Pueo Keffer

The S-1 is out, so circling back to the six things I said I'd poke at. TL;DR: Unit economics are unequivocally better than the bear narrative scripted over the last quarter. Ironically, we have a nice refresh of the AWS debate circa 2010, arguing duration risk vs. margin risk. 1/ NDR by enterprise cohort: No pretty cohort graph, but blended NRR is ~900% (ex Pro/Max, off tiny FY25 base). $100k+ customers went ~200 (FY24) --> ~1,500 (FY25) --> 11,900+ (Sep '26), and revenue from that group is growing faster than the customer count, so this is pure NDR expansion… this shows in customer concentration w/ top 5 customers falling from ~36% to ~21% of revenue in just two quarters. 2/ Compute payback: We can work with the S-1 #’s here! GM % on customer-serving compute went from negative (FY24) --> ~40% (FY25) --> ~85% (1H26). Backing 2H25 out, each half incr. gross profit covered the prior half's training/research compute.. ~1.2x (2H25), then ~3.8x (1H26). Payback went from ~10 mos (FY25) to < 2 mos (1H26). 3/ Inference cost per unit of capability: Unfortunately, no $/token breakdown 😤.. But serving used ~27% of the compute pool in 1H26 (vs ~44% prior year) while revenue grew ~14x. Opus 5.5 cut customer cost per task ~40% "while preserving margins," so serving cost per task fell (at least) as fast. 4/ Rev mix by tier: Not disclosed 🥲. They did highlight that customers typically use higher-capability models, and over 80% of 3Q26 revenue came from model versions released in the prior six mos. That's a bit backward-looking, as the open-source debate has gotten louder since 1H, but still directionally positive for Ant. 5/ FCF at maintenance R&D: Q2 already ran a ~19% adj. operating margin while 1H spend on “non-serving compute” was running at ~$13B annualized. Let’s assume they keep ~25% of it for post-training and refresh (and that the rest is redeployable/cancellable). You see a bridge to a ~40% adj. OI margin. But looking back at #4, ~80% of revenue lives on new models, and fwd compute is already committed (below). 6/ RPO vs compute duration: RPO is $10.8B, ~75% inside 12 months. Compute commitments plus signed leases are ~$589B, with terms out to 20 years, up ~5.6x since December. That's ~55x RPO and ~13x Q2 run-rate revenue. The S-1 discloses this and explains how they are mitigating it via multi-chip fungibility, partner financing, and Anthropic Direct. AWS didn't reach Anthropic's 4-year run-rate until its 15th year, but the IPO debate will fester around whether demand compounds faster than a heaping pile of fixed commitments. Bidding starts at >$1.5T 🚀🚀