Gooaye EP709, 28 Oct 2026: I Thought Models Would Be Commoditised Into Utilities — This Filing Broke My Causal Chain

Contents

1. Models would be commoditised into utilities — I believed that for nearly two years
2. Gross margin went from the low teens to 85% while prices kept falling
3. Two things I still worry about after the good numbers
4. ”Aren’t these AI companies buying revenue, so they collapse when the subsidy stops?”
5. “I don’t own AI stocks, so a crash isn’t my problem”
6. What is still open
7. Worth a look
8. The one thing to take with you

To assert something without having tested it is folly; to rely on it when you cannot establish it is deceit.

—— Han Feizi, “Xianxue” (Warring States period; my own translation)

On Gooaye EP709, 28 October 2026, Mong-Kung Hsieh spent half the episode on Anthropic’s freshly filed prospectus. The figures he read out: gross margin has gone from the low teens a little over a year ago to 85% today, so a dollar spent on compute now brings back twice the revenue it did; and existing customers spend nine times more a year later, against the 1.1 times he says is typical for a software company. He admitted on air that this overturned his own expectation — he had believed models were a fiercely competitive product and price competition would cap the business. The condition sits on the other side: he also read out what the company owes going forward, four to eight hundred billion dollars depending on how you count, so these numbers hold only while the growth rate does. Every figure below is as spoken on the show, and only counts once checked line by line against the filing.

Models would be commoditised into utilities — I believed that for nearly two years

My old chain was tidy. Foundation models converge, the gap between the few that matter narrows, a customer switches supplier by editing one line of config, and so prices fall forever. I even had evidence: the unit price of each model generation kept dropping, and dropping hard enough to make the whole thing look like selling electricity. Follow that through and the conclusion is that these companies grow revenue fast while subsidising every dollar of it, and the money gets made above them in applications, or below them in machines.

That chain sat in my head for nearly two years without a review. Its convenience was its explanatory power: revenue surging meant subsidy, losses meant the cost of the price war, a price cut meant here we go. Every new fact slotted in — and a frame that accommodates everything has usually stopped working.

Gross margin went from the low teens to 85% while prices kept falling

The revenue line he read out roughly doubles each quarter: about 400 million dollars in a single quarter early last year, 11.5 billion in the second quarter of this year, a rise of some thirty-odd times. The core business made money in that second quarter, 970 million dollars. On customers he gave three figures: about 12,000 customers paying over 100,000 dollars a year, against 900-odd a year ago and only a few dozen three years ago; roughly half the Fortune 500 are customers, and what they pay Anthropic is a small slice of their total spending; and heavy users writing code with Claude Code spend more than 120 times what non-users do.

What stopped me was the margin line. A little over a year ago this business ran in the low teens, so a dollar of compute came back as a few cents of revenue. Now it is 85%, and a dollar comes back as two. Prices falling and margin rising, side by side, is where my gap shows up: I wired “unit price falls” straight into “margin gets squeezed” and skipped the cost column in between. The compute needed to serve a dollar of revenue has fallen faster than the price, so each dollar keeps more even as the price list keeps coming down. Hsieh’s version is that he too believed they would end up competing each other down, he just did not expect the runway to be this long.

The other column I had misread is the loss. The show mentions eighty-odd billion dollars of accounting losses, which people online used to argue the company loses money after all. His explanation: about seventy of those eighty billion come from the valuation rising so far. The notes it issued to the likes of Google and Amazon in exchange for compute are tied to the share price, so as the company becomes more valuable those notes become more expensive, and the accounting picks up a larger expense — no cash left the building. To judge whether a company is burning money, look for cash going out the door behind the loss rather than at the size of the net loss line.

The counterweight belongs here too, because it also holds: 130-odd billion dollars of cash and some 1.1 billion a year of interest income does not mean it has enough. The show is explicit that once the compute it has committed to buy is counted, it still has to keep raising.

Two things I still worry about after the good numbers

He told a phone story in the same episode that I think has the same shape. Apple’s folding phone was in his hands about half an hour before he offered it to his wife, because so much of it was unoptimised; she said the optimisation would come later. That evening he watched his kid fold it to about a hundred degrees on a table like a small laptop, watching video on top while scrolling something else below, and he kept it. He also noted that Face ID was mocked when it launched, and now people complain that a fingerprint alone is not enough. He describes himself as someone who changes his mind easily — same with stocks, a position he loved can be cut the moment an event makes him dislike it, and it does not bother him.

”Aren’t these AI companies buying revenue, so they collapse when the subsidy stops?”

If you carry a version of my old chain, swap it for three columns: which way gross margin has moved over the past year, whether a dollar of compute brings back more revenue than it used to, and whether existing customers spend more in year two than year one. All three rising looks nothing like buying revenue — and for the net loss line, first ask whether cash went out or something got revalued.

The limit is on the show as well. He sums up the bear case in one sentence: the results are good, and shareholders can still be diluted without limit. The leases he read out are Broadcom at 160-odd billion dollars, SpaceX GPUs at 120-odd billion (about 3 billion a month from December), Google at 100 billion, Amazon at 100 billion, Microsoft at 40-odd billion. The clause that matters is on the Broadcom side: up to 42 billion dollars of that rent can be paid in notes convertible into stock at a 10% discount to market, and there is a switch in the terms — miss the revenue target and the rate goes to 13% with the conversion price dropping to a 15% discount. In plain words, when the share price falls and results disappoint, the same rent costs more shares and holders get diluted faster. He compares it to Palantir being criticised years ago for handing stock to staff, and says it need not be a problem so long as the company keeps delivering growth. Read that backwards and the failure condition is written right there: the moment growth stops, the dilution and those hundreds of billions of commitments arrive together.

“I don’t own AI stocks, so a crash isn’t my problem”

I used to comfort myself with this one. His line is that if any link in the chain blows up, do not expect the things you hold with low AI exposure to be fine. The reason is purchasing power rather than valuation: staff who joined over the past few years become wealthy once their stock can be sold, and that money turns into consumption out in the economy, while those hundred-billion-dollar leases are each other’s preconditions, so one broken link means others stop getting paid. He then pointed at something already happening — laptop makers are cutting their forecasts again.

What is still open

He does not answer the valuation question. The show mentions the company wanting to raise 100 billion dollars at a valuation of 1.5 to 2 trillion, with listing around the 16th of November as he recalls it, and says whether that is expensive has to wait for more information. His script is a hot stretch from now through the listing, then the familiar headlines about staff selling stock and all the market’s money being drained, and he says he cannot time it closer than a week or two either way. What I am least sure of is how long 85% gross margin holds: it is the most important column in this filing and the one nobody has an answer for.

Worth a look

• Gooaye EP709, 28 October 2026, the source for these notes
• Anthropic’s filing with the SEC — gross margin, customer counts, retention, cash and purchase commitments are all in there to check against what was said on air
• Filings and material announcements from Broadcom, Google, Amazon and Microsoft, for the terms of the compute leases and the convertible notes
• Monthly revenue and shipment data from laptop and consumer electronics makers, for whether the order cuts are deepening
• Historical distributions of net revenue retention in software, for where “nine times” sits against peers

The one thing to take with you

The moment you change your mind, what moves you is a specific image or number, not thinking it through.

Hsieh did not change his view because he reasoned harder about whether models commoditise; he changed it because the margin column went from the low teens to 85%. He did not keep the folding phone because he worked out what folding is worth; he kept it because he saw his kid prop it at a hundred degrees on a table. Something lands on you first, and the thinking follows.

Here is something I have tried that has nothing to do with investing: pick a view you held firmly three years ago and have since dropped — about a family member, about a job you used to look down on, about a kind of spending you used to call wasteful — and write down the specific scene or number that made you let go. If you can write it, you know what it takes to move you. If all you can write is “I came round to it later”, that change was probably made for you by someone else. When I did this, all three of mine came from something I saw with my own eyes, and not one came from something I read.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.