↵ select ↓ ↑ navigate esc close

marginal revolution read out: The Anthropic IPO

marginal revolution read out ·

Listen: https://serve.buildtall.systems/3a3f09ed299ed607c4f78541da96c798ef6dcd0d4f1b0d04b8fad5be6c0603f9.mp3

Source: https://marginalrevolution.com/marginalrevolution/2026/11/the-anthropic-ipo.html?utm_source=rss&utm_medium=rss&utm_campaign=the-anthropic-ipo


This is marginal revolution read out, and the post is The Anthropic IPO. The author, writing on Marginal REVOLUTION, begins with a personal note: he is bullish on the commercial prospects of AI companies, including Anthropic, and serves on its advisory board, though with zero governance role and zero access to internal information. But he sees today's IPO as a signal of bearishness on extreme superintelligence, despite what any involved parties might say.

The reasoning is simple. If superintelligence were arriving soon, why would a company go public and accept all the extra scrutiny and regulation? Debt would be the obvious alternative, because you would not want to cut others into such an amazing deal. The author adds another layer: true superintelligence would likely raise real interest rates, which raises the return to borrowing today, because you could reinvest later at higher rates if you did not need the money for compute.

Now, debt would be relatively costly for Anthropic. But even so, the author argues, they could have shrunk the IPO and borrowed more at the margin. He cites a Claude estimate that Anthropic would have had to pay about ten to eleven percent on that debt. That leads to his key inference: presumably Anthropic thinks pending advances, whatever you call them, will not yield ten to eleven percent at the margin. To him that seems wise and prudent, but he is taking the estimate under advisement and notes that the average return on American corporate capital is about seven percent.

He adds an addendum: maybe there is a clever three-dimensional chess strategy where Anthropic does the IPO, then borrows lots at a much lower rate, say six to seven percent, and still comes out ahead by busting through a credit constraint. He says maybe, but it strikes him as a rather strained possibility. This is the core of the post, and it sets up the two external links that follow.

The first linked item is titled ten to eleven percent on that debt, which is the Claude share the author references. The author's commentary is the sentence that introduces it: now raising money through debt would be relatively costly, but Anthropic could have shrunk the IPO and borrowed more at the margin, and Claude says they would have had to pay about ten to eleven percent on that debt. The fetched page itself contains almost nothing usable, just Anthropic PBC corporate registration boilerplate in Korean, so the substance here is entirely the author's use of Claude's debt-cost estimate.

The second linked item is titled some sanity from Theo, an X post by Theo Jaffee. The author introduces it simply by saying, 'Here is some sanity from Theo.' Jaffee's post responds to an OpenAI research acceleration blogpost and argues that the evidence does not support the claim that artificial superintelligence is imminent. He notes that agent workdays have grown very jaggedly, swarm sizes haven't grown much, and spending on tokens hasn't actually outpaced spending on human workers, although the growth rate is clear. There also hasn't been an acceleration in research experiments per experimenter, or in internal ECI scores. Most surprisingly, task success rates for tasks under fifteen minutes haven't grown much, and the share of tasks done by agents across time horizons is low and mostly not growing. Jaffee says it is possible Anthropic is much further along in automating AI R&D than OpenAI, and that METR saw this, and we cannot rule out an intelligence explosion. But he argues there has been a huge belief cascade in AI and AI safety that seems largely unwarranted by today's evidence. It is entirely possible, he writes, that full automation of AI R&D cannot be achieved imminently, or that it does not massively accelerate capabilities, or that gains concentrate in easily verifiable tasks and fail to generalize to the messy real world. These are all open research questions. The author's decision to include this as 'sanity' reinforces his own skepticism about radical near-term superintelligence.

Across these links, the theme is uncertainty about the pace of extreme AI. The Anthropic IPO can be read as a market signal that the company itself does not expect superintelligence to arrive fast enough to justify avoiding public markets, while Theo Jaffee's post adds independent evidence that the field's most dramatic timelines may be belief more than measurement. The author's tone is cautious but not dismissive: he remains bullish on commercial AI, but sees the IPO as a reality check.

  1. The Anthropic IPO
  2. ten to eleven percent on that debt
  3. some sanity from Theo