Is Anthropic Worth $2 Trillion?
Evan Armstrong — formerly the writer of Every's Napkin Math, now running his own publication The Leverage — co-published this S-1 teardown with Every; Every's Kate Lee edited it. It is a secondary analytical read of the same S-1 covered in the vault's primary-source note ([[2026-10-26-anthropic-s1-teardown]]), not a restatement of it: Armstrong's numbers and framing (NRR, true gross margin, customer-tier power law, "Hollywood not SaaS") largely don't appear in that filing-table note.
Why this is in the vault
An independent analyst's gloss on Anthropic's S-1 that surfaces numbers the raw-filing note doesn't (900% net revenue retention, the 45%-not-85% "true" gross margin, the customer-concentration power law) and frames the valuation bet in terms directly useful for assessing the vendor at the center of the founder's main bet.
The core argument
Armstrong's frame: foundation-model companies behave like movie studios, not SaaS companies — "you are only as good as your next model." Three pieces of evidence: (1) models older than six months earn under 20% of revenue, so the revenue engine depends on a release cadence holding forever; (2) Jevons paradox is structural — every release cuts cost per task 25-45%, so Anthropic must grow usage faster than it cuts price, against a compute cost base contracted years in advance; (3) growth is already decelerating from 144% QoQ (Q2) to ~45% QoQ (Q3), partly real competition, partly a deliberate cutoff of customers violating usage policy (5-7% of revenue, mostly labs distilling Claude).
Other load-bearing claims: Anthropic's headline 85% gross margin only holds if you exclude frontier-training compute from cost of revenue; include it (Anthropic itself describes compute as one fungible pool) and gross margin is 45% — closer to a hardware business than a SaaS one. The 11,900-customer figure understates concentration: the real "relationship count" that matters is closer to 250 accounts above $10M/year, following a clean power-law where each 10x step in spend cuts the customer count ~7x. The $80.7B "loss" headline is non-cash — the real operating loss was $693M, with $72.4B being a mark-to-market charge on Amazon's convertible notes that vanishes post-IPO. Compute spend scales from ~$22B/year today to ~$155B/year by 2028, mostly non-cancelable, against revenue that is consumption-priced and cancelable at will — a long-term-debt-vs-short-term-revenue mismatch Armstrong flags as the real risk, not customer concentration. Verdict: the $2T bet requires believing Anthropic keeps out-innovating everyone, functionally forever.
Mapping against Ray Data Co
Armstrong's two numbers worth carrying forward: Fortune 500 customers spend under 0.1% of total expenses with Anthropic even where they're already paying $100k+/year (the top 10 Fortune 500 accounts spend under 1%), and revenue concentration in the largest accounts (Cursor/Copilot-adjacent) is falling, not rising (36% → 21% of revenue for the top five, Q1→Q3). Both corroborate, from an independent analyst with no stake in Anthropic's framing, the deployment-gap thesis underneath the founder's main bet: capability is not the bottleneck, getting enterprises from pilot to production is. That's the same thesis the primary-source S-1 note already mapped to phData's Frontier Academy angle and to The Denominator — this note adds an outside-observer's version of the same evidence, useful specifically because it isn't Anthropic's own selling document making the claim.
The "Hollywood not SaaS" framing and the 45%-true-gross-margin point are also a sharper caution than the primary-source note carries: if frontier-model economics really are hardware-like (amortizing a perishable six-month asset) rather than software-like, that's a data point against treating Anthropic cert stacking as a stable long-term moat and in favor of the founder's existing read that the scarce asset is production deployment stories, not more certifications.
Related
- [[2026-10-26-anthropic-s1-teardown]]
- [[2026-10-19-ark-invest-anthropic-revenue-records]]
- [[2026-06-11-phdata-anthropic-partner-tier-cert-leverage]]
- [[project_credibility_for_phdata_sales]]
- [[project_phdata_cert_escalator_path]]