We track no Anthropic ticker, and by our own rules we cannot: there is no audited revenue, no margin line, no share count. Yet no private company runs through more of the numbers we publish. Its revenue is the largest component of the market pool our SpaceX model prices the model-layer fight against. Its compute commitments are an $83 billion book we decomposed contract by contract in August. And its private marks have already passed through public income statements: a $53.4 billion line in Amazon's June quarter, two-thirds of Microsoft's earnings beat — lines we stripped out in print at the time.
The IPO — whichever of the two press calendars survives — is the single event that reprices all three at once: the revenue gets audited, the compute book gets funded, and every holder's mark gets a public closing price instead of a financing round. Both calendars put that event within about two months. This piece is the map of what it lands on, drawn from things we have already published and verified, before the prospectus prints.
One discipline note up front, because everything below depends on it. Anthropic's official record here is three pages: two financing announcements — February 12 and May 28 — and a June 1 confidential-submission notice. Every other number in circulation — the $2 trillion, the $65 billion run-rate, the $15 billion credit facility, the bank slate — is unnamed-source press reporting, and it stays labelled that way here.
The company record, in full
The February 12 announcement of its Series G: $30 billion raised at a $380 billion post-money valuation, led by GIC and Coatue, with the company stating "our run-rate revenue is $14 billion, with this figure growing over 10x annually" — plus a segment detail nobody quotes, Claude Code alone at a run-rate "over $2.5 billion", and a customer count over $1 million a year that went from a dozen to more than 500 in two years. The May 28 announcement of its Series H: $65 billion raised at a $965 billion post-money valuation, with run-rate revenue that had "crossed $47 billion earlier this month". Read the two pages as a series and the official record is already remarkable: the company's own numbers mark revenue at $14 billion in February and past $47 billion in May — 3.4 times in about three months — and mark the company at $380 billion, then $965 billion, fifteen weeks apart. The May round included $15 billion of previously committed hyperscaler investment, of which Amazon's $5 billion is the only allocated piece — Microsoft's and Google's shares are deliberately not broken out. The same page lists the compute partnerships that frame everything below: Amazon for up to five gigawatts, Google and Broadcom for five gigawatts of next-generation TPU capacity, SpaceX for access to its Colossus clusters — and, in a detail almost nobody quotes, Micron, Samsung and SK hynix as "infrastructure partners". The June 1 notice says a draft S-1 was confidentially submitted, shares and price not set. That is the entire official record.
Two numbers wearing the same jersey, so watch the swap: the $65 billion raised in May is company-official; the $65 billion revenue run-rate reported for end-July is a leak — Bloomberg, CNBC and Reuters have each reported it, citing their own sources, from what Bloomberg describes as a regular update Anthropic sent its own investors, and the company has not commented. The leaked series interleaves cleanly with the official one — about $9 billion at the end of 2025, past $30 billion in April, the official $47 billion in May, $65 billion in late July — and it carries the only year-over-year figure on any record: a June quarter above $11.5 billion against $787 million a year earlier, unaudited. Same numeral, different quantities, different reliability — and a run-rate is an annualised pace, not audited revenue.
Arc one: the revenue is already in our SpaceX model
Our SpaceX model prices the frontier model-layer market as a pool: $18.85 billion a quarter — Anthropic $11.6 billion, OpenAI $6.7 billion, Grok $0.55 billion — with the model's own caveat attached: press-reported, none of it audited. That pool is the denominator of the share-shift question we published in August: Anthropic and OpenAI's combined $18.3 billion June quarter is 2.3 times all of SpaceX's revenue, and Anthropic's $11.6 billion — up 143% sequentially — is 61% of the pool.
The stored figure holds up well against the official record: the company's own May pace, $47 billion annualised, is $11.75 billion a quarter — within about 1% of the June quarter we carry. The leaked July pace does not: $65 billion annualised is $16.25 billion a quarter, and our stored quarter sits roughly 29% below it. If the leak is right, Anthropic added $18 billion of annualised pace in two months — a 38% jump. We are holding the pool at the June-quarter actual anyway, for a reason worth stating: a quarterly actual and an instantaneous pace are different bases, and splicing a leaked pace into a pool of quarterly actuals would trade consistency for recency. The S-1's audited half-year is the one event that upgrades the basis rather than churning it; the pool gets refreshed then, on numbers with an auditor's name attached.
Arc two: the compute book the IPO funds
In August we decomposed every Anthropic compute contract with priceable terms: 1,185 megawatts across four deals, roughly $83.1 billion of contracted value — $133 billion counting a $50 billion Fluidstack commitment that has no priceable terms — splitting cleanly into two products: powered shells at ~$2.37 million per megawatt-year on twenty-year leases, and Vera Rubin compute at ~$15.5 million on six-year terms. Only TeraWulf names Anthropic in its own disclosure; the rest is press-attributed. Annualised, the priced book is about $10.6 billion a year of committed spend.
Against that stands the financing everyone is aggregating this week: a $15 billion revolving credit facility that Bloomberg reports as "set to finalize" — not signed, on any public record. Put the two published numbers together and the proportion is the story: $15 billion covers 18% of the priced book, 11% counting Fluidstack, about 1.4 years of the annualised committed spend. The bars above the article show it. A revolver at that scale is working capital, not the war chest; the raise being discussed for the IPO — as much as SpaceX's $85.7 billion or more, per Bloomberg's people familiar — is the instrument sized like the book.
And the priced book is a floor, not the total. The Series H partnerships — Amazon's up-to-five gigawatts, the five-gigawatt TPU program, Colossus — sit entirely outside it, unpriced. So does the largest reported commitment of all: our August read of the sector's financing structures carries Anthropic's pledge of up to one million Google TPUs and press reports of a roughly $200 billion, five-year Anthropic–Google programme — a reported figure more than twice the entire priced book. One of the partnerships now has a public conversion rate: on its September earnings call, Broadcom disclosed for the first time that it earns $20–30 billion of content per gigawatt of customer compute, and named a customer roadmap that includes Anthropic at one gigawatt in 2026, five in 2027 and an incremental ten in 2028 — the disclosure our Broadcom analysis was built on. Multiply management's own two numbers and Anthropic's five-gigawatt 2027 program alone implies $100–150 billion of Broadcom content — arithmetic on the company's disclosures, not guidance, and it dwarfs the entire priced book. It also cuts the other way: Broadcom counts six accelerator customers but put named gigawatts on only three — Anthropic, OpenAI and Meta, with Anthropic on track to become its largest by 2027 per the same call — so the Broadcom model's AI line is, in meaningful part, a bet on this IPO's proceeds arriving. Hock Tan gave the dependency its image on that call, likening Anthropic and OpenAI to "two geniuses stranded in the middle of Mongolia who need help getting to college" — his explanation, as reported from the call, for why only two of his six accelerator customers need outside financing to keep pace.
Arc three: the marks are already in public earnings
Our Microsoft analysis put the mechanism in one sentence: "Microsoft doesn't revalue this stake every quarter; it updates the number only when a fresh financing round gives it a new price to point to, so this isn't a repeatable quarterly tailwind." An IPO is that trigger — for every holder, simultaneously, at whatever price the market actually clears.
The May round already showed what one repricing does. Amazon's June quarter reported EPS of $5.75, up 242% — and the release itself says net income included $53.4 billion of non-operating pre-tax other income, "primarily from our investments in Anthropic". We stripped it at the time: the repeatable number was $27.5 billion of operating income, up 43%. Microsoft's quarter: a $3.2 billion gain on its Anthropic stake added $0.33 of a roughly $0.50 beat — two-thirds of the beat was this one private mark-up.
The third name everyone adds to that list needs more care, and the precision matters. Alphabet's headline EPS this season was indeed two-thirds mark-to-market — roughly $6.26 of the $9.11 reported — but its confirmed $94 billion position is tied to the SpaceX IPO, not to Anthropic. The two private companies distorting hyperscaler earnings this year are SpaceX and Anthropic, and Alphabet's published distortion is the other one. Its Anthropic exposure is nonetheless the largest on any record: court filings in the Google antitrust case, as press-reported, put its stake at about 14% — ownership capped at 15%, no voting rights, no board seats — and Bloomberg reported in July that the stake is carried at around $124 billion.
We read the filing behind that number. Alphabet's June-quarter 10-Q discloses $124.3 billion of non-marketable equity securities under the measurement alternative — a bucket that never names Anthropic, in the same note where Alphabet does name its SpaceX shares, a position roughly nine times smaller. The chain, precisely: the $124.3 billion is Alphabet's disclosed total; the roughly $124 billion carrying value for the Anthropic stake is Bloomberg's reporting; and reading the stake as substantially the whole bucket is our inference, labelled as such. The note even carries Microsoft's repricing rule in filing language — the value is "adjusted upward or downward to fair value upon observable transactions for identical or similar investments of the same issuer" — which is exactly what an IPO print is. The money going in now has a date, too: Google's 2026 commitment was a separate April instrument, up to $40 billion — $10 billion immediately, the rest on milestones, per CNBC — which is why the Series H hyperscaler line names no Google amount. Set beside the roughly $200 billion of spend reportedly flowing the other way, the pair is the most circular relationship in this story — and the S-1's related-party section is where the two directions finally net out on paper. And the 14% predates this year's two rounds, which raised $95 billion of new capital; if Google did not participate pro rata, today's percentage is lower. The prospectus's shareholder table settles it.
What would a $2 trillion print do to these lines? Here is where most analysis quietly invents stake percentages — only Alphabet's is on any record, and we are not going to invent the others. What can be said is now arithmetic, because the previous step sits on the company's own pages: the May round repriced Anthropic from February's $380 billion to $965 billion, a step of $585 billion, and that step generated every gain above. A $2 trillion listing would be a step of $1.04 trillion — 1.8 times larger in dollars than the one that just put $53.4 billion into Amazon's quarter. Amazon's booked gain was seventeen times Microsoft's on the same event, the only public signal of their relative exposure, and the step lands in whichever quarter the listing prices. The symmetric warning comes from the precedent: marks reprice down, too. SpaceX listed at $1.77 trillion and fell roughly 50% from its post-IPO peak — below its IPO price — within seven weeks. Holders who mark at a rich print ride it in both directions.
What $2 trillion is, in the only language we can check
The table above the article is the comp set, from our own published models and pieces. Alphabet trades near 9.3× trailing revenue, Tesla at 11.75× trailing sales as we priced it in August, SpaceX around 14× implied, Broadcom near 20× and Nvidia near 21× trailing enterprise value to revenue. Against those: Anthropic's own May mark on its own May pace — $965 billion on $47 billion — is 20.5×, Broadcom-Nvidia altitude. The $2 trillion talk on the leaked July pace is 30.8×, above every name in the set.
Three honesty clauses before anyone quotes that. First, the bases differ: our comp multiples divide market prices by audited trailing revenue; the Anthropic rows divide a rumored valuation by an annualised leaked pace — and on a genuinely trailing basis, summing the quarters actually reported, $2 trillion would run north of 60×, roughly double the headline figure. Second, the bull defence uses forward numbers: on the year-end pace of $100–120 billion that the FT reports investors using as the startup's preferred measure, $2 trillion is 17–20×; on the internal 2028 forecast of $190–200 billion that Reuters attributes to unnamed sources, it is about 10×. Those denominators are one and two-plus years out, unaudited, and leak-sourced — that is the pitch being made, not a multiple that exists. Third, none of the Anthropic rows involves a market price. The one precedent for a private AI-adjacent giant meeting an actual tape at this scale is the SpaceX line above.
What an IPO buyer would actually own
The governance is unlike anything in the comp set. By Anthropic's own description, its Long-Term Benefit Trust holds a dedicated share class — Class T — whose function is electing directors, with authority that phases in on time and funding milestones toward electing a majority of the board, and trustees who elect their own successors. By press accounts the board crossed to a trust-elected majority in April. The trust today has three members — Neil Buddy Shah (chair), Richard Fontaine, and Ben Bernanke, who joined in July — after Mariano-Florentino Cuéllar stepped down in August to become Anthropic's Chief Global Affairs Officer. Per the FT, four of the seven current directors are trust-appointed, including Reed Hastings and Vas Narasimhan; the trust's Class T carries no economic stake, per the same reporting; and the appointment right survives the listing. Add the super-voting founder shares the same coverage reports, and the shape is clear: a public buyer would get economic exposure to the revenue above with, by design, little claim on control. One counterweight is written into the trust's own design: "failsafe" provisions let sufficiently large stockholder supermajorities change the trust and its powers without the trustees' consent, at thresholds that rise over time. How the S-1's risk-factor section words all of this is one of its most-awaited pages.
The offer's reported structure differs from the recent mega-listings too. Per The Information's reporting, the deal would include a secondary component — existing holders selling at the listing — where SpaceX and Cerebras both listed with no selling stockholders; lockups are reportedly under discussion at longer than the standard 90–180 days; and employee sales would run through pre-scheduled trading plans. Which holders sell, and how much, is exactly the kind of fact only the prospectus can settle.
What we will not do, and what we will
We will not publish an Anthropic fair value. A forward model needs audited revenue, a margin structure and a share count; Anthropic has published none of the three. Building one from leaked run-rates would manufacture precision we do not have — so the number is absent on purpose, and this piece prices nothing.
What we will do, committed here: read the S-1 the day it prints, the way we read CXMT's. The audited revenue against the $47 billion and $65 billion paces above. The gross margin against the GPU complex — the number that, as PitchBook argued in June, reprices the sector either way; its analysis pre-registered the benchmarks the S-1 lands against — compute cost of $0.71 per revenue dollar in the March quarter, projected $0.56 in June, a path toward roughly 44% — and, on its scenario math, margins printing below 35% would compress fair value by 70 to 81 percent. The share count against the $965 billion mark. The cover page against both banks lists. The selling-stockholder table against the structure reports — and against Alphabet's court-filed 14%. The risk-factor pages against the one advance claim on record: CNBC's sources say the filing will name the backlash against AI and data centers as a risk factor. The queue matters as well: OpenAI filed its own confidential S-1 on June 8 — a week after Anthropic, at a reported $852 billion, with Goldman Sachs and Morgan Stanley advising, the two banks atop Anthropic's reported slate — and by late-June reports was leaning toward 2027, so Anthropic's audited margin prints first and, per PitchBook, recalibrates OpenAI's eventual offering before it prices. And when it trades, it joins coverage the way Klarna did — on reported numbers.
What to watch
- The gross margin line in the S-1. The first audited margin of a frontier lab at scale — the single number with the widest repricing radius, for Anthropic and for everyone selling it compute. PitchBook's June path toward roughly 44% is the pre-registered test it lands against.
- Audited revenue against the ledger. The S-1's income statement versus the company's $47 billion May pace and the leaked $65 billion July pace — and versus the $11.6 billion June quarter our SpaceX model carries, which gets refreshed on the audited basis.
- The cover and the selling-stockholder table. Banks and their order, the exchange, the ticker, and whether existing holders sell — the reported break from the SpaceX and Cerebras pattern.
- The trust's risk-factor language. How Anthropic itself describes three trustees electing a board majority over public shareholders' economics.
- Mark season in the listing quarter. Amazon's and Microsoft's Anthropic lines reprice on the print — and so does Alphabet's stake, about 14% per court filings and carried around $124 billion by Bloomberg's reading of its filings, the largest single exposure on any record.
Anthropic's raises, valuations, run-rate figures, trust design, hyperscaler and compute-partnership details are from its own February 12, May 28 and June 1 announcements and its published trust description — its only official statements. The holder gains and EPS figures are as each company reported, covered in our June-quarter analyses; the SpaceX pool figures, the 1,185 MW / $83.1 billion compute decomposition and every model number are our published assumptions, press-reported at source and none audited. Broadcom's per-gigawatt content and customer roadmap are management's statements on its September 2 earnings call. The $65 billion July run-rate and its waypoints (Bloomberg, CNBC, Reuters), $2 trillion talk (FT, NYT, Reuters attributions), $15 billion revolver status (not signed on record), offer structure (The Information), the FT's board details, the court-filed Alphabet stake and Google's commitments in both directions, OpenAI's confidential filing, Hock Tan's call remarks and PitchBook's margin benchmarks are press reporting rather than company statements. The $124.3 billion bucket and SpaceX-shares note are from Alphabet's June-quarter 10-Q; the Anthropic attribution is Bloomberg's, the whole-bucket reading our labelled inference. All multiples, ratios and step arithmetic are ours, with as-of dates as published in the linked pieces. No live price appears here.
