Price the Recursion

Last Monday, Anthropic filed confidentially to go public. The number attached to it was $965 billion, which quietly makes the company worth more than OpenAI for the first time, on the back of a fresh $65 billion round and a run rate headed past $50 billion by next month.

Three days later, the same company published a research note arguing that the thing it sells might soon need a globally coordinated pause.

Read that again. The S-1 and the fire alarm went out the same week, from the same building.

Here's the deal. I have read a lot of risk-factor sections. I have never read one where the company's own research arm says the product line might, if it works exactly as intended, warrant the entire industry agreeing to stop.

The note is called "When AI Builds Itself."

The argument is simple, and it is not hype. AI is now doing the work of building AI, and that work is compounding. Claude wrote more than 80% of the code merged into Anthropic's own codebase as of May. Their engineers ship roughly 8x the code per quarter they did a couple of years ago. On open-ended coding problems, Claude Code's unattended success rate went from 26% last November to 76% in May. The tasks the model can finish on its own stretched from four minutes in early 2024 to twelve hours this spring, and they think multi-day jobs land before the year is out.

Then the part that actually matters. Pointed at optimizing its own training code, the model went from a 3x speedup last May to a 52x speedup by April. Handed an open research problem, agents recovered 97% of the performance gap in a week. The humans recovered 23%.

That is the loop. The model writes the code that trains the model that writes better code. Jack Clark put a number on it: a 60% chance that an AI system fully trains its own successor by the end of 2028. The recursion has a base case now, and the base case is this year.

The flywheel is the bull case and the bear case at once.

This is what makes the company so strange to price. For a decade I have watched investors hunt for the compounding loop, the thing that makes the next dollar of input return more than the last one did. Anthropic just published theirs in a blog post. AI improving AI is the cleanest flywheel anyone has ever drawn. If you are writing the check, that slide is the whole bull case. It is the moat and the terminal value sitting in the same diagram.

It is also, word for word, the risk disclosure.

The exact mechanism that justifies a near-trillion-dollar valuation is the exact mechanism the safety note says could run faster than any institution is ready for. That is not two findings. It is one finding that two desks read in opposite directions. Growth sees a self-improving asset. Risk sees a loop with no obvious off switch. Same chart.

I have never had a deal where the upside case and the downside case were the same sentence.

Implementing usage-based pricing successfully requires more than just a pricing strategy. It requires financial and operational infrastructure capable of handling dynamic pricing models, real-time usage signals, and increasingly complex monetization approaches.

In this guide, you'll learn  ⤵

Download the Guide

You cannot underwrite a pause.

Here is the actual problem for anyone pricing this. Anthropic's recommendation is not "stop." It is conditional: a slowdown would be a good thing if it were verifiable, and only if multiple well-funded labs across multiple countries agreed to it at the same moment. They know that coordination is not coming while everyone else is sprinting toward the same IPO window. So the warning manages to be sincere and inert at once. It changes nothing about the roadmap.

Now sit in the investor seat. How do you discount a cash flow when the company's own scientists assign real probability to a world where the prudent move is for the whole industry to halt? You can't. There is no row in the model for "founder-endorsed global stop, timing unknown, triggered by the product working." So the market does what markets always do with a risk they cannot size. It ignores it and prices the flywheel.

And honestly, I get it. Back in February they quietly dropped the flagship safety pledge, the one that promised safeguards before training the next system, and swapped it for transparency and matching whatever competitors do. The stated reason was that a unilateral pledge "wouldn't actually help anyone" if everyone else kept going. That is not a safety argument. That is a competitive one. The pause note runs on the identical logic. We would stop, but only if everyone stops, and everyone won't, so we ship.

Is the warning a disclosure or a billboard?

The cynical read writes itself. You file your S-1 on Monday, you publish a note on Thursday saying your technology is so powerful it might need to be paused, and the warning quietly becomes the pitch. Nothing sells frontier capability like a credible founder who says he is scared of it. The skeptics are already calling the safety story IPO positioning, and I cannot fully argue them out of it.

But I think the lazy read is the cynical one. The numbers in that note are too specific to be theater. The 52x is real. The 80% is real. The loop is closing whether or not it helps the roadshow. The uncomfortable version is that both things are true at the same time: the warning is honest, and it is also excellent marketing, because the world we built rewards being honestly terrifying. That is the part I cannot get out of my head.

So what do you do with a company that is its own short thesis? You probably buy it anyway. The flywheel is real, the revenue is real, and "this might be too dangerous to continue" has never once lowered a frontier valuation. I will watch this one price like everyone else, telling myself I am being clear-eyed about a risk row I already know the model does not contain.

The base case landed this year. We are pricing the recursion as growth. I am not sure we know how to price it any other way.

— SWEdonym