What Anthropic’s IPO teaches us about the real dynamics of frontier AI labs
In late October 2026, Anthropic filed an S1 full of controversial figures—1,200% revenue increase, net loss of $14 billion, valuation of $1.8 trillion. The document has become something of Rorschach test as writers and pundits read into it what they want: “Anthropic’s risk factors include the end of humanity,” wrote Fortune (the document doesn’t actually say this), and “93 pages of risk,” reported CBS News.
But let’s set that aside. More interesting, we think, is what Anthropic’s S1 reveals about the nature of the business they now see themselves as in. At the frontier, AI labs face very different dynamics than software or app companies ever did. Labs are building the tools to rapidly disrupt all knowledge work. In so doing, they make their own expensively-trained models irrelevant every six months.
This makes for a punishing, self-imposed, auto-accelerating cycle of increasing capital intensivity. Rory O’Driscoll, Partner at Scale, discusses these dynamics in the latest episode of 20VC.
So, do we think extinction threats are already priced into a stock?
Frontier AI CEOs have stolen headlines and imaginations lately with grim tales of the future. “The world should accept some bad things happening,” Sam Altman mumbled to a reporter at Decoded. "I won't lie to you—there are real dangers," said Anthropic CEO Dario Amodei. And while that leading quartet endorsed the idea of a detente, none have taken real steps toward it. Which begs the question, what good is all this dark posturing? Because, it has now reached a point where the public is defiling AI billboards and states are suing the model makers. Just this past week, Florida’s Attorney General asked the courts to block OpenAI from developing any new models.
“I think it is super interesting to see the attorney general in Florida basically sue to stop OpenAI and Anthropic. They’re saying, ‘You told me in this legal document that this was not safe. Therefore, I’m entitled to believe you. And you should stop.’ This IPO will be a real public participatory moment, in a way that doesn’t happen when a semiconductor stock goes public. Everyone’s going to have an opinion here.”
Launching amidst such a cloud of concern is new territory for tech companies. But perhaps only if you fail to read the S1 for yourself, or fail to realize that the purpose of an S1 is to raise concerns early, to reduce exposure later. Anthropic’s actual S1 text is solemn, and speaks in measured but hopeful terms about humanity’s bright future if they “get it right.”
“They have to spew all those negatives just to cover their ass. Though I’m not sure who’s going to sue you if the world ends. But anyway, people are reading that and saying, ‘Huh, this is the first S1 I’ve ever seen that poses existential risk, and where the product lies to me.’ But they have to do it.”
All criticisms aside, the S1 is historic, both in scope and solidity.
“It was a great S1. And frankly, all the haters can go pound sand. The burden of proof is definitely on the naysayers. I found it great on almost every dimension and I think it’ll be wildly well-received. They didn’t give the Q3 revenue number, but it was great—better than expected—on margins, operating income, even operating income after stock-based compensation. It was great on customer concentration, compute efficiency, cash on the balance sheet. Really, a top-tier performance. Stunning performance in terms of scale and speed. This thing’s gone up-and-to-the-right faster than anything else in human history—in five years you add up to a trillion dollars? That’s as good as it gets in venture.”
Anthropic is devoted to disrupting Anthropic
Once upon a time, there was only SaaS, and it had a playbook: You invested in expensive engineers to build software that you then kept and resold at diminishing incremental costs. Your investment was mostly upfront, in jockeying to control the market. Once you got the flywheel going, you could coast for years, decades even. This is not at all how AI labs like Anthropic must operate.
“Eighty percent of Anthropic’s revenue comes from models that are under six months old. That means the shelf life is only six months. Note that they didn’t give us a gross margin calculation, which they could have done, and put compute costs in one place—but they do have a footnote about how 70-80% of compute costs are for training, not for inference. That tells me that these models are wildly expensive to build. Once you build them, they are fairly profitable to serve. But you only get to serve them for six months. Then you have to invent a new one, which requires another massive investment.”
From one perspective, that’s terrifying. And from another, it’s an advantage—the unstable market careens toward ever faster models. But if you’re Anthropic or OpenAI, you are the one setting the pace.
“If they miss even once, they are screwed. On the other hand, you forever keep yourself in the best position to build the best stuff. They said that very clearly in the S1: ‘We want to build new stuff.’ Very few companies are going to be able to compete at that frontier. Each next turn of the crank takes $8 billion or $16 billion. There are very few organizations that can write a $16 billion compute check. So while you could read that short shelf life as scary, I would argue that it’s an advantage if you’re the leader and you’re willing to be aggressive. You keep it so others have to spend billions and billions just to play.”
Is there a future for open-source models?
Government agencies and big companies are shutting out the open-weight, open-source models coming out of China because they fear data privacy and intrusion. The concerns will only grow more acute as the power balance tips toward bad actors. In October, PHP developers found a critical flaw in the language underpinning 80% of all websites, thanks to AI. Some argue that because of these risks, open-source models are on their way out; that no serious organization will run the risk of exfiltration by China, or sloppy, hackable software. But others argue there will always be demand for doing things cheap and fast.
“The inference market is really big. We added up all the open source inference competitors and they only had $5 billion, whereas the closed-source two-frontier models had $100 billion. As long as the customers of the frontier models see the cost advantage of directing even 10% of their spend to open-source, or can put up with the ‘brain death’ of open source to save money, there’s going to be demand for that.”
And what of the tech workers? Companies are deciding between them and compute
All the tech layoffs are changing in tenor. A few years ago, companies bragged about replacing workers because of AI efficiencies, even when it was clear that for some, it was just opportunistic cost-cutting. But now, the layoffs are becoming more ornate, drawn-out, and choreographed. Microsoft just announced that instead of laying people off, it would extend a buyout offer to 60,000 employees. Those who elect to leave win generous reward packages.
“I think it’s very interesting that Microsoft thinks they can do with 10% fewer employees. It’s possibly true. You do wonder about how much of it is really AI efficiency, and how much it’s just that they never needed them. But I think it’s that these large companies are getting the depreciation bill for CapEx, and are going to have to find it somewhere. Earnings this year are a golden period where people selling semiconductors are booking the gains, but the people buying the CapEx aren’t booking the depreciation. So everybody’s winning. But I think as you look to 2027, that depreciation is going to start hitting Microsoft, Google, and Meta, and they’re really going to be pushing back on headcount. The AI bill is coming due, and this is how it’ll be paid.”
