Anthropic S-1拆解：900% NRR真相与AI算力军备竞赛

Plus Broadcom backstopping Anthropic’s chip debt, Meta’s $95B equity raise, the PHP2Shell flaws, and an IC fight over Instinct at $30B and OpenAI at $1.25T

• This week had one story that mattered more than everything else combined: Anthropic finally filed its S-1. Harry Stebbings, Rory O’Driscoll and I went through what’s actually in it, the one number Rory finds scary, and the number I think is worse that isn’t in the filing at all.
• From there we got into Microsoft offering exits to more than a quarter of its company,
• The bill for AI capex coming due in 2027,
• A Qwen model trying to reach a server in Beijing, and
• A mock investment committee where Harry wanted Instinct at $30B and I wanted to put 20% of the fund into OpenAI.

#1. Anthropic’s S-1: $16.3B in Six Months, $974M of Q2 Operating Income, and $131.7B in Cash

Anthropic made its prospectus public on Monday, filing to list Class A shares under the ticker ANT. First-half 2026 revenue was $16.3B, up 1,300% from $1.16B a year earlier. Most of it came in Q2 at $11.6B, the quarter Anthropic posted its first operating profit of $974M. Adjusted operating margin went from negative 159% in the first half of 2025 to positive 10% in the first half of 2026. Cash and equivalents were $112.4B at June 30, plus $19.3B in short-term investments. Net revenue retention was approximately 900% as of September 30. The scary-looking $80.7B net loss is mostly paper: $72.4B of it is the rising value of convertible notes. The one gap: Q3 revenue is left blank in the filing, though Reuters has reported Q3 revenue of $16.8B.

Rory’s read: It was a great S-1, and the haters and worriers have to go pound sand, at least for now. The burden of proof sits with the naysayers. Better than expected on margins, on operating income even after stock-based comp, on customer concentration, on compute efficiency, and stunning on the balance sheet. You have to work to find questions.

Jason’s read: Be careful with the 900% NRR. It almost certainly includes pilots and trials, so the real durable number is more like 300% to 400%. That’s still the best I’ve ever seen. Within the next two quarters, I don’t see them missing. Add the brand, Nasdaq-100 inclusion, and retail demand, and it’s hard to bet against a strong debut.

On the day-one pop: Rory expects it to be wildly well received. It goes into the QQQ fast; the S&P 500 takes longer because of the profitability requirement. Rory doesn’t think it needs the index lift SpaceX arguably did. People will just want to own it.

#2. Five Years From Founding to IPO, Versus 24 for SpaceX

Anthropic was founded in 2021. Its IPO, targeting a roughly $2 trillion valuation, is slated for November. Bessemer’s Byron Deeter went on TV calling it the greatest investment in the firm’s history, at a firm that dates back more than a century.

Jason’s read: Folks are almost underestimating Anthropic. Five years to IPO versus 24 for SpaceX isn’t just unprecedented. Time value to liquidity still matters. Markups are fun, but getting out of the greatest IPO of our lifetimes in under six years changes how you look at everything else. A seed deal is 18 years to full distributions. Why would I spend my time on anything that’s 10 to 12 years to an IPO? The honest counter: it’s one deal in a fund, and there are still two more years to distribute.

Rory’s read: Every founder and every investor is asking the same thing right now: why bother with anything else? The size, scale and speed mean Anthropic will dominate venture returns for a long time, and everything else looks like more work for less return. All credit to the founding team and the early investors. Greatest investment, biggest check, fastest liquidity.

#3. Rory’s Scariest Number: Most Revenue Comes From Models Less Than Six Months Old

Rory’s read of the filing: roughly 80% of revenue comes from models less than six months old. The S-1 doesn’t break out a gross margin. Compute costs sit in one bucket, with a footnote suggesting 70% to 80% of compute goes to training, not inference. Infrastructure spend climbed to $8.95B in the first half from $2.33B a year earlier.

Rory’s read: Models are wildly expensive to build and fairly profitable to serve, and you only get to serve them for about six months before you have to invent the next one. That is nothing like B2B software, where you build something and sell it for 10 to 15 years. You can read that as terrifying: miss once and you’re done. Or you can read it as an advantage for the leader. Anthropic spent roughly $9B on compute in the first half and will spend far more next. Very few companies can write a $10B to $12B compute check. If the price of each turn of the crank keeps rising, the field keeps shrinking.

Jason’s read: That’s Elon’s point. Most of the neolabs just die after this. They can’t raise the capital to survive the next cycle.

#4. Broadcom Backstopped $63B of Chip Debt. Meta Went to Equity for $95B.

Bank of America, Blackstone, Citi and Morgan Stanley are providing $63B to buy Broadcom’s AI chips and networking gear for Anthropic, with Broadcom providing residual value support to the senior lenders. The senior debt is about $43.5B, or 69% of the total. Separately, Meta launched roughly $35B of public equity offerings and plans an approximately $60B at-the-market program. The underwritten portion is $22.5B of Class A common and $12.5B tied to mandatory convertible preferred. Meta reported a $2.12B cash burn in Q3, its first negative free cash flow in over a decade.

Jason’s read: The worst metric of the week wasn’t in the S-1. It was that Broadcom had to step in and guarantee most of this debt so Anthropic could buy its chips. We just read the S-1. It doesn’t get much better than Anthropic as a credit. And the debt market still needed the chip vendor standing behind it. Combine that with banks finding it harder and harder to lay off neolab debt, and the conclusion is that debt is an exhaustible source here. That’s the one thing to be nervous about.

Rory’s read: I agree the banks have hit their limit on debt. What’s actually happening is the industry is equitizing. Anthropic is about to issue on the order of $100B of stock. Meta is doing up to $95B. Google issued around $80B earlier this year. SpaceX did its IPO. A class of companies has a cheap cost of capital and is willing to take dilution to keep building. So at the macro level, the spend keeps going. Elon’s comments about debt drying up were him kicking competitors while they’re down, from behind one of the world’s most inflated market caps. Very on brand, and you have to respect it.

Rory on Meta and chips specifically: Meta’s core business is one of the best on the planet and throws off enormous cash. This was the first quarter capex outran it, so they went to equity. That’s unrelated to last week’s chip selloff. Since the S-1 dropped, chips have been bouncing, memory is up, Korea was up 10% on Tuesday. CoreWeave scared people two weeks ago, but CoreWeave is a $20B to $30B company. Anthropic is a $2T company that will have around $200B in the bank after this deal, and it’s going to spend it on chips.

#5. Microsoft Offers Buyouts to 60,000 Employees, 27% of Its Workforce

Microsoft offered voluntary buyouts to around 60,000 employees, roughly 27% of its 223,000 workforce, in an October 26 memo from Chief People Officer Amy Coleman. The offer covers level 70 and below in the US, Canada, India and the UK, reaching as high as VP. Employees have until early December to decide and leave December 31. In April, more than 30% of those eligible took Microsoft’s first-ever voluntary retirement program, and in July it announced 4,800 layoffs. Microsoft has forecast $175B in capex for calendar 2026.