Skip to the short version
ANTS-1, Annotated

Updated November 3, 2026 for Amendment No. 1. What changed →

The information in this preliminary prospectus is not complete and may be changed. These securities may not be sold until the registration statement filed with the Securities and Exchange Commission is effective.

The trillion-dollar prospectus, annotated.

Anthropic's IPO filing runs to more than 400 pages. Its first amendment, filed November 2, fills in two big blanks: third-quarter revenue of $17.0 billion and $895.4 billion committed to computing power. Quarterly revenue is up 45-fold in six quarters; a first operating profit in Q2 turned back into a loss in Q3; and a $104 billion nine-month loss is almost all paper. The share price is still a blank. Here is what the filing says, how to read it, and what its sharpest critics and defenders are arguing about.

Amendment No. 1 S-1/A filed November 2, 2026

What the first amendment changed

A week after the public S-1, Anthropic filed an amended version with third-quarter results and the compute total it had left blank. The share price is still missing. Every page reference on this site now points to the amended filing.

    The five-minute version

    Seven things to know before you read anything else

      Growth S-1/A pp. 136, 150, 160–161, 187, 199

      From $376 million a quarter to $17.0 billion

      Anthropic's revenue grew 45-fold in six quarters. Almost all of the jump came in 2026, and growth slowed in the third quarter.

      Before you look: how much revenue did Anthropic book in the three months to June 30, 2026?

      Quarterly revenue, $ billions

      Q3 2026 was blank in the original S-1. The dashed line marks our estimate from that filing's statement that Q3 revenue was 86 times Q4 2024's $197 million. The amendment reports $16.96 billion.

      Source: S-1/A p. 160 (quarterly statements of operations)

      The amended S-1 shows revenue of $4.6 billion for all of 2025 and $33.2 billion for the first nine months of 2026. Third-quarter revenue alone was $16.96 billion, up 1,127% on a year earlier and 47% on the second quarter. The company credits successive model and product launches, especially Claude Code and the Opus releases, which pulled spending up across existing customers.

      How we got Q3 right. The original S-1 left Q3 revenue blank but said it was 86 times Q4 2024's $197 million, which works out to about $16.9 billion. The amendment reports $16.96 billion, and the "86 times" sentence is still there.

      Growth slowed from 144% in Q2 to 47% in Q3. Price reductions on new model versions and contractual discounts took about $600 million off Q3 revenue; without them growth would have been about 52%. It has slowed before: Q4 2025 grew 48%. After Opus 4.5 launched in November 2025 at a lower price, growth jumped to 132% in Q1 (the company cites several factors) and 144% in Q2. The filing also warns that usage dips around summer and winter holidays and that releases are uneven, so quarters will swing. Still extraordinary, but slower, and the bears have noticed (see the debates).

      Who is paying

      Customers paying at least $100,000 a year (the S-1's stand-in for "enterprise") went from about 20 at the end of 2023 to more than 11,900 at the end of September 2026. Together they produced $24.5 billion, 74% of nine-month revenue. More than 1,900 pay over $1 million and about 250 pay over $10 million. About 79% of nine-month revenue was usage-based (paying per token through the API) and 21% came from subscriptions such as Claude Pro, Max, Team and Enterprise.

      Customers paying $100,000 or more over the trailing twelve months

      Approximate counts as stated in the filing ("~", "more than").

      Sources: S-1/A pp. 136, 144 and the MD&A chart "Rapid growth of enterprise customer base"; the June 2026 point is from the original S-1 (orig. p. 148)

      Three caveats the bulls skip

      • Net revenue retention of "approximately 900%" is not a SaaS number. A customer counts if it paid at least $1 in the base quarter. Measured against a tiny base a year ago, the ratio explodes. Typical software companies report 110–130%p.2.
      • Some revenue came from customers who broke the rules. An estimated 5–7% of revenue in the first nine months of 2026 came from customers later found to be violating the usage policy or commercial terms. Enforcement stepped up in August and September and is expected to reduce that revenue.
      • More than 40% of the money arrives through rivals' clouds. 44% of nine-month revenue came through Amazon, Google and Microsoft marketplaces (48% in the first half), up from 11% in 2023. Those same companies hold 52% of Anthropic's receivables.

      The growth is real, though the 2026 quarters are unaudited. But a quarter of it rests on five customers, more than 40% of it is collected by competitors, and part of it came from customers the company is now cutting off.

      Unit economics S-1/A pp. 135, 151, 160–163

      Each dollar of compute brought in $2.08 in Q2, and $1.72 in Q3

      In early 2025 Anthropic spent $2.51 on computing for every dollar it earned. By mid-2026 that had flipped, and the second quarter was its first with an operating profit. In the third, demand outran the capacity it had contracted, it bought pricier short-term compute, and the ratio fell back.

      The income statement doesn't split costs into "cost of revenue" and operating expenses. It reports one big line, compute and infrastructure, which covers research, model development, serving customers and Anthropic's own use of Claude. A note to the accounts does regroup costs by function. Cost of revenue, mostly the compute that serves customers, was $4.8 billion in the first nine months of 2026, an implied gross margin of about 85%. Most compute, though, is booked as research and development ($19.2 billion), and that is what produces the models customers pay for. So the chart below sets the whole compute line against revenue, quarter by quarter.

      Revenue and compute spending per quarter, and revenue earned per dollar of compute

      Two panels on separate scales. The top panel compares dollars. The bottom panel shows the ratio between them.

      Source: S-1/A p. 160; ratio computed from the same table
      Operating margin by quarter

      Switch between the reported figure, the same figure without the non-cash donation-matching charges ($2.0B in Q3), and the company's adjusted version, which also adds back stock-based pay and a $1.5B copyright settlement.

      Source: S-1/A pp. 160, 163

      Why the ratio jumped, and why it fell back

      Compute costs here behave like rent. Anthropic signs up for capacity 12 to 18 months ahead, and the expense starts when the capacity is delivered, whatever demand turns out to be. In the first half of 2026, demand outran the capacity that showed up, so revenue per compute dollar soared.

      In the third quarter the company covered the gap by renting. When demand or research needs exceed its long-term capacity, it buys short-term compute that "currently costs significantly more". Q3 included $5.0 billion of it, against about $1.5 billion in the whole first half (the nine-month total was $6.5 billion). Compute costs rose $4.3 billion as revenue rose $5.4 billion, and revenue per compute dollar fell from $2.08 to $1.72. The company expects to use more short-term compute in the near term, which it says will compress margins.

      The bill is about to step up. One contract alone, with SpaceX's AI unit, runs at about $3.1 billion a month once fully delivered, which the company expects in December 2026. That is about a third of the entire third-quarter compute line, every month. The next few quarters will show whether revenue keeps pace with the capacity arriving.

      Why "adjusted" flatters. It leaves out stock-based pay ($1.07B in Q3) and the donation-matching charge ($1.96B in Q3). Had the IPO happened on September 30, restricted stock units would have added an $8.2B catch-up charge at once and $24.1B more over about two years.

      Because we typically procure compute 12 to 18 months in advance of availability, these expenses are generally fixed in the near term and increase as committed compute becomes available to us. Compute and infrastructure expenses as a percentage of revenue decreased from 251% for the three months ended March 31, 2025 to 48% for the three months ended June 30, 2026, as revenue growth outpaced the growth of our committed compute cost base, and increased to 58% for the three months ended September 30, 2026, reflecting the delivery of additional compute.

      Q2 showed revenue can cover the compute bill twice over. Q3 showed how fast that ratio moves when demand runs ahead of contracted capacity. It will be tested again as hundreds of billions of dollars of contracted capacity arrives.

      The $104 billion loss S-1/A pp. 160–164, F-7

      Why the company's best nine months show its biggest loss

      Anthropic lost $1.3 billion running its business in the first nine months of 2026. It lost $104.0 billion on paper. The difference comes from a pile of IOUs that grew more valuable as the company did.

      From revenue to net loss, $ billions

      Bars step down for costs and up for income. The red bar is the accounting charge.

      Source: S-1/A p. F-7 (consolidated statements of operations)

      Between 2023 and 2025, Amazon and Google lent Anthropic billions through convertible notes. These are loans that turn into shares at a price fixed in advance. As Anthropic's valuation climbed, the shares those notes would become were worth more. Accounting rules require the company to record that increase as an expense, called a change in fair value. That line was $94.3 billion in the first nine months of 2026. Most of it ($59.6 billion) landed in the second quarter, when the Series H round reset the valuation; the third quarter added $21.9 billion.

      A further $10.4 billion of "other expense" this year came from valuing new financing facilities from Google, Amazon and Broadcom, $8.4 billion of it in the second quarter. No cash left the building for either item. Interest earned on the cash pile added $2.4 billion going the other way.

      The perverse part. The better the company does before the IPO, the bigger this loss looks. It stops when the notes convert into stock at the IPO. Amazon's notes had $6.2B of principal outstanding at September 30 but are carried at $130.5B. They convert into non-voting shares, capped at 12% ownership unless Amazon waives the cap.

      The other side of the trade. Amazon, which holds the notes, records their rise in value in "other comprehensive income", not profit: $140.3B of unrealized gain by September 30. Its profit included $62.8B of 2026 gains, mainly from marking up its Anthropic preferred shares, nearly all in the first half (Amazon 10-Q).

      The expense was largest in the three months ended June 30, 2026, reflecting the increase in our enterprise valuation in connection with our Series H financing. We will continue to recognize changes in fair value until these instruments are converted into equity, repaid, or otherwise settled.

      What the loss does tell you

      The paper loss tells you little about the business. It does tell you who now owns the company. Each markup is value that moved from existing shareholders to the noteholders. The real cash question sits in the compute contracts, the next section.

      Read "net loss" here as "how much richer Amazon and Google's IOUs got." The operating line is the one that shows how the business is doing: positive in Q2, negative in Q3, and about +$1.4 billion in Q3 without the donation charge.

      The compute bill S-1/A pp. 6, 31, 173–176

      We estimated $855 billion. The filing now says $895.4 billion.

      Anthropic has committed about $895.4 billion to chips, data centers and cloud capacity, mostly as minimum payments. $69.8 billion is due in the next 12 months: about a year of revenue at Q3's pace, or 1.8 times Q3's compute spending at an annual rate. It estimates another $150–175 billion on deals without a fixed price, which the total leaves out.

      As of and subsequent to September 30, 2026, we have entered into purchase commitments and lease agreements, including short-term commitments, as well as leases not yet commenced, with an aggregate committed spend of approximately $895.4 billion, substantially all of which is related to compute and infrastructure. These commitments include approximately $447.5 billion expected to be incurred before December 31, 2030, and substantially all of the remainder is expected to be incurred between 2031 and 2045. Approximately $69.8 billion of these commitments are due in the 12 months following September 30, 2026.

      How the commitments stack up, $ billions

      The filing's own components of the $895.4 billion. Hover or tap a block for its source.

      Sources: S-1/A pp. 173–174 (contractual obligations), 139 (cash), 160 (revenue)

      The filing's own breakdown, at September 30: $313.8 billion of purchase commitments and $6.2 billion of leases already running, in its contractual table; $17.4 billion of short-term minimums; and $451.2 billion of leases signed but not yet started. Signed after September 30: $100.2 billion of long-term deals and $6.6 billion of short-term ones. Total: $895.4 billion. Not included: an estimated $150–175 billion for capacity without a fixed dollar commitment, mainly additional TPUs and AMD systems, whose price and volume aren't final.

      How binding is it? $213.8 billion of the total can be cancelled or depends on conditions before it starts, "such as the counterparty's ability to secure financing, facilities, or regulatory approvals". Only $4.3 billion of the $69.8 billion due in the next 12 months is of that kind. New in the amendment: some suppliers themselves depend on outside financing, and if it dries up, capacity "may be delayed, repriced, reduced, or not delivered".

      How close we got. The original S-1 left the total blank. Adding up its June 30 footnotes and the deals signed afterwards, we got about $855 billion. The filed total is about $40 billion, or 5%, higher. The two aren't directly comparable: ours started from June 30 balances, the filing from September 30. The filing doesn't bridge the two. Pushing the total up: deals signed or enlarged since (Google's minimum rose by $6.8 billion) and $17.4 billion of short-term minimums the original S-1 left blank. Pulling it down: payments made in the third quarter. The rest is timing and rounding.

      Who it's owed to is spelled out deal by dealp.170. Some are firm minimums Anthropic must pay "regardless of actual usage", with cash true-ups for any shortfall. Others are "up to" amounts or can be cancelled.

      The named compute deals, $ billions

      Shaded by how binding each is. Hover or tap for terms.

      Source: S-1/A pp. 174–176 (MD&A, "Infrastructure Service Arrangements" and "Infrastructure Equipment Leases")
      The bill grew $91 billion in the last 25 days of drafting

      Long-term compute deals signed after June 30, as stated in the same footnote of each draft and the original S-1. The September 4 draft used a June 1 starting point, so it is left out. The amendment counts deals signed after September 30 instead ($100.2 billion), so it isn't comparable and isn't shown.

      Sources: DRS/A filed Oct 1 and Oct 9, 2026; original S-1 p. F-65

      Can they pay?

      The amendment gives three anchors: $69.8 billion due in the 12 months after September 30, $447.5 billion by the end of 2030, and the rest by 2045, plus a year-by-year table for $320 billion of it. We spread the commitments to match those anchors: the table as filed, short-term minimums within about a year, and the leases and new deals ramping up as they start in 2026–2028 and running about ten years. Then set your own assumptions for revenue growth and for how much revenue each compute dollar earns ($1.72 in Q3 2026). The tool shows the year the contracted bill outruns what revenue can carry, and how much of the cash pile the gap would eat.

      Contracted compute vs. what revenue can carry, $ billions per year

      Illustrative model. Bars: the filed commitments, spread by us to match the filing's totals. Line: compute spending that revenue can carry at your chosen efficiency.

      2026 base: $54B (nine-month actuals of $33.2B plus ~$21B assumed for Q4, 25% above Q3). Growth then halves each year, to a 5% floor.

      Q1 2025: 0.40×. Q2 2026: 2.08×. Q3 2026: 1.72×. Higher means more efficient.

      $150–175B for capacity without a fixed price, mainly TPUs and AMD systems, paid after delivery. We add the midpoint, spread over 2027–2035.

      Model: our spread of the S-1/A totals (pp. 6, 173–175). Not a forecast.

      Anthropic has bet that demand keeps compounding. At Q3's efficiency, paying for 2027's commitments (about $75 billion in our spread) out of revenue alone would take about $130 billion of revenue, 2.4 times our 2026 estimate. With the $131.8 billion of cash and short-term investments absorbing the gap, growth of about 70% in 2027 is enough; the company's TPU and AMD estimates raise that to about 90%. The pressure then moves to 2028–2030, when our spread puts $110–130 billion a year. Much slower growth means new money or cancelled contracts.

      The money goes round S-1/A pp. 57, 74, 166–176, F-16, F-39–F-42

      Its biggest investors are also its biggest suppliers

      Amazon and Google have put money into Anthropic, and Broadcom and AMD have agreed to lend or invest. Each sells it far more in chips and cloud capacity, and several of those investments only unlock as the partner delivers its own hardware.

      Critics call this circular financing: a supplier funds its customer so the customer can buy from the supplier. Defenders call it ordinary vendor financing that lets a fast-growing buyer lock in scarce capacity. The S-1 gives enough detail to judge each relationship on its own terms.

      Money in vs. money owed back, $ billions

      Left: cash invested or lent, plus committed facilities still to come. Right: minimum compute spending Anthropic has committed to that partner. Select a partner for the details.

      Sources: S-1/A pp. 166–176, F-39–F-42; investment totals are our sums of disclosed tranches

      Four details worth knowing

      • Amazon's next $15 billion arrives in $5 billion steps "following the achievement of the relevant compute capacity milestones", meaning capacity Amazon delivers. Amazon can also settle its notes by netting them against money it owes Anthropic from marketplace salesp.165.
      • Google's $20 billion facility is sized to TPU capacity delivered and to revenue targets. Google also holds an option to buy up to $10 billion of stock alongside the IPO at the IPO price.
      • Broadcom leases Anthropic about $161 billion of TPU hardware and offers a $42 billion facility whose only purpose is to lend Anthropic the rent. It accrues 9% interest (13% if revenue targets are missed) and converts into stock at a 10% discount to the market pricep.166.
      • AMD sells Anthropic more than $20 billion of GPU systems and has agreed to buy up to $5 billion of its stock as delivery milestones are metp.163.

      Hiding in plain sight. Amazon and Google each own more than 5% of Anthropic. They hold non-voting Class N shares, and the related-party rules key off voting stock, so neither appears in the "Related Person Transactions" section. Their deals are in the MD&A insteadp.282.

      The circularity is real and disclosed. It ties Anthropic tightly to a few partners who are also its competitors, but most of the commitment runs one way: Anthropic owes its partners several times what they have put in.

      Who controls Anthropic S-1/A pp. 108–119, 246–257, 292–310

      One share with 50.1% of the vote, and a trust that picks the board

      Public investors will own a slice of the profits but have almost no say. The filing calls the structure "novel" with "no direct public company precedent."

      There are five classes of common stockp.292. Class A is what you can buy: one vote per share. Class N is the same economically but carries no vote, and it's held mostly by strategic partners such as Amazon and Google. Class F is a single share held by an LLC of the seven co-founders. It always carries 50.1% of the stockholder vote and has no economic rights. Class T is a single share held by the Long-Term Benefit Trust, which elects four of the seven directors. Class E gives employees a vote on a few governance matters.

      Who owns the profits, who casts the votes, who picks the board

      Economic shares are approximate: the S-1 leaves ownership percentages blank, so they are derived from disclosed share counts.

      Sources: S-1/A pp. 12, 117, 251, 260, 293; F-31, F-39–F-42 (derived)

      What happens if…

      The rules only make sense when you test them. Pick a scenario.

      The trust

      The trust has four trustees, with a fifth seat vacant: Neil "Buddy" Shah (chair, CEO of the Clinton Health Access Initiative), Richard Fontaine (CEO of the Center for a New American Security), former Federal Reserve chair Ben Bernanke, and Harvard physicist Christopher Stubbsp.245. Each is paid $400,000 a year and may not own Anthropic stock. Anthropic funds the trust and consolidates it on its own balance sheetp.246. The trust's directors are Reed Hastings, Novartis CEO Vas Narasimhan and former Microsoft CFO Christopher Liddell, plus one open seat. The founders' three seats are Dario Amodei, Daniela Amodei and Spark Capital's Yasmin Razavip.256.

      Small print, big deal. Anthropic must give the trust five days' notice before it can "transfer control of an AI model exceeding capabilities thresholds" set by the board each quarter. A transfer without that notice is voidp.285.

      You are buying an economic stake in a company run by its founders and overseen by a trust, with a legal duty to weigh "the long-term benefit of humanity" alongside your returns. That is the deal on offer, and it is disclosed plainly.

      The founders' letter S-1/A inside cover, 10 unnumbered pages

      "Never let your company be worth a dollar more than it is actually worth"

      Before the table of contents, all seven co-founders signed an eight-page letter to investors. It is printed as images, so most text searches of the filing miss it. The amendment leaves it unchanged apart from small typographic fixes and adds two pages after it: third-quarter highlights and an adoption timeline. These are the passages that matter, with our notes.

      Fill in the blank S-1/A cover, pp. 12, 286–288, 330

      The cover says "between $ ____ and $ ____ per share." What would you have to believe?

      Bankers will set a price range in a later amendment. Until then, you can run the test a valuation professor would: start from a market value and work out the future the price assumes.

      Reverse valuation: the revenue Anthropic would need

      Move the sliders. The chart compares the revenue required in the target year with what today's largest companies earn.

      Implied price per share shown below, on ~16.9B fully diluted shares (our estimate).

      When growth slows and the company is priced like a big, steady tech firm.

      Microsoft and Alphabet run around 30–36%. Compute-heavy businesses run lower.

      Large-cap tech has typically traded around 25–35× earnings.

      Higher for riskier bets. Long-run stock market returns are around 8–10%.

      New shares for staff, partners' stock-settled facilities and the public benefit fund.

      Method: required revenue = value × (1 + return)years × (1 + dilution)years ÷ (P/E × margin). Comparators: trailing-12-month revenue, StockAnalysis, Oct 26, 2026.

      Risk factors S-1/A pp. 28–120

      88 risk factors, and several no prospectus has carried before

      Risk factors are the part of an S-1 where lawyers list everything that could go wrong. Most are boilerplate. Anthropic's include disclosures about its own models behaving badly, written in securities-law language.

      New in the amendment: AI speeding up AI research

      Anthropic's Responsible Scaling Policy sets capability thresholds that trigger extra safeguards. The amendment says the company may soon meet the one for AI research and development, and that it is weighing rules to keep people in charge of building its next models. These passages are new; the original S-1 did not contain them.

      The debates

      Nine arguments about this IPO, heard from both sides

      For each fight: what critics say, what defenders say, what the filing itself shows, and where we come down.

      Who's saying what

      The expert map

      Finance and AI analysts on the filing and the business behind it, including reactions to the November 2 amendment. Filter by stance or field. Each card links to the original.

      How we got here, and what happens next

      Funding rounds, six secret drafts, and a staircase of unlocks

      Private valuations by round, $ billions (log scale)

      On a log scale every tenfold rise takes the same space. On a normal scale everything before 2026 would be flat.

      Sources: S-1/A pp. 286–288 (round prices); press reports for post-money valuations
        Six confidential drafts, the public S-1 and its first amendment: what changed

        Highlighted cells mark a change from the previous version. Blank means no change noted. From our comparison of the six confidential drafts the SEC released on October 26, and of the November 2 amendment.

        Sources: DRS (Jun 1), DRS/A No. 1–5, Form S-1, Form S-1/A, SEC EDGAR
        When locked-up shares can be sold: share of each group's holdings released

        Assumes listing on Nov 19, 2026 and earnings releases in late February, May, August and November 2027. The filing ties most steps to earnings dates, not fixed days.

        Source: S-1/A pp. 311–312, 322–328; dates are our assumption

        Plain-English glossary

        The jargon, decoded

        Dotted-underlined terms anywhere on this page show these definitions when you hover or tap them.

        Method and sources

        How this was made

        The S-1/A (Amendment No. 1, accession 0001628280-26-068676) and the original S-1 (accession 0001628280-26-067168) were downloaded from SEC EDGAR, along with the six confidential drafts released on October 26. They were converted to text with printed page numbers kept, and every figure here was checked against that text. Page chips (p.160) open the amendment on SEC.gov; chips marked "orig." open the original S-1. In most browsers they jump to the quoted words.

        Derived numbers are labelled as ours and show their arithmetic. Outside claims name their source and date. Where sources disagree, we say so. The compute model and the valuation tool are illustrations built on stated assumptions, not forecasts.

        This is a preliminary prospectus. The price range, share counts and several figures are still blank and will change in later amendments. Figures for 2026 quarters are unaudited. Nothing here is investment advice, and this site is not affiliated with Anthropic or its underwriters.