the stall is a trend, and trillion-dollar promises do not survive trends. Two: the language. Every time "jobs will vanish" becomes "augmenting workers" in a keynote, that is the sales forecast being cut in public using vocabulary as the press release. Three: the listings. When the biggest names sprint for the public markets while growth visibly slows, you are not watching a coronation — you are watching a handover.

Source: Analysis adapted from Meerkat Explains — "The Claude Situation Is a Total Sh*tshow..." (July 2026). Revenue figures cited are third-party tracker estimates, not audited accounts.