Anthropic's filing spends more building models than running them, which is the case for Meta's Muse bet
Two free Stratechery items close the week: an episode reading Anthropic's filing for what it says about training costs, a flattering margin and Meta's Muse spending, and the Friday email indexing the rest of it, including a burst Chinese AI bubble.
Two Stratechery items cleared the twenty-four hours to 08:00 this morning. The Friday episode of Sharp Tech is free on YouTube, and the Friday email that indexes the week is free to read. 12 Lenny's Newsletter has published nothing since Wednesday's four-agent benchmark, which yesterday's issue covered. 3
Both items open on the same document: Anthropic's IPO filing, published Monday. The episode spends half its length inside it, and the email leads with what that reading produced.
AI economics
The Numbers Behind Anthropic's IPO, AI Bubble Questions, Meta's Spending Continues — Sharp Tech, with Ben Thompson and Andrew Sharp
Published October 30, 2026, 18:00 Shanghai. Free on YouTube at 1 hour 38 minutes; the Stratechery page carries only the note that the show is now on video. 4
The first half reads the filing, and three claims come out of it.
- The cost that decides who stays is training. Anthropic's filing groups expenses by function. For the six months to June 30 it reports $9.7 billion of research and development against $2.4 billion of cost of revenue. 5
The footnotes assign customer-serving compute to the first line and model-development compute to the second. 5 Serving cost more than doubled from $1.02 billion a year earlier. Research and development went from $1.85 billion to $9.7 billion over the same six months. 5
Thompson reads the larger line as a cost that competition sets. A lab spends to stay at the frontier, and the spending stops when the lab stops competing. 1 So "pace the frontier for safety" and "pace the frontier to make sure our company stays in business" arrive as the same budget. 1
- The headline loss is a mark, and the margin flatters. The $80 billion loss the episode opens with is a real line in the document. Net loss for the half reads $80.7 billion. 1 Inside that figure sits a $72.4 billion fair-value change on convertible notes, derivative liabilities and SAFEs. 5
The mechanism the episode describes: Amazon lent money that converts into Anthropic shares. Each new raise reprices Amazon's stake, and the increase is booked against Anthropic as a loss. 1
The second adjustment runs the other way. Anthropic sells Claude inside AWS, Google Cloud and Azure and pays each provider a fee. The filing's own phrase is "paying a fee to each CSP". 5
That fee sits in sales and marketing, a line that ran $3.8 billion for the half against $239 million a year earlier. 5 Thompson puts the fee near a third of the revenue booked through those channels, so the 85% gross margin leaves the channel cost outside it. 1
- Meta's Muse bet and Meta's spending are one decision. Meta said on October 26 that it plans to raise $95 billion in stock offerings to fund AI infrastructure. 6 Among the things the company said the money supports is "its AI agent Muse". 6
The episode's argument for reading the two numbers together: Muse sits at the interface layer between users and every model. Holding that layer means owning the model underneath it. 1 A company that rents the model hands the provider a reason to want the same interface. 1
Thompson puts it plainly. Believe in Muse and you should be happy Meta is spending money. Want the spending to stop and your real position is that Meta should leave Muse alone. 1
The publisher's chapter list, with start times, all from the episode description: 1
| Segment | Starts |
|---|---|
| Anthropic Opens the Kimono | 00:00 |
| Anthropic's Obsession with OpenAI | 34:40 |
| AI Bubble Questions | 43:00 |
| A Brief 1Password History | 59:50 |
| Muse as the Ultimate Founder Bet | 1:03:17 |
| Elon's Role in the Ukraine War | 1:09:20 |
| Ben Sees You Can See Everything | 1:22:22 |
| Emotional Thoughts on the Brewers World Series Win | 1:31:40 |
The week in one index
2026.44: At Long Last… — This Week in Stratechery, by Andrew Sharp
Published October 31, 2026, 01:00 Shanghai. Free, and an email with three picks and a list of links; it has no audio version. 2
- It opens where the episode ends. Sharp's first pick walks from the filing to Tuesday's Update and to this week's Sharp Tech. His own sentence explains the pairing: the conversation covers "why the SEC forced Anthropic to help us understand Meta's Muse bet". 2
- Its new argument is about China. Sharp China opened on "the climate of fear across the Chinese bureaucracy", with Xi Jinping's anti-corruption campaign running further than his earlier supporters expected. 2
The same episode turns to a fight over Mao Zedong's record and Kylian Mbappé. Sharp reads that as evidence that "rumors of a Chinese century have been greatly exaggerated". 2 Its own title puts the STAR 50 AI bubble in the past tense. 7 A written version sits on Sharp Text. 8
- The rest of it is a pointer list, and the Stratechery pointers are spent. The four Updates and Articles it lists are the ones this daily has read all week. 2 The unread material sits on the other shows: two Dithering episodes, two Asianometry episodes, Sharp China, and Greatest of All Talk. 2
Other items worth seeing
- Kylian Mbappé's Little Red Card, Sharp's written account of the Mao uproar and the Chinese reaction to it. 8
- Five Questions on the Fifth Plenum; Mbappé Hysteria and a Chinese Own Goal; The STAR 50 AI Bubble Has Burst, the Sharp China episode behind that post. 7
- Dithering on the end of the internet and on touchscreen MacBook Pros, and Asianometry on how Linux and Windows ate UNIX and on Silicon Valley's jagged edge. 2
What you can check from outside
The expense table is public, and it carries the four lines the episode argues over. The footnote under it puts channel partner costs in sales and marketing, where Thompson says the cloud fee sits. 5
The cloud arrangement is public too. The filing describes direct contracts with customers made "while paying a fee to each CSP", which is the mechanism behind that line item. 5
Meta's $95 billion offering is separate from the episode and dated October 26. It starts with $35 billion in underwritten public equity offerings, then $60 billion through an at-the-market program. 6
Two things rest on the speakers. One is the estimate that the channel fee runs near a third of the revenue booked through those channels. The other is the reading that a mark against Amazon's convertible note is money that never left the company. The figures behind both are in the filing; the interpretation is Thompson's. 15
What it changes for you
- Reading a vendor's price cut: a cheaper token moves inference cost, which is the smaller line in this business. The training bill decides who is still selling models next year.
- Reading hyperscaler earnings this quarter: use operating income. Investment marks now swing net income in both directions, with Google's SpaceX stake marked down and Amazon's Anthropic stake marked up. Neither movement describes the business underneath. 1
- Vendor diligence: ask what share of a model vendor's revenue arrives through a cloud marketplace, and where that marketplace's fee sits in its accounts.
- Building with agents: give each agent its own password-manager vault instead of pasting keys into a chat window. The tip came out of Muse's new 1Password integration. 1
Where to start
- Thirty-five minutes: the opening segment to 34:40, which is the filing, the cost split and the loss.
- Six minutes: Muse as the Ultimate Founder Bet at 1:03:17, where the spending argument is laid out.
- Five minutes: the free email, if the weekend's reading list is what you came for. 2
References
- 1
- 22026.44: At Long Last… — This Week in Stratechery
stratechery.com
- 3Lenny's Newsletter feed
lennysnewsletter.com
- 4
- 5
- 6
- 7
- 8Kylian Mbappé's Little Red Card — Sharp Text
sharptext.net
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.