25/ Reinvent, don't entrench From the founders' letter in the S-1: "We will be constantly seeking to reinvent our business, releasing new classes of products whose aim is to dwarf the economic value of our existing ones." "Relative to a typical business, we will generally devote less effort to entrenching our current products and business models, and more effort to defining new ones, because that is where the technology's greatest possibilities lie." 26/ The viability timeline "AI chatbots became viable around 2022, and AI coding agents around 2025; AI agents for general knowledge work and scientific discovery are gradually becoming viable in 2026." The S-1 describes Claude Cowork as extending Claude Code's agentic architecture to knowledge workers broadly. It also says model versions released in the six months to 9/30 accounted for over 80% of Q3 revenue. 27/ How big they think the market is Knowledge work: >$40T of economic value globally in 2025 (company estimate, built from ILO / BLS wage data) Knowledge workers: >1B by 2028 (IDC) AI's cumulative economic impact: $22.5T by 2031 (IDC); Anthropic believes it could be even greater Pharma spend alone: >$2T in 2026 28/ A country of geniuses in a data center Anthropic believes "Powerful AI" will become widely available within a few years. Its definition: - Matches or exceeds top human experts across most professional and scientific disciplines - Uses any interface a human can, including direct computer control - Works autonomously for hours, days or weeks - Coordinates across many instances, with humans and with other AI systems - Controls lab equipment, robots and other machines The founders describe millions of virtual agents able to transform software, physics, math, biology, medicine, finance and more. 29/ Why the compute bets look so large The founders say Anthropic will often make forward investments, especially in compute, that may look irresponsibly large at first glance. The reasoning: based on the scaling laws, AI's economic value is on an exponential curve, so near-future value will be much larger than current value. Their analogy is pharma: very large upfront R&D that only makes sense in light of later returns. At Anthropic the same process runs much faster and at much larger scale. The hedge: give up some upside for downside protection, e.g. offloading excess compute if demand is overestimated, improving efficiency, or raising capital to ride out temporary slowdowns. 30/ The growth caveats sit right next to the vision The company expects AI capabilities to keep advancing for many years, with demand concentrated at the frontier. It also expects the revenue growth rate to decline, possibly significantly and faster than anticipated, and seasonality to become more visible as growth slows. 31/ Revenue they choose not to take - No ads in Claude, even though the founders acknowledge ads would likely be a large source of revenue - Claude Mythos Preview was not made generally available over cybersecurity risk, and its capabilities were released later as Claude Fable 5. The delay meant forgoing substantial revenue, and they gave away ~$100M in Claude credits to cyber defenders - There may be times they choose to advance more slowly than they could, to make systems safer, even at a competitive disadvantage 32/ The historical analog they chose 2000-2025: global GDP ~$34T -> $118T+ (more than tripling) Top 10 public tech companies' revenue: $68B -> $2.5T+ (~38x, ~11x the rate of global GDP growth) IT spend: ~$1.2T (2006) -> $4.3T (2025), as cloud went from 0 to $1.9T The founders believe the impact of an exponential in machine cognition will be vastly greater than the exponential in computer chips (Moore's law). 33/ On investor expectations Before the IPO, the founders spoke with many leaders who had been through one. The advice they heard from several of them: "never let your company be worth a dollar more than it is actually worth" Their own principle: underpromise and overdeliver. They compare it to how they recruit: an "anti-sell" pitch, designed to put off candidates who aren't a good fit.

11/ Q3 revenue is blank in the S-1, but you can back it out “Revenue for the three months ended September 30, 2026 was 86 times our revenue for the three months ended December 31, 2024” Q4'24 revenue = $197M, so Q3'26 ≈ $16.9B. That lines up with the $16.8B shared at the investor day. 12/ New models drive almost all revenue Model versions released in the six months to 9/30 made up over 80% of Q3 revenue. Old models fall out of use fast. Both the release cadence and the risk of a model being restricted now directly affect revenue. 13/ Christmas matters for the exit ARR Customer spend in the two weeks ended Jan 2, 2026 was ~20% below the two weeks before. A run-rate measured over the last 7 days of the year will be heavily affected by the holidays, so be careful with the Dec’26 “exit ARR”. 14/ The October run-rate is left blank on purpose “Our Run-Rate Revenue was $ as of October 31, 2026” Q3 revenue, liquidity and total compute commitments are also blank. Expect them in the S-1/A, together with the price range. 15/ Anthropic runs on Claude Over 90% of code written by Claude (Sep’26) Code per engineer is 45x 2024 levels The value of internal Claude usage grew 250x+ from Jan to Sep, and since April it has exceeded monthly personnel cost (ex-SBC) TTM revenue per employee went from $1.4M to $6.6M+ 16/ Margins will compress before they expand Compute is bought 12–18 months ahead, and costs hit the P&L before the revenue that capacity will generate. The company expects margins may compress when big new capacity comes online. SpaceXAI alone is ~$3.1B a month from Dec’26.