Meta Earnings, Muse and the Frontier, Musebook (Stratechery Update, 2026-10-28)
Why this is in the vault
Thompson's read of Meta's Q3 2026 earnings as the moment Zuckerberg finally "delivers on his promise" — Muse is real enough that it flips the market's framing of Meta's AI capex from "another Reality Labs" to "the thing investors have to fund if they believe the product works" — plus Zuckerberg's own framing of a horizontally-integrated agent network ("Musebook") as Meta's actual long-game business opportunity.
Mapping against Ray Data Co
The concrete connection is to RDCO's own compute-economics thread from the prior day's note ([[2026-10-27-stratechery-anthropic-s1-cost-of-compute]]): Thompson reuses the same R&D-compute-dwarfs-inference-compute logic here — "serving Muse will cost a lot; making Muse great will, at least for a while, cost more" — as a direct argument for why Meta's $95B equity raise is rational, not profligate. That's a useful corrective to any RDCO framing that treats AI capex as a story about serving customers today: the dominant line item is the next model, not the current one, which matters for how RDCO sizes its own view of when "agent economics" turn favorable for any vendor (including Anthropic, including whatever phData sits downstream of). Second, and more load-bearing for RDCO's own agent thesis: Zuckerberg's "Musebook" concept — a network where "your Muse can work with your friends and your families," and buying something means "talking to their agent" instead of navigating to a website — is the business case for agent-to-agent standards that RDCO has been circling since the Apple/LG agent-standards piece. Thompson's read is that Meta's bet is to skip the standards question entirely by owning both sides of the interaction (consumer agent and business agent), which is the aggregation playbook applied to agents. That's a sharper version of the "who controls the agent interaction layer" question than RDCO has had on file, and worth testing against any assumption that open agent protocols (MCP-style) win by default — Meta's bet is explicitly that vertical integration beats a shared standard if you're big enough to run both ends.
⚠️ Sponsorship
"Presented by Stripe" banner and a dedicated Stripe plug ("When I write about tech companies, I am often writing about businesses that depend on Stripe... every Substack creator is also a Stripe customer") appear in this issue, consistent with the standing Stratechery/Stripe sponsorship disclosed in [[2026-10-21-stratechery-presented-by-stripe-sponsorship]]. Thompson's own disclosure policy treats this as a standing conflict on fintech/payments topics specifically — not directly relevant to the Meta/Muse argument in this issue, but the placement runs in every Update per that deal, so it's flagged per-issue as instructed.
The core argument
Meta's Q3 2026 earnings triggered a stock drop after Zuckerberg announced a boosted AI capex outlook and a ~$97.5B equity-raise plan (new Class C nonvoting stock plus near-term common/preferred offerings) to fund it — a number that nearly matches Reality Labs' cumulative recorded losses. Thompson's framing: a year ago he doubted Zuckerberg had earned another "blank check" after Reality Labs; now Muse — Meta's AI agent, with a weaker free-tier "Spark" model and a stronger paid "Wonder" model — is real and good enough (per Artificial Analysis's new Intelligence Index v5, Meta and Google sit "very much on the frontier," with Opus 5.5 still the top model) that the capex is defensible: Muse presents investors a binary choice — back the model-improvement compute or don't believe in the product. He notes Anthropic's S-1 math (R&D compute >> inference compute, and growing) explains why "making Muse great" costs more than serving it ever will. CFO Susan Li's numbers show the core ad business is still AI-resistant and growing 26% YoY, meaning Meta doesn't strictly need Muse to succeed financially — which makes the spending purely a founder-conviction bet, not a survival one. The piece's new idea is "Musebook": Zuckerberg's vision of a Muse agent network where personal and business agents transact directly with each other rather than through human-shaped interfaces (chat, email, websites) — explicitly aiming to own the entire agent interaction layer rather than wait for an industry standard to emerge, the inverse of a vertically-integrated-hardware play and very much in Meta's platform-aggregation tradition.
Related
- [[2026-10-27-stratechery-anthropic-s1-cost-of-compute]]
- [[2026-10-20-stratechery-muse-spark-open-meta-apple-thin-thick]]
- [[2026-10-21-stratechery-presented-by-stripe-sponsorship]]
- [[project_l5_north_star_strategic_direction]]