Why this is in the vault
Packy's argument that the three standard "AI is not a bubble" claims are exactly steps 1-3 of the classic capital/technology cycle (railways, fiber, oil, autos), and that moats decide who profits - a clean historical frame for the investing capital-cycle thesis.
The core argument
- Three anti-bubble arguments are the cycle's opening steps. AI is useful and adopted, revenue is real and growing fast, labs can be profitable. Granting all three describes the setup of a ten-step cycle: miracle tech, early profits, bottomless demand, capital and competitors flood in, everyone overbuilds, prices collapse, capital flees, a low-cost survivor consolidates. Steps 8-10 often but not always play out. He cites Perez and Chancellor as precedent.
- Historical cases. Airlines (Buffett's line about Orville and capitalists), British railways 1840s (incumbents invested "in self-defence"), 1990s fiber (Qwest, Global Crossing, Level 3; bandwidth prices fell roughly 70% a year), 1860s oil (Pithole, Barnum's skepticism). Winners were cost-and-scale consolidators: Rockefeller, Ford, Vanderbilt.
- Consumers capture the surplus. Nordhaus: producers captured about 2.2% of social surplus from innovation 1948-2001. Investors in obvious-demand categories tend to overestimate appropriability.
- Exceptions. Cornered resources (regulation, patents), non-obvious demand (AWS), and long-compounding process power (ASML). Where you can just throw money at the problem, the jungle applies.
- Anthropic IPO. He declines to call a bubble or a top. Notes Anthropic's S-1 compute commitments of about $900.9B through 2045 against Dario's earlier "off by a year, you destroy yourselves" remark; bull case is self-improving models winning every market. Explicitly "no idea" which outcome.
- Thesis line. "The more miraculous your technology, the more crucial your moats." He favors Vanderbilt/Rockefeller/Ford types: cost and scale advantages that can plug in better technology rather than be threatened by it. Closes with a personal "psyched as consumer, fraught as investor" stance.
- Verification note: figures (S-1 commitment, price declines, market caps) are Packy's citations from the newsletter body, not independently checked here.
Mapping against Ray Data Co
Strong fit with the investing capital-cycle work ([[project_investing_markov_capital_cycle]]): the ten-step cycle is a ready-made phase-labeling template for AI infrastructure, and "position horizon, not day-trading" matches his survivor-consolidates framing. It also bears on the founder's phData-main bet and how durable any agent-services edge is when the underlying model is commoditizing: advantage has to come from something other than access to the miracle (workflow, data, distribution), consistent with the moat notes already filed. Caution: the essay is a framework, not a signal - it explicitly takes no position on Anthropic's price, so it should not feed any trade or allocation proposal directly. Relevance to RDCO as a business is indirect (no action item).
⚠️ Sponsorship
Paid sponsor block: Reducto (document parsing, r-1 mini, $1 per 1,000 pages, up to $5,000 migration credits). Standard ad relationship, no connection to the argument. Structural bias: Packy is an investor (Not Boring Capital) and ends with a stated preference for certain company types and a link to companies he likes; no specific holdings disclosed in the body. Treat the portfolio-adjacent tilt toward "cost-and-scale" physical-world companies as a perspective, not neutral analysis. Also self-referential links to his own prior essays.
Related
- [[2026-04-10-jaya-gupta-anthropic-moat]]
- [[2026-10-22-stratechery-moffett-spacex-bubbles]]
- [[2026-03-16-stratechery-agents-over-bubbles]]
- [[2026-07-21-commoncog-how-moats-are-built]]
- [[2026-06-03-alphasignal-anthropic-ipo-filing]]
- [[project_investing_markov_capital_cycle]]