Anthropic, AI and The “Numbers” Problem

About a week ago I got a ping. Someone wanted to know if I knew someone who wanted $10 million of Anthropic common stock as a forward contract at $1 trillion. My first reaction was that we are so deep in a bubble that when we look up, all we see are sparkling, endless orbital data centers.

If you look beyond the bubble argument, you start to realize all the numbers — including what Anthropic has itself revealed — are so large and so ridiculous that you cannot tell whether anything is real. Unless we get to the SEC filing for an Anthropic IPO, buying common or contracts at $1 trillion today is buying the press release, not the financials.

Or you are believing that professional investors pumping money into Anthropic are doing their due diligence. My experience says they are not. There is nothing more dangerous than an investor punch-drunk on FOMO.

Still, that late-night ping came just ahead of the news that Anthropic was closing a round at a pre-money valuation above $900 billion. Not quite a trillion, but close. There are enough secondary buyers and sellers trading the common at an implied trillion. It does not matter what the real value of the stock is, or what the underlying fundamentals or logic are. The headlines dictate the momentum.

Even though I love Claude and what it does for me (for $200 a month), I have a tough time believing the numbers out of the company. I see Dario Amodei as more righteous than his peers, but equally disingenuous. Telling the US government or the Pope about mythos, AI, and its ability to cause havoc is just confessional cosplay. Sam Altman could learn a thing or two from him.

When it comes to Anthropic, the punters are betting on the bull case, meaning revenue. The company says its annual run rate has gone from about $9 billion at the end of 2025 to over $30 billion this spring, driven mostly by Claude Code and enterprise API. Sources close to the company put the current figure closer to $40 billion. Roughly quadrupling. So we are being asked to buy Anthropic at somewhere between 22x and 25x that number. That is the number every investor in this round is believing and tying their hopes to. It is also the number nobody outside Anthropic can verify.

Every bubble argument I have read this year is fighting on the wrong terrain. Capex-to-revenue ratios. Circular financing. Depreciation cliffs. The numerator is doing all the work. The question is not whether 25x revenue is too much. The question is whether the revenue is the revenue.

Run rate is a press-release number.

I would believe it only after it shows up in SEC filings, and there too I am expecting part fact, part fiction. Just like the SpaceX IPO filing.

I would love to know when gross differs from net, or get a better estimate of the retention curve, or a disclosure of how much is paid cash versus burned credits. I would love to know how much is related-party. A company picks a recent month, multiplies by twelve, hands it to a reporter, and the reporter writes it down because there is no other number on offer.

A $40 billion run rate with 40 percent net revenue retention is a different company from a $40 billion run rate with 130 percent net revenue retention. From outside, the two are indistinguishable. The valuation gets the same multiple in both cases. The first company is worth a fraction of the second. This is not a small problem. It is the entirety of the problem.

Past bubble debates were arguments about multiples. In 1999 you could argue about whether 20x revenue was too much for Cisco, but Cisco’s revenue was Cisco’s revenue. Audited, breakable into segments, attached to a profit-and-loss statement that lived in a 10-Q. SaaS in 2021, same thing. ARR was a soft number compared to GAAP, but at least the underlying contracts had a recognizable shape. The fight was always the multiple. The base was the base.

Whether it is OpenAI or Anthropic, revenue cannot be resolved without an S-1.

A fortyfold expansion in eighteen months. There is no historical revenue cycle this resembles. Frontier-model demand is too new to model and too entangled with its own suppliers to read cleanly.

I also think the circular-financing framing is wrong. It is an easy analogy with Lucent and its vendor financing during the broadband build-out. The right question is simpler: how real is this revenue? Not many are asking it. We have already seen through Amazon, Microsoft, and Google filings that the math does not add up cleanly.

Anthropic committed up to $100 billion to AWS over time in exchange for Amazon’s investment. It has a parallel arrangement with Google. OpenAI made a comparable commitment to Azure. Some portion of every AI lab’s revenue is the cloud provider paying the lab to consume the cloud provider’s compute. Some portion of the cloud provider’s AI revenue is the lab paying the cloud provider back. Both numbers go into both top lines.

This is where the broadband bubble analogy holds. I spent months trying to unpack the deals between fiber providers, Lucent, and the CLECs, and what Nortel was doing to mask its own numbers. It took a couple of years to see clearly. By then the world had already sorted it out. The rising number of telecom bankruptcies told the story.

The same is true today. The frontier labs. The hyperscalers. Nobody wants to report the real data. They all want to look great ahead of the biggest technological shift of their lifetimes.

On May 18, Anthropic’s enforcement of its long-standing transfer restrictions rattled the secondary market in its shares. The policy itself — voiding any sale not approved by the board — dates to at least February 2026. Publicly traded funds touting Anthropic exposure sold off. Investors who had bought into special-purpose vehicles and forward contracts to get “exposure” learned that some of what they owned was, in Ike Brannon’s phrase, “well-disguised bets on the company’s valuation increasing rather than bona fide ownership of transferable shares themselves.”

This is another layer of opacity. It makes revenue beside the point. Just bet on valuation going up. You don’t even need Polymarket to channel your greed. Anyone and everyone is betting on the valuation — no different from GameStop or other meme stocks, except the scale is insane. Forwards on Anthropic common have become a sub-asset class of their own, with their own price discovery, separate from the company’s own mark.

Welcome to 2026, or 1926. Either way, most will find out that even humans hallucinate. And worse than the machines they bet on.

May 29, 2026.