Anthropic’s release earlier this week of an S-1 for its planned IPO has become a Rorschach test—bulls and bears project onto it their hopes, dreams, and fears—as well as a test of reading ability, which many people failed.. At the same time, most of what got attention was things that had been known for weeks, through earlier leaks. Markets run on expectations, and differences from them, so it's more useful to focus on the things that people somehow missed or mispriced.
Here are five things that caught our attention:
- Anthropic Is a Bank With a Money-Losing AI Side Gig
- Extraordinary Growth, Uncertain Durability
- Shareholders Supply Capital, Not Control
- The S-1 Describes Anthropic 1.0, Not 2.0
- Growth Slows as the Coding Era Ends
1. Anthropic is a bank with a money-losing AI side gig
Anthropic held $112 billion in cash and equivalents at June 30, plus another $19 billion in short-term investments. It generated $1.1 billion in interest income on that balance in the six months. Granted, this won't get you a gig at PIMCO or Doubleline, but it is massive.
To put in context, Anthropic's interest income is larger than their operating losses of $700m+. It is also larger than many regional U.S. banks. While the interest income provides a cushion beneath operating losses, it doesn't change underlying operating economics. For now, being well-funded is itself a meaningful source of earnings. That benefit will go to zero as the cash gets spent.
2. Expansion is extraordinary. Durability is less clear.
About 80% of first-half revenue came from customers paying by usage—through the API, and so directly exposed to the winds of tumbling token prices. At the same time, more than 80% of third-quarter revenue came from models released within the preceding six months.
These are remarkable growth figures, but they also describe a business in a rapid race against churn, continually renewing its product and trying to earn its costs again. In a world where models are increasingly converging, the ability to charge premium prices on the latest models will continue to decline sharply.
3. Shareholders supply capital, not control
A single Class F share, held by the seven co-founders, carries 50.1% of the vote. Public shareholders will not control director elections. As ever, these sorts of covenant-lite approaches to stewardship fond of among tech companies seem like a great idea in good times, and then a disastrous one when unhappy shareholders want change.
For now, the governance proposition is straightforward: buy economic exposure, leave control with the founders. Anthropic is a drive-share, and investors are along for the ride.
2. The S-1 describes Anthropic 1.0, not 2.0
Anthropic currently leases massive amounts of compute, hundreds of billions of dollars worth, most of which commitments are not cancellable. Further, most of those commitments are to its competitors, like Amazon, Microsoft, and Google, all of whom sell competitive services, and all have a "house take" on their part of Anthropic sales.
The subtext of the S-1 is that this dependency worries Anthropic, as it should. Microsoft's regularly hobbled competitors in its heyday, insofar as they were distributed on its Windows platform. It famously once threatened to cut off a certain Andreessen-run company's air supply, which caught the attention of the Department of Justice.