Why analyst margin figures diverge — which margin each source is talking about, verified claim by claim

Status: VERIFIED 2026-08-23 (America/Dawson_Creek 2026-08-22). Attributions below were established by two independent, blind research arms over the same brief — a GPT-5.6 Pro deep dive (broker request pr-20260823T020913Z-d0fe55, 2026-08-23) and a Fable 5 native-search arm (WebSearch / WebFetch / Brave; the Exa index was out of credits on this host and is stated as such in the arm's provenance header) — then merged by the director. Both arms quote the sources' own words with dates; where the arms disagree the disagreement is shown. Labels: STATED = the source states the object or basis in its own words · IMPLIED = the source's own numbers or method imply it, but it does not say so · OUR READING = our attribution, with the reason given. This ladder grades the ATTRIBUTION — which margin a source is talking about — not whether the source's figure is true. Arithmetic in §2 was re-derived independently by both arms. The owner's reading that prompted this page was stated as probable, not as fact; the verdict on it is in §4.

1. Four different objects get called "margin"

The first two are the same unit (a token served) on two procurement bases, comparable only when model, token mix, cache rate, latency target, batch, utilisation, fleet membership and retained serving overhead are held fixed — which is what the two cards do. The third and fourth are different objects. The durable editorial rule this page adopts: never inherit a basis from an analyst's name; attach the basis to the individual claim, and where the source is silent say "basis not stated by the source; shown under both bases."

2. The arithmetic that links the two unit bases

Let C be the lab's own all-in cost per token on owned accelerators and P the price per token. Own-accelerator margin m_own = 1 − C/P. A lessor whose cost for the same capacity is C and who earns margin m_L on its revenue charges R = C / (1 − m_L), so the renting lab's margin is

m_rent = 1 − R/P = 1 − (1 − m_own) / (1 − m_L), equivalently (1 − m_rent) = k · (1 − m_own) with k = 1/(1 − m_L) the rent-over-cost multiple; and in reverse m_own = 1 − (1 − m_rent)(1 − m_L).

Worked at m_own = 95% (P = 20 C):

So the framing "a 95% own-accelerator margin survives a lessor's cut at ~90%" is right: the markup multiplies the small cost side, not the price side. Two conditions the identity rests on, stated so nobody reads m_L as a named lessor's accounting gross margin: (i) it assumes the lessor's relevant cost base equals the modelled owned-capacity cost C — real lessors and self-owners differ in procurement discounts, financing, useful life, utilisation, residual value, PUE and shared overhead, so in practice k is rent ÷ the lab's modelled TCO, and "m_L" is 1 − 1/k, an implied quantity; (ii) the multiple applies only to the cost the lessor replaces — retained serving costs (orchestration, networking, support) should be added separately rather than scaled.

This page's own numbers, read that way: at the page's defaults the rental-inclusive stress case reads ≈58% and the same state owned reads ≈88%, i.e. k = 0.4211/0.1226 ≈ 3.43 — a 3.43× rent/TCO multiple, equivalent to a 71% implied lessor margin or a 243% markup over the page's modelled TCO. Card 1 prints it as an implied spread over modelled TCO, never as an observed lessor margin. (Re-measured from the shipped engine at each release.)

3. The cited claims, attributed one by one

4. The owner's reading, verdict by analyst

So the divergence is not different analysts disagreeing about one margin. It is different sources switching among four objects — a rented-hour token margin, an owned-TCO token margin, a cohort contribution margin, and company or product accounting GM — and the same source (SemiAnalysis) legitimately using more than one. The two cards on the front page exist so that the first two can be seen side by side; the rest must stay labelled as what they are.

5. Corrections to the main page queued by this verification

1. "this corpus does not establish that the 90–95% allegation is Anthropic-specific" (§1 passage on Zephyr) — contradicted by his 2026-06-24 post ("adopt Anthropic/OAI tactics … (95% GM)"); the page should say the allegation IS lab-specific in his words and that its basis is what remains unstated.
2. "Anthropic turned its first gross profit in Q2 (June 2026)" (RAISE update) — the only retrievable text is a session summary; it conflicts with SemiAnalysis's own written record (38% inference GM in 2025, "mid 60s" by 2026-05-27, "Operating Income profitable in 2Q after adjusting for stock-based compensation") and with Patel on the Sequoia podcast ("net income profitable excluding stock-based compensation"). Relabel as an unverified summary-level claim or drop the "first gross profit" phrase.
3. "~30% (late 2025) to ~55% today … ~50% to ~65%" — only the endpoints (~55%, ~65%) appear in the retrievable summary; the starting points were found in no source by either arm. Cite them or drop them (The Information's 33% is a different, "adjusted" cut and cannot stand in for "~30%").
4. The Opus "north of 80%" citation can be upgraded from "quoted-secondary" to the Sequoia transcript, which specifies the object ("for the API price", on an Opus token) and the rent framing.
5. The −94% / 38% / mid-60s series should carry SemiAnalysis's own label ("inference gross margins" — a business line's GM, not a unit figure).
6. Two Zephyr posts in the corpus ("75–80% GM", 2026-06-27) answer a reply about CXMT, not Anthropic; nothing on the page may derive an Anthropic figure from them.
7. The site's key-findings line "TeorTaxes: Serving Opus ≤ $4/1Mt, 90→95% via batch" fuses two posts; record them separately with their bases as above.
8. DeepSeek V4 "70–80%" is an API-product margin with an undisclosed cost basis, not "company-side".

These are queued for the release-notes pass (tranche 5) because they move pinned front-page bytes; the research page above is the record until then.

6. Method and provenance

Brief: im-arc/dives/analyst-basis-brief.md (seven questions; the page's fifteen cited passages and the prior draft attached as content). GPT Pro arm: broker request pr-20260823T020913Z-d0fe55, gpt-5-6-pro, 2026-08-23, full report archived as im-arc/dives/analyst-basis-gptpro.md. Fable arm: im-arc/dives/analyst-basis-fable.md, 2026-08-23, WebSearch/WebFetch/Brave (Exa out of credits — stated), X posts via the fxtwitter mirror with UTC timestamps. Convergence: every verdict above is shared by both arms; the only differences were coverage (the Fable arm surfaced the 2026-06-24 Zephyr post and the CXMT contamination; the Pro arm supplied the InferenceX/Rubin method statements, The Information's chart-note definitions and the PitchBook cut) and the 67.6%-row rounding (Pro: 84.6% at exact inputs). Two X posts and one Morgan Stanley note could not be retrieved by either arm and are labelled accordingly. SP-real-terms-comparisons: no dollar comparison across years is made on this page.