Anthropic (ANT) — What the S-1 Says a Dollar of Claude Is Worth

Anthropic filed to go public on October 26, 2026. The filing shows a company whose revenue rose from $4.6 billion in 2025 to $16.3 billion in the first half of 2026, whose second quarter was the first with an operating profit, and which costs about 15 cents to serve for each dollar it sells. It also shows about $657 billion of named compute commitments landing fastest in 2027–29, a founders' share that carries 50.1% of every general vote, and a trust with no stock that elects a majority of the board. Fair value $50.72 a share, probability-weighted, $833 billion: 14% below the $58.90 price of the last private round and less than half of the $2 trillion the wires discuss.

TON618 Capital Research · As of 26 October 2026 · Download PDF · Download model

On the morning of Monday, October 26, Anthropic, PBC filed its registration statement with the SEC for a listing on the Nasdaq Global Select Market under the ticker ANT, with Goldman Sachs and Morgan Stanley leading. There is no price range, no share count and no figure for the quarter that ended in September; those arrive by amendment. What the filing does give is three audited years and a half-year: revenue of $41 million, $386 million and $4,585 million in 2023–25, then $16,286 million in the six months to June 2026; a $693 million operating loss for that half that turned into a $974 million operating profit in its second quarter; $131.7 billion of cash and short-term investments; and a net loss of $80.7 billion, of which $72.4 billion is the accounting mark on convertible notes that rose in value because the company did. The last private round, Series H in May and June, was priced at $589.0095 per share before a ten-for-one split, or $58.90 after it, about $968 billion for the whole company on the count this note uses. This note asks what the business the filing describes is worth, what the company has promised to spend to build it, and who decides.

Terms used throughout, defined once. S-1: the registration statement a company files to sell shares to the public; DRS: the confidential drafts filed before it. Consumption revenue: usage billed by the token (the filing's "pay-for-intelligence"); subscription revenue: prepaid seats (Pro, Max, Team, Enterprise). Compute: the processors, memory, networking and power used to train and run models; the filing books all of it as one operating expense. Cost of revenue: the filing's functional line for the compute and overhead that serve customers; the serving margin is one minus cost of revenue over revenue. CSP: a cloud services provider (Amazon, Google, Microsoft), which here is investor, compute vendor, sales channel and customer at once. Run-Rate Revenue: the company's annualized measure (13 times 28 days of consumption plus 12 times monthly subscriptions). RPO: remaining performance obligations, contracted revenue not yet recognized. NRR: net revenue retention, this year's revenue from last year's customers over last year's. PBC: a Delaware public benefit corporation, whose directors must balance stockholders against a stated public purpose. LTBT: the Long-Term Benefit Trust, five unpaid-in-stock trustees who hold one Class T share and elect a board majority. Class F: one share held by the founders' LLC with 50.1% of the general vote. Class N: non-voting shares held by strategic investors, the Amazon notes and the Public Benefit Fund. FCFF: free cash flow to the firm, the cash the business generates before financing. Cost of equity: the return shareholders require, built here as the ten-year Treasury yield plus beta times an equity risk premium. Terminal growth: the growth rate assumed forever after the explicit forecast. Reverse DCF: solving the model backward for what a given price requires.

Anthropic was incorporated in Delaware in January 2021 as a public benefit corporation by seven researchers who had led GPT-3's engineering, the scaling-laws work, RLHF and interpretability research at OpenAI. Its charter's stated purpose is "the responsible development and maintenance of advanced AI for the long-term benefit of humanity," and the filing calls the company "first and foremost a research lab." It launched commercially in the first quarter of 2023 with 43% of that year's $41 million of revenue from one customer, shipped the Claude 3 family in 2024, Claude Code as a research preview in February 2025, and Claude Cowork in January 2026. Today the model line has four classes: Mythos (with Fable), Opus, Sonnet and Haiku. Mythos was first released in April 2026 as a gated research preview through Project Glasswing, a program for vetted cybersecurity work; Fable 5, the generally available configuration "with our production safeguards, which block or limit performance on certain sensitive topics," followed in June, with Mythos 5 for Glasswing participants; Mythos 5.1 and Fable 5.1 shipped in September.

The scale as of September 30, 2026: more than 11,900 customers paying over $100,000 a year, 1,900 over $1 million and about 250 over $10 million; roughly 40 million smaller customers; every company in the Fortune 10, over half the Fortune 500 and over 30% of the Global 2000. Revenue per employee is "over $6.6 million," which with the filing's 170% headcount growth puts the staff at about 5,100–5,200. Over 90% of the contributions to the company's own codebase in September 2026 were written by Claude. Regrettable attrition was about 1% over the trailing year, and all seven founders remain.

What it says it is appears most plainly in the eight-page founders' letter inside the filing. It states two categories of decision that "may run counter to investors' expectations": "forward investments (especially in compute) that may seem irresponsibly large at first glance," defended by the scaling laws and a comparison to pharmaceutical research, and "decisions that depart from profit maximization." Its two examples are the ones this note prices: holding Mythos Preview back from general release "meant forgoing substantial revenue and ceding ground to competitors (we also gave away around $100 million in Claude credits to cyber defenders)," and "we have also committed to keeping Claude free of advertising," which "would likely be a large source of revenue for us." It adds that "there may be times when we choose to advance more slowly than we could in order to make our systems safer, even if that puts us at a competitive disadvantage." The letter's own advice on the offering: "never let your company be worth a dollar more than it is actually worth."

Revenue was $376 million, $787 million, $1,382 million and $2,040 million in the four quarters of 2025, then $4,731 million and $11,555 million in the first two of 2026. Consumption, billed by the token, is 80% of the mix; subscriptions 20%; the United States 72%. The third quarter is blank in the public filing, but two sentences pin it: trailing-twelve-month revenue "increased approximately 1,200%" from September 2025 to September 2026, and revenue per employee went from $1.4 million to over $6.6 million while headcount grew about 170%. With the undisclosed fourth quarter of 2024 bounded at $97–174 million, the growth sentence gives $16.0–17.0 billion for the third quarter of 2026 and $34–35 billion of trailing revenue (the revenue-per-employee route gives $15.3–16.3 billion, and the "approximately" in both sentences can widen either); the Financial Times, via Reuters, reported $16.8 billion on October 15, inside the range. Note what the range implies: growth of 144% from the first quarter to the second, then 39–47% from the second to the third. The company's own run-rate metric was more than $47 billion "earlier in May"; Bloomberg reported $65 billion at the end of July, a figure the company has not published.

Who pays. Customers above $100,000 a year produced 74% of first-half revenue ($12.0 billion); the 40 million smaller customers the rest. Four customers have at times exceeded 10% of revenue: Customer A (12% of 2025, 13% of receivables), Customer B (12% of 2025, 18% of the first half of 2025), Customer C (43% of 2023, 11% of the first half of 2026) and Customer D (11% of 2023). All are unnamed, and the filing does not say whether any is an investor or a compute vendor. No customer exceeded 10% in the nine months to September, and 79% of third-quarter revenue came from outside the top five. Remaining performance obligations were $10.8 billion at June 30, about 75% of it due within a year: a backlog of about one quarter of a year's revenue at the second-quarter pace. The book is short, and the filing says so: most revenue is consumption, and "a growing number" of enterprise customers sign non-cancelable annual or multi-year minimums.

Two deductions the filing makes itself. "Approximately 5% to 7%" of revenue for the nine months to September came from customers later found to be violating the usage policy or commercial terms; enforcement tightened in August and September and the company expects the share to fall. And "for the two weeks ended January 2, 2026, total customer spend declined by approximately 20%," the seasonality the company expects to show as growth slows. Models released in the six months to September produced over 80% of third-quarter revenue: the revenue rides the release cadence.

Circularity, as a number. The filing names no customer, so revenue from the cloud companies that also own, supply and distribute Anthropic cannot be measured. The other three legs can. Money in: Amazon and Google have put about $41 billion into Anthropic's notes and preferred stock through June 2026 (Amazon about $18 billion, Google about $23 billion), with $35 billion of undrawn facilities and Google's $10 billion option at the IPO. Money out: at least $114.5 billion committed to Google for compute, an estimated $125–150 billion more of TPU capacity "from TPU providers, including Google," $161.2 billion of Broadcom lease rent for TPU systems, and at least $110.0 billion committed to Amazon. For every dollar the two have invested, Anthropic has committed to spend about $5.50 with them on compute, more than $8 with the TPU estimate. Money through: 48% of first-half revenue was sold through the three clouds' marketplaces, where the CSP collects and remits and Anthropic pays a channel fee; sales, marketing and partnerships, mostly those fees, is the fastest-growing cost line at 19% of revenue. Revenue from the CSPs as customers: undisclosed; the filing says they "are also our customers" and that their "purchasing decisions may be influenced by their own competing offerings." The filing's own risk factor calls these relationships "complex dynamics that could give rise to conflicts of interest."

The filing has no gross-margin line: compute is one operating expense that "supports the full range of workloads," and the company says it plans and assesses it "on a total basis." Its functional table gives the serving margin anyway. Cost of revenue, which "substantially all" consists of compute for customer-serving workloads plus some personnel and overhead, was $752 million on $386 million of revenue in 2024, $2,773 million on $4,585 million in 2025 (a 40% serving margin) and $2,380 million on $16,286 million in the first half of 2026 (85%). Compute as a whole was $7,326 million in 2025 and $8,945 million in the half; by quarter it fell from 251% of revenue to 48%. So serving was 38% of the compute bill in 2025 and 27% in the half; the rest, $6.6 billion in six months, went to research, development and internal use, the cost of staying at the frontier. The channel fee is the other cost that scales with revenue: sales, marketing and partnerships, mostly those fees, rose to 19% of revenue in the half from 12% in 2025, $2.7 billion of the increase being fees to the cloud marketplaces.

What the filing says happens next is the hinge of the valuation: "As significant incremental compute is brought online in future periods, the associated expense will be recognized in advance of the full revenue that the compute will ultimately support, and we expect operating margin and Adjusted Operating Margin may compress in those periods." Compute is contracted 12 to 18 months ahead, is "generally fixed in the near term," and "compute and infrastructure expenses as a percentage of revenue will depend primarily on how our revenue scales relative to our total committed compute cost base." Short-term compute, at a higher unit cost, is expected to rise in the near term. The 85% serving margin is real; the operating margin is the ladder in §4 against the revenue in §2.

The other lines: personnel $3,152 million in the half, of which $2,185 million was stock compensation (13% of revenue) and $967 million cash (6%); other operating and administrative $1,714 million (11%), including the donation-matching charges that reached $2.0 billion in the third quarter alone. Unrecognized stock compensation stood at $26.6 billion on restricted stock units (of which $2.9 billion relates to units whose service condition is satisfied or partially satisfied) plus $2.8 billion on options, with $5.8 billion more granted after June. The model expenses stock compensation, at 10% of revenue in 2027 falling to 6%, and does not also dilute for it.

The public filing leaves the aggregate committed spend blank ("more than $ "), the September-30 contractual-obligations table blank, and the amount due in the next twelve months blank. It names the agreements:

That is $657 billion named, or $782–807 billion with the TPU estimate. The filing's own category totals run higher because they include data-center leases, colocation and smaller agreements: about $265.6 billion of fixed or minimum commitments at June 30, $321.4 billion of leases signed but not yet commenced (from $50.2 billion six months earlier), and about $245 billion of new minimum obligations plus $21 billion of short-term agreements entered after June 30, together $853 billion. For scale: the last table the filing fills in, at December 31, 2025, totaled $54.6 billion. The $518 billion the Reuters leak reported on September 28 appears in no draft, no exhibit and not in the public filing (section 12 of the S-1 draft; the leak's counterparty figures match the August and October drafts, and the public figures are larger). Akamai's $11.6 billion commitment, announced September 24, is not named.

Scheduled by the house from each agreement's terms (ramping monthly rates for the cloud contracts, SpaceXAI at $3.1 billion a month from December 2026 to its cap, the Broadcom leases from ready-for-service), the named ladder is $25 billion in 2026, $93 billion in 2027, $156 billion in 2028, $160 billion in 2029, $119 billion a year in 2030–31 and $50 billion in 2032. Compute expense was $7.3 billion in 2025 and $8.9 billion in the first half of 2026. The company describes the reason: it is "building out our own compute and infrastructure and vertically integrating across the compute supply chain," under a program it calls Anthropic Direct that leases chips and hosts them in dedicated facilities; its first data center opened this year.

How it is paid for, and the strings. Beyond the $131.7 billion of cash: Google's $20 billion facility to December 2030 (draws tied to third-party compute agreements and revenue targets, settled in Class N shares at the 30-day average price, a 14.99% ownership cap, reduced if Google buys $10 billion at the IPO); Amazon's $15 billion for 30 months after the IPO, in $5 billion steps as capacity milestones are met, with a 12% cap; Broadcom's $42 billion solely to pay rent under the 2026 lease, at 9% paid in kind and convertible at 90% of the average price, rising to 13% and 85% "during a period in which certain revenue performance conditions are not met," with 150% of principal in cash on a change of control; AMD's up to $5 billion of stock; a $15 billion revolver with a $5 billion minimum-liquidity covenant from the fourth quarter of 2026; and the IPO, for which the wires say up to $100 billion and the filing says nothing. The parent guarantees its two leasing subsidiaries; up to $0.8 billion of data-center guarantees and a $500 million power-contract guarantee sit beside them. The filing says the build-out will "be a significant use of net cash in operating and investing activities for the foreseeable future" and that "net cash used in these activities may increase regardless of whether our financial performance improves."

In the model, compute cost in each year is the larger of a target share of revenue and the cash floor of the ladder, delivery-adjusted by case: the firm contracts (Microsoft, Google, Amazon, the Broadcom leases) in full in every case; SpaceXAI in full through 2027 and then in full, at 85%, or not at all (bull, base, bear) because either party can end it on 90 days' notice; the capacity-dependent agreements (Nscale, Lambda, AMD) in full, at 85%, or not at all; the TPU estimate in full, in full, or at half. The floor binds in 2027–29 in every case and through 2031 in the bear case. In the base case the company loses $24 billion, $47 billion and $17 billion in 2027–29 as the ladder lands, then earns 19% in 2030 and 23% in 2031. In the bear case it loses money through 2031 and burns about $156 billion in 2027–31, more than its cash: that is the case in which the facilities, the revolver and the IPO proceeds are not optional. The floor costs $5.22 a share, weighted, against a model that lets compute be a share of revenue.

The count. The income statement's weighted average of 2,025 million shares counts only the common. At the IPO the preferred (8,559 million as converted, after the ten-for-one split), the Amazon notes (1,840 million Class N at fixed prices of $2.09 and $5.05, converting "at the lower of" those and the IPO price, within Amazon's 12% cap), the options (1,797 million at an average exercise price of $2.89), the restricted stock units (1,050 million, which vest on effectiveness) and the Public Benefit Fund's 847 million Class N shares join the 2,188 million common outstanding: 16,281 million economic shares, or 16,427 million with the 146 million issued after June (new unit grants, the October donation-matching settlement, two acquisitions). That is the count this note divides by; it reconciles to the filing's own "fully diluted" basis of 16,937 million once the unissued plan reserves are added (847 million is 5.00% of it). Not counted: the IPO's new shares, Google's $10 billion placement, AMD's $5 billion and any facility draws, all blank or future. The Series H price of $58.90 is $968 billion on this count, the "$965 billion" reported in May.

The wedge. Class A (one vote; 9,410 million at the IPO) holds 57.8% of the economics but at most 49.9% of the general vote, by construction. Class N (6,871 million) holds 42.2% of the economics and no vote; it is the strategic investors, the Amazon notes and the fund, and converts to Class A only on transfer to an unaffiliated holder. Class F, one share held by the founders' LLC, carries 50.1% of every general vote, a share the filing says "will not be impacted by future issuances of our capital stock that may cause economic or voting dilution." Class T, one share held by the LTBT, elects the Class T directors, "a majority of the authorized members of our board of directors"; the founders' letter puts it as four of seven. The company is a Nasdaq "controlled company" by virtue of the trust. What a Class A holder can decide on matters put to stockholders generally: nothing the Founder LLC opposes, and not the board majority.

The trust and the sunset. The LTBT is a Delaware purpose trust of five trustees, one seat vacant: Neil Shah (chair; also an initial manager of the Public Benefit Fund), Richard Fontaine, Ben Bernanke (since June) and Christopher Stubbs (since October), each paid $400,000 a year by the company, which funded the trust with $8.4 million this year. Trustees may hold no stock; they choose directors and "neither the LTBT nor the Trustees have decision-making authority in the day-to-day management or strategy setting." The trust's terms change only by 80% of the board and 80% of the trustees, or, over the trustees' objection, the founders' LLC plus a majority of Class A plus a vote of an employee class (Class E) to be issued, or 85% of Class A if the founders' share is gone. The founders' share is redeemed 450 days after fewer than three "Continuing Founders" remain for 90 consecutive days, or earlier if a supermajority of the continuing founders agrees to redeem it or the Founder LLC dissolves; a founder stops counting on resignation, death, removal for cause, an approved leave, or holding less than 2% of what they held at the IPO. Seven today; the structure survives the loss of four. The draft-by-draft record shows the structure was assembled during the SEC process: no Class F existed in the June 1 draft; it entered on July 10; Class E, the amendment threshold and a first founder sunset on August 10; the 90-day test and the 450-day transition in the public filing.

Who benefits most. The non-voting strategic class holds 42% of the economics, Amazon and Google first among them (their stakes are not disclosed; the "21% and 15%" in the press is reporting). The seven founders hold 10.4% of the economics, 1,686 million Class A shares, and have pledged 80% of it to philanthropy (9% of fully diluted shares, the letter says) while holding, through one share, a majority of every general vote. The public holds Class A economics and a vote that cannot decide. The Public Benefit Fund holds 5%, and the board may issue up to 2% more each time the share price closes 50% above the last milestone for 90 of 120 trading days, starting from the IPO price. Prices paid, per post-split share: $3.00 (2024), $5.61 (early 2025), $14.10 (September 2025), $25.91 (February 2026, and the April employee tender), $58.90 (May–June 2026).

The Amazon notes, and the loss they explain. Amazon's $6.2 billion of convertible notes converts at the IPO into 1,840 million Class N shares at $2.09 and $5.05 a share, either side of the $3.00 paid in the 2024 round, 11% of the pre-offering economic count, subject to Amazon's 12% ownership cap, with any excess rolling over and converting quarterly at the lowest of the fixed price, the IPO price and the ten-day volume-weighted average price. At the Series H price those shares are worth about $108.4 billion, against the $108.6 billion the balance sheet carries the notes at; the $72.4 billion charge in the first half is the notes' fair value rising with the company's valuation, from $41.7 billion at the start of the year, with $5.5 billion of notes converting along the way. In the house's reading it is a transfer inside the capitalization table rather than a cost of the business: the operating loss in the half was $0.7 billion, and the second quarter earned an operating profit of $974 million. The charge stops on the portion of the notes that converts; any balance the 12% cap leaves as rollover notes keeps being marked until it converts.

When the holders may sell. The lock-up is not one cliff but a schedule of timed releases in four holder groups, each step a percentage of that group's own lock-up holdings (shares, options, units and convertibles). The filing leaves blank the share count of every step and the fully diluted share of the two stockholder groups, and gives no size for the other two; those are filled at the amendment. The percentages, in the filing's underwriting section, are not blank. Counted from the date of the prospectus and from each results release:

The reader's use for it: supply arrives in steps from the first week and no single day carries it all; the two 180-day groups are fully free at the half-year mark, the holders above 0.5% are nine-tenths free before the year is out, and the founders may sell up to 29% at the half-year mark, about 489 million shares and share-equivalents, 3.0% of the pre-offering economic count, a ceiling rather than a plan, with each founder having made a voluntary, revocable pledge to dedicate 80% of personal equity to philanthropy. The founders' remaining 71%, about 1.2 billion share-equivalents on the beneficial-ownership count, is the last step, taken on the same trigger as the 9.1% remainder of the holders above 0.5%; which step is largest depends on group sizes the filing leaves blank. Which holders sit in each group, how many shares each step is, and how many days of trading each would take, are the amendment's and the first prints' to settle.

The filing's own evidence that frontier position is the moat: models released in the six months to September produced over 80% of third-quarter revenue; "with each release, we have seen most customers utilize the latest version"; 250 model versions, features and applications launched in nine months; the cyber benchmark ExploitBench went from 40% (Opus 4.8) to 78% (Mythos 5) to 91% (Opus 5.5) in about four months. Each model helps build the next ("closing the loop"), and Claude writes over 90% of the company's own code. Four chip families (AWS Trainium, Google TPU, NVIDIA through Microsoft, SpaceXAI, Nscale and Lambda, AMD from 2027) run under an orchestration layer the company says "took years and significant investment to build" and "reduces dependence on any single vendor." Distribution runs through all three clouds, systems integrators, over 150,000 certified partner practitioners, and the case studies in the filing (Salesforce: over 60% of 83,000 employees on Claude, output per developer up 151%; Mercado Libre: 40% of 2025 production code; Verizon 33,000 engineers; Lilly 50,000 seats). Ramp's card data had 43.8% of US businesses paying Anthropic against 40.6% paying OpenAI in September 2026, and 60% of token spend in October against 36%; a card-spend sample, not the market.

The same cadence has a brake its owner controls. Mythos was held to a gated preview because of "its advanced capabilities in cyber vulnerability discovery"; Fable ships with safeguards that route some requests "to a less capable Claude model"; the company "recently chose to delay pre-training a new frontier base model for more than a month" for weight security; and in June 2026 the Commerce Department "imposed worldwide export restrictions on our Fable 5 and Mythos 5 models, which we disabled for all customers until those restrictions were lifted." The filing says future models "exhibiting such capabilities may be subject to access constraints" and that the commercial viability of gated programs "may depend on sustained governmental and industry coalition support." This is the one moat whose owner has said in writing it will narrow it when it judges it must.

What erodes it: price, first. On Ramp's token-spend sample, the blended price paid for Anthropic tokens fell 19% in a month to $0.80 per million in the week to October 11; open-weight models price below it (DeepSeek V4 Flash at $0.14 and $0.28 per million tokens in and out, about a quarter of that blended price; GLM-5.2 at $1.40 and $4.40 and Qwen 3.6 at about $0.33 and $1.95 are not cheaper than that blended price on the output leg); the BenchLM index has frontier prices up 61% in a year and mid-tier prices down 41%, the split Anthropic's four-class pricing exploits and open weights attack from below. The filing concedes the mechanism in its own words: open-weight models "could exert pricing pressure on our offerings", and customers "may conclude that incremental performance gains at the frontier do not justify the additional cost". So far the threat is to price, not to share: on Ramp's adoption data from more than 85,000 US businesses, 0.3% paid DeepSeek in September 2026 against 43.8% paying Anthropic, and the model carries the threat as the compute share of revenue, 60% in 2027 in the bear case. Second, the clouds that distribute Claude also "develop and deploy competing AI models," and the distribution deals "give them visibility into our pricing and commercial terms." Third, OpenAI (a reported $852 billion mark in March, seeking about $1.4 trillion, listing "not 2026") and Google. Fourth, talent that can "join or found competing companies immediately."

The directors of a Delaware PBC must balance stockholders' pecuniary interests, the interests of those materially affected, and the stated purpose; "there is limited judicial precedent." On February 27, 2026, "in response to our publicly stated position that our technology should not be used for mass domestic surveillance or fully autonomous weapons," the President directed federal agencies to stop using the company's technology and the Department of War designated it a supply-chain risk; the company won most of its case in the Northern District of California on August 27, lost 2-1 in the D.C. Circuit on September 25, and has asked for rehearing. Government revenue is not quantified. The June export restriction (above) "could recur." The FTC issued a civil investigative demand in October on consumer protection, privacy "as well as the safety and security of our products," and Axios reported on October 26 that the FTC has demanded testimony from Anthropic executives. The external evaluators the company indemnifies, "without a cap on our liability" in some cases, are unnamed in the filing; anthropic.com named a six-person METR team with employee-like access on October 15. The copyright docket: Bartz settled for $1.5 billion (approved July; $1.2 billion still accrued at June 30), the first Concord case was argued on summary judgment on October 21, a second wave of music-publisher suits names two executives, Reddit's contract claims survive, four Canadian actions and the Bartz opt-outs continue; none is quantified. The company endorsed California's SB 53, tests with the US and UK AI safety institutes, and faces EU general-purpose-model enforcement from August 2, 2026. The filing mentions lobbying once and gives no figure. The company's own reports to the Senate under the Lobbying Disclosure Act show federal lobbying spend rising from $0.36 million in the first quarter of 2025 to $2.11 million in the third quarter of 2026 ($1.56 and $1.97 million in the two quarters before it, with one dip in the fourth quarter of 2025), with nine lobbyists named in the latest report and defense, energy and consumer safety among the issues listed; reported giving to a political group this year is $40 million.

Every length below is an argument a reader can dispute; the sensitivities in §9 give the value of each alternative. The rule is the house's: a stage lasts as long as the thing that drives it, dated by an observable, and is checked against base rates (across US companies in 1951–98, above-median sales growth survived five years running in 6.3% of firms and ten years in 0.9%).

Terminal growth is 4.0% a year in the base and bull cases, the Federal Reserve's September 2026 longer-run projection of real growth (2.0%) plus inflation (2.0%), which a company that has become a large share of knowledge-work spending cannot outgrow forever; 3.0% in the bear case, for a company losing share. Every fade lands exactly on its terminal rate.

Against the company's own talk. The base case reaches $113 billion in 2027, $167 billion in 2028, $303 billion in 2031 and $450 billion in 2036. Reuters reported in August that the IPO case "rests on $190–200B revenue forecast for 2028": the base case sits about 14% below it, the bull case ($244 billion) about 25% above, the bear case ($92 billion) at half. The FT's reported investor expectation of a $100–120 billion exit run-rate for 2026, and the company's own estimate at its October 15 investor meeting that the run-rate passes $100 billion by year-end (from the $67 billion reported at the meeting, as Damodaran relayed it on October 22), imply a fourth quarter of $25–30 billion; the house's bull-case fourth quarter is $23.8 billion, base $19.8 billion, bear $17.0 billion, because the filing says growth decelerates, the year-end dip is 20% over two weeks, and 5–7% of revenue is being purged. The length the price implies: at the house growth rates no stage length up to nine years reaches the Series H price (nine years gives $58.39, fifty cents short, because the stage ends at a 15% rate); the price is not a longer boom but a faster one: 3.2 points more growth in every stage-one year of every case for Series H, 12.8 for $1.5 trillion, 19.4 for $2 trillion.

Method. Free cash flow to the firm over ten explicit years (2027–36) on a 2026 stub built from the first-half actuals and the implied third quarter, discounted at mid-year, with one cost of capital in every case, stock compensation expensed, terminal value by Gordon growth, cash of $131.7 billion plus a $2 billion house estimate of second-half cash generation added, the $1.2 billion Bartz accrual and $0.4 billion of finance leases deducted, divided by 16,427 million shares; the terminal value is discounted from the end of 2036 (discounting it from mid-year too would add $1.85 a share, $52.58). A dividend model has nothing to discount; free cash flow to equity was negative in every year through 2025 and the capital structure changes at the IPO; a multiple on revenue growing 1,300% is a growth assumption in disguise, so it is a cross-check below, never blended.

Cost of capital 12.25%. The ten-year Treasury at 5.00% on October 23 plus a beta of 1.51 times the house's 4.8% equity risk premium; the beta is the midpoint of the hyperscalers' median of 1.24 (Microsoft 1.09, Alphabet 1.15, Amazon 1.32, Oracle 1.58) and the AI pure plays' 1.78 (NVIDIA 1.58, Palantir 1.99, CoreWeave 1.56, Nebius 2.22), two-year weekly betas against the S&P 500 after the Blume adjustment; no interest-bearing debt survives the IPO beyond any Amazon rollover notes held back by the 12% cap, which convert quarterly, and $0.4 billion of finance leases. The house used 10% for SpaceX, a company with a cash-generating Starlink business and a listed market history; Anthropic has neither, and a beta set from its own peers is the honest substitute. Every result is shown at 10%, 11%, 12.25% and 13%. Tax is 21% after the $8.2 billion of federal net operating losses the filing reports at December 31, 2025, plus the first-half 2026 loss.

Cost structure, base case (share of revenue, 2027 → stage end → terminal). compute 52% → 42% → 38% (first-half actual 55%, second quarter 48%; at maturity about 15 points to serve and 25 to research, a hyperscaler cloud's cost structure); channel and marketing 17% → 12%; cash personnel 6% → 5%; stock compensation 10% → 6%; other 7% → 4%; public-benefit spend 1% throughout. Terminal operating margin 34% base, 41% bull, 23% bear.

Weighted 25/50/25, fair value is $50.72 a share, $833 billion. The weights are the house convention for a triad whose base case is the most likely path and whose tails are judged equally likely; a reader who weights the bull case at a third and the bear at a sixth gets $58.49, within a dollar of the Series H price. The bear case's terminal value exceeds its operating value because the explicit years are losses: enterprise growth slows to 55% in 2027 and 22% in 2028, the firm contracts land anyway, the company loses $39 billion, $33 billion, $30 billion, $24 billion and $18 billion in 2027–31, and the equity is mostly the cash on hand plus a 23%-margin business a decade out. The base case loses money in 2027–29 while the ladder lands, then earns 19%, 23%, 25%, 27%, 30%, 32% and 34%. The bull case grows into the ladder by 2029. The proponent's case, on its own inputs (bull growth, a 10% cost of equity, a seven-year consumption stage) gives $167.62, $2.75 trillion; the house weights that path at a quarter and discounts it at 12.25%.

The read from the business. Return on capital is −342%, −246% and −138% at the three quarter ends, with the spread to the hurdle widening from −354 to −150 points: the mechanical label is recovering. The loss narrowed because first-half revenue grew 1,300% against compute's 283%. The lens is the wrong one for a company whose capital sits in its vendors' data centers until Anthropic Direct capitalizes it, and when the 2027 deliveries land the series will fall before it rises.

Sensitivities (weighted, $ a share). Cost of equity 10% / 11% / 13%: $74.17 / $61.80 / $45.63. Terminal growth half a point either way: $48.37 / $53.38. Consumption stage three, seven or nine years: $36.64 / $54.58 / $58.39. Stage-one growth ten points a year lower or higher in every case: $29.34 / $80.65. Compute share of revenue five points higher or lower every year: $43.23 / $57.84. Revenue path at 0.8 or 1.2 times: $38.08 / $62.03. No commitment floor at all: $55.94. End-year instead of mid-year discounting: $50.08. The single greatest point of sensitivity is stage-one growth; the second is the cost of capital.

What each anchor requires. The Series H round ($58.90; $968 billion) is the house path at an 11.3% cost of equity, or 1.14 times the path at 12.25% (base revenue of $347 billion in 2031 instead of $303 billion; 46% compound growth in 2026–31 instead of 42%), or 3.2 points more growth a year, or compute 5.8 points lower as a share of revenue every year. $1.5 trillion is the path at 9.0%, or 1.75 times it ($529 billion in 2031; 59% compound), or 12.8 points more growth a year, or compute 31 points lower. $2 trillion is the path at 7.9%, or 2.38 times it ($722 billion in 2031; 69% compound), or 19.4 points more growth a year; no reduction in the compute share reaches it. Damodaran's own break-even, in his August 20 post, is the same arithmetic on a ten-year horizon: at the "rumored pricing for the IPO" of $2 trillion, "allowing the company premium pricing margins (after-tax operating margin of 30%) and above-average risk (cost of capital of 10%), the company will have to generate close to $1.2 trillion in revenues, if the AI market matures in ten years, and close to $2 trillion, if the wait is 15 years"; the house path reaches $450 billion in 2036. On October 22 he wrote that a case for $1.5 or $2 trillion needs "a surge in operating margins to 30% or higher", a story he calls a reach but "a plausible one", and that his own valuation waits for the prospectus.

The peers, as a cross-check. On trailing revenue and the October 23 closes, NVIDIA trades at 17.1 times sales (83% growth, 65% operating margin), Palantir 75 times, Snowflake 21, Microsoft 11.6, ServiceNow 11.4, Alphabet 8.5, CoreWeave 7.9, Oracle 6.5, Amazon 3.5; the median is 11.4. Anthropic's trailing revenue to June ($19.7 billion) makes a trailing multiple meaningless; on the second-quarter run-rate of $46.2 billion the house fair value is 15.2 times, Series H 18.1, $1.5 trillion 29.6 and $2 trillion 40.4; on 2027 base revenue of $113 billion, 6.2, 7.4, 12.1 and 16.6 times. The peer median applied to 2027 revenue gives $1.28 trillion of enterprise value, which is the problem with the method in one line: it pays NVIDIA's multiple on a year that has not happened. The private marks are a separate band: Anthropic's rounds at $183 billion (September 2025), $380 billion (February) and $965 billion (May); OpenAI's reported $852 billion (March) and the $1.4 trillion it is reported to be seeking; xAI's $250 billion inside SpaceX; and the $1.5 trillion Reuters reported investors discussing on October 15 after third-quarter growth slowed to 45%, down from the $1.5–2 trillion range in the Wall Street Journal.