Anthropic (ANT) S-1 Analysis: $11.6 Billion in a Quarter, an 85% First-Half Gross Margin, and a Compute Bill That Says the Profit May Not Last

Filing Snapshot: Results vs. Expectations

There is no consensus to beat. Anthropic has never reported as a public company, the S-1 contains no guidance, and the price range, share count, September-quarter revenue and total committed compute are all blank in this preliminary filing. The only meaningful comparison is with its own history, and that history is unusually well documented: three audited years, two unaudited half-years and six unaudited quarters that foot exactly to both.

Anthropic, PBC is a Delaware public benefit corporation founded in January 2021 by seven co-founders, all of whom previously worked at OpenAI. It sells access to its Claude models through its own platform and applications, including Claude Code and Claude Cowork, and through the three largest cloud marketplaces. Ernst & Young audited 2025; Grant Thornton audited 2023 and 2024. It had approximately 5,300 employees at September 30, 2026.

The Print

The Gross Margin the Income Statement No Longer Shows

Anthropic reports its expenses by nature, compute, personnel and so on, rather than by function. It changed to that presentation during 2025, and in doing so removed cost of revenue and gross profit from the face of the income statement. The functional view survives in a supplemental note, and it is the most useful table in the filing for understanding the business model.

"During the year ended December 31, 2025, the Company changed its presentation of operating expenses from a functional classification under which cost of revenue, gross profit (loss), research and development, sales and marketing, and general and administrative were presented as separate line items, to a classification based on the nature of expenses."
— Anthropic, PBC Form S-1, Notes to Consolidated Financial Statements

"Substantially all of cost of revenue consists of compute and infrastructure costs associated with customer-serving workloads."
— Anthropic, PBC Form S-1, Notes to Consolidated Financial Statements

The table resolves what looks like a contradiction. Serving paying customers cost $2,380 million in the first half of 2026, a derived gross margin of 85.4%, up from 39.5% for 2025 and 12.4% a year earlier. That is a software-grade margin. Research and development, which the note says consists primarily of compute for model development, research and internal usage plus research personnel, cost $9,741 million, 59.8% of revenue. Sales and marketing includes the compute behind free access to Claude as well as channel fees. At most about 27% of the $8,945 million of compute and infrastructure expense in the half sat in cost of revenue.

Revenue Assessment

The top line is not in question. Revenue of $16,286 million for the half grew 1,300.3%, the June quarter alone produced $11,555 million, and the filing implies the September quarter added about $16.9 billion more. The quality questions are about durability rather than existence: 80.2% of first-half revenue was consumption-based and priced per token, roughly 48% came through cloud marketplaces, and the company says over 80% of September-quarter revenue came from model versions less than six months old. We take those up under Revenue Composition and the Key Topics below.

Margin Assessment

Two margins tell two different stories, and both are true. The gross margin of 85.4% says inference is already a highly profitable product. The operating margin of negative 4% for the half, and positive 8% for the June quarter, says the cost of staying at the frontier absorbs almost all of it. Research and development fell from 138.2% of revenue in 2025 to 59.8% in the first half because revenue grew faster, not because spending slowed: it rose 426.5% year over year. The company says operating margin for the first nine months of 2026 expanded significantly year over year. Whether it keeps expanding depends on whether research compute starts to grow faster than revenue again, and the compute book discussed below suggests it will.

Earnings Assessment

Net loss is not a useful measure for this company yet. Of the $80.7 billion first-half loss, $72.4 billion is the change in fair value of convertible notes and related instruments and $8.8 billion is other expense, primarily from the 2026 note purchase agreements. Both are non-cash and both arise because the company became more valuable. Adjusted Net Income of $2,671 million strips them out, but it also strips out $2,185 million of stock-based compensation, which is a real cost. Our preferred bridge is GAAP operating income plus interest income: negative $693 million plus $1,111 million, or a derived pre-tax operating profit of about $418 million for the half, before the charges that the offering itself will trigger.

Revenue Composition

Anthropic reports a single segment. The notes split revenue three ways, by pricing model, by billing geography and by channel, and add a single backlog figure. Together they describe a business that is overwhelmingly usage-priced, mostly American, sold about half through other companies' clouds, and only lightly contracted in advance.

Consumption and Subscription

Four-fifths of revenue is pay-for-intelligence: tokens metered and priced by model. Subscriptions, which include the Claude Pro and Max plans, grew faster in the first half, 1,690% against 1,229%, and rose to 19.8% of revenue. The filing describes the two as a bridge, with subscribers who exhaust a plan's usage buying more capacity directly. A growing number of enterprises sign annual or multi-year minimum-spend contracts, but the company is clear that these still result primarily in consumption revenue.

"However, our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years."
— Anthropic, PBC Form S-1, Risk Factors

Assessment: that sentence is the mismatch at the center of the model. Revenue can be turned down in a week; the compute that serves it is committed for years. Consumption pricing is also why growth has been so fast, because a customer can scale from a pilot to production without renegotiating anything. It cuts both ways, and we would expect quarterly revenue to be visibly more volatile than a seat-based software company's once growth slows.

Channel

"During the six months ended June 30, 2025 and 2026 (unaudited), revenue generated from sales through CSP marketplaces to customers represented approximately 44% and 48% of total revenue, respectively."
— Anthropic, PBC Form S-1, Notes to Consolidated Financial Statements

Roughly half of revenue now arrives through the marketplaces of Amazon, Google and Microsoft, up from about 11% in 2023 and 32% in 2024. The cloud providers act as collection agents: at June 30, 2026, receivables from marketplace sales were 58% of Anthropic's net receivables. Separately, amounts payable to cloud service providers, which are also its main compute suppliers, were $2,810 million. Anthropic keeps the customer contract and the pricing, acts as principal and records the revenue on a gross basis, and pays each provider a fee, which sits in sales, marketing, and partnerships.

Assessment: the channel is a genuine distribution advantage, because enterprises can buy Claude inside cloud budgets they have already committed. It also concentrates revenue, receivables, compute supply and a large part of the shareholder base in the same three counterparties, some of which compete with Claude directly. The financial statements contain no related-party note, and neither they nor the related-party section quantify revenue from, or payments to, Amazon or Google.

Geography

The United States produced 72.2% of first-half revenue and international customers 27.8%, measured by billing address. The international share was higher in 2024, at 31.9%, and has stayed between about 26% and 30% since the start of 2025 while the company expands its regional offices. The filing names international expansion as a meaningful growth driver.

Assessment: a U.S.-weighted mix is normal for an American software company at this stage and is not a concern on its own. It matters here because the largest regulatory event in the filing, the federal government's February 2026 order to stop using Anthropic's models, is a U.S. event.

Contracted Backlog

"As of June 30, 2026 (unaudited), remaining performance obligations amounted to $10.8 billion."
— Anthropic, PBC Form S-1, Notes to Consolidated Financial Statements

Remaining performance obligations of $10.8 billion, about 75% of it expected within twelve months, compare with $11,555 million of revenue in the June quarter alone. Deferred revenue was $1,650 million. No comparable backlog figure is given for any earlier date.

Assessment: contracted backlog covers less than one quarter of current revenue. For a company with roughly $265.6 billion of fixed or minimum purchase commitments at the same date, the gap between what customers have committed to Anthropic and what Anthropic has committed to its suppliers is the clearest single measure of the operating leverage, in both directions, that public shareholders are buying.

The Six Quarters the Half-Year Hides

The S-1 includes six quarters of unaudited detail. We footed every line: the four 2025 quarters sum exactly to the audited full year on revenue, each expense line, operating loss and net loss, and the two 2026 quarters sum exactly to the six-month totals. This is the company's own arithmetic, and it is where the story of the filing actually sits.

Three things in this table deserve more weight than they usually get. First, the March 2026 quarter is the hinge. Compute grew only $442 million sequentially while revenue grew $2,691 million, and the operating loss narrowed by more than $1.1 billion. That pattern, revenue compounding on a compute base that was contracted 12 to 18 months earlier, is the mechanism the company describes in its own words, and it does not repeat automatically. Second, the June quarter's incremental operating margin over March was 38.7%: $2,641 million of additional operating income on $6,824 million of additional revenue. Compute absorbed 31.8% of the incremental revenue and sales, marketing, and partnerships another 16.1%. Third, personnel fell from 58% of revenue in Q1 2025 to 16% in Q2 2026 even though stock-based compensation rose twelvefold. Revenue simply grew faster than every cost line except sales, marketing, and partnerships.

The net loss line is a different story and deserves to be separated cleanly. Every quarter's net loss is dominated by the change in fair value of the convertible notes Anthropic issued to Amazon and Google, which are marked to the company's own valuation. When the Series H round lifted that valuation in the June quarter, the mark-to-market charge reached $59.6 billion, on top of a further $8.4 billion of other expense tied to the 2026 note purchase agreements. The $66.3 billion net loss in a quarter that produced a $974 million operating profit is accounting for the company becoming more valuable, not cash leaving the business.

Assessment: the quarterly table makes the bull case and the bear case at once. The bull case is the revenue line: $376 million to $11,555 million in five quarters, with growth re-accelerating in 2026 rather than fading. The bear case is the compute line: it added fewer dollars than revenue in only the last two quarters, and the commitments discussed below suggest it is about to grow much faster. Anyone annualizing the June quarter's 8% operating margin is annualizing the best ratio of revenue to compute this company is likely to report for some time.

The Compute Book: The Forecast the Filing Does Not Print

The S-1 contains no guidance, and the two numbers that would come closest to one are blank: the total of Anthropic's committed compute spend, and the by-year table of contractual obligations. What the filing does print, agreement by agreement, is the size and term of its largest compute contracts. Added up, those contracts are the most informative forward-looking disclosure in the document, because nobody commits several hundred billion dollars to compute without a view of the revenue it will serve.

"In addition, the majority of our compute is contracted under multi-year commitments that we typically procure 12 to 18 months in advance of availability. When customer demand or research needs exceed this long-term capacity, we procure short-term compute, which typically carries a higher cost per unit of compute."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

The annualization is ours. Contract values are divided by their stated terms, the SpaceXAI figure is the company's own monthly run-rate multiplied by twelve, and the Nscale, Lambda and AMD terms are our assumptions where the filing says only "into" a year. Real schedules will be lumpier than a straight line. The direction is not in doubt, though, and the filing adds that Anthropic is party to other compute arrangements and leases it does not describe, so the table omits some obligations, although several rows are maximums or can be cancelled.

The notes to the financial statements give category totals that go further than the named agreements. At June 30, 2026, Anthropic had approximately $265.6 billion of fixed or minimum purchase commitments and approximately $321.4 billion of leases that had not yet commenced. After June 30 it entered into new, and amended existing, agreements with approximately $245 billion of minimum payment obligations, plus approximately $21 billion of short-term commitments. On our arithmetic those categories add to roughly $853 billion before the $125 billion to $150 billion estimate for 2028 to 2031 deliveries, which takes the total toward $1 trillion. Because some of the later agreements amend earlier ones, the categories may overlap. The company does not add them up itself, the MD&A total is blank, and some arrangements are subject to conditions. At year-end 2025 the equivalent figures were $54.6 billion of purchase commitments and $50.2 billion of leases not yet commenced.

"In addition to the transactions above, subsequent to June 30, 2026, the Company entered into new, and amended existing, lease and commercial agreements primarily to secure additional compute and infrastructure capacity, with aggregate future minimum payment obligations of approximately $245 billion, with some arrangements subject to certain contractual conditions being met."
— Anthropic, PBC Form S-1, Notes to Consolidated Financial Statements

The nearest step is the most concrete. SpaceXAI, formerly X.AI LLC and now a subsidiary of Space Exploration Technologies Corp., is expected to deliver full capacity in December 2026 at roughly $3.1 billion a month. That is about $9.3 billion a quarter for one supplier, against $5.6 billion of total compute and infrastructure expense in the June quarter. Microsoft's commitment starts in November 2026. Broadcom's TPU leases add more capacity from 2027. Unless revenue keeps compounding at the 2026 pace, the compute-to-revenue ratio that produced the June quarter's profit is likely to move back up in the first quarters Anthropic reports as a public company.

"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."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

The company also expects its use of short-term compute, which costs more per unit, to increase in the near term and to compress margins. It adds that this cycle will make margins fluctuate even if the underlying economics of the business continue to improve.

Two features of the contracts cut the other way and deserve equal weight. The SpaceXAI agreements become cancellable by either party on 90 days' notice after their initial periods, and the Nscale and Lambda fees accrue only as capacity is delivered and accepted, so a meaningful slice of the book is flexible rather than take-or-pay. And the company is moving part of its compute to what it calls Anthropic Direct, leasing chips and hosting them in dedicated facilities, which separates the data center commitment from the accelerator commitment and defers the larger outlay. The cost of that flexibility is a different cash flow profile: the company expects capital expenditures to increase significantly and to produce significant net cash use for the foreseeable future.

"Our compute commitments are substantial and generally must be satisfied regardless of our revenue or utilization, while our access to additional capital depends on a number of conditions, many of which are outside our control, including market conditions and investor demand."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

The funding to meet the book is substantial. Anthropic had $112.4 billion of cash and cash equivalents and $19.3 billion of short-term investments at June 30, 2026, after raising $141.1 billion of capital to date. In September it replaced a $3.25 billion revolver with a $15 billion facility that carries a $5.0 billion minimum liquidity covenant from December 2026. Note purchase agreements give it up to $77.0 billion of further partner financing: $20.0 billion from Google, $15.0 billion from Amazon in $5.0 billion increments tied to compute milestones, and $42.0 billion from Broadcom to finance rent on the TPU leases. AMD has committed to buy up to $5.0 billion of stock as its GPUs are deployed. Operating cash flow turned positive at $3.9 billion in the first half of 2026.

The Broadcom facility is the one to read closely. Its notes accrue interest at 9.00% paid in kind and convert into stock at 90% of the 30-day volume-weighted price. The conditions for drawing it include revenue growth tests the company says sit significantly below its recent growth rates, and the penalty for missing them is explicit.

"Any notes issued during a period in which certain revenue performance conditions are not met bear interest at 13.00% per annum and convert at 85%, rather than 90%, of the applicable price."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

It is worth being precise about what most of this compute is for. In the first half of 2026, cost of revenue, which the company says is substantially all compute for customer-serving workloads, was $2,380 million, against $8,945 million of total compute and infrastructure expense. The remainder trained and evaluated models, ran research and internal usage, and served free users. That split is why an 85% gross margin and a near-zero operating margin can describe the same half-year. It also means the commitments are not simply capacity waiting for demand. A large share is the cost of the next models, which has to be paid whether or not the next models sell.

Assessment: the compute book is both the strongest evidence of management's conviction and the main reason we do not annualize the June quarter. A company that has signed for up to roughly $637 billion by name, and roughly $853 billion of fixed or minimum obligations by category on our reading, has built for a revenue base of roughly $250 billion to $325 billion a year once the contracts are all running, on our arithmetic. That is a coherent plan given a revenue line that grew from $376 million to $11.6 billion a quarter in five quarters. It is also a plan with an asymmetric failure mode: if demand or pricing disappoints, a large share of the cost stays, and the Broadcom financing that backstops part of it converts into stock at a discount to the market price. The IPO price is a bet that the plan is delivered in full. Our valuation gives it partial credit.

Key Topics from the S-1

1. The Headline Growth Claim Ends in a Blank, and the Filing Lets You Fill It

"Claude has experienced the fastest enterprise adoption of any technology in history, as demonstrated by revenue growing from $0 at commercial launch in the first quarter of 2023 to $ for the three months ended September 30, 2026."
— Anthropic, PBC Form S-1, Prospectus Summary

The headline growth claim in the prospectus overview is missing its number. The September quarter will be filled in by amendment, but the filing already pins it down. The Business section says September-quarter revenue was 86 times the revenue of the December 2024 quarter, and Risk Factors gives that quarter as $197 million. That puts the September quarter at about $16.9 billion, up about 47% sequentially. Two other disclosures agree: trailing revenue per employee of over $6.6 million with approximately 5,300 employees implies at least roughly $16.7 billion, and trailing revenue rose approximately 1,200% between September 2025 and September 2026.

Assessment: the 86-times figure is rounded, so the derived quarter could sit anywhere from about $16.8 billion to $17.0 billion, and it is consistent with the September-quarter figure reported in the financial press. Sequential growth therefore slowed from 144% in the June quarter to the mid-40s in September. That is still extraordinary. It is also a return to the pace of the December 2025 quarter after two quarters of acceleration, and it is consistent with the deceleration the company itself predicts.

"Our recent growth may reflect early-stage adoption dynamics, the nascency of the AI market, and rapid initial customer expansion that may not be sustainable over time. We expect our revenue growth rate to decline over time, and it may decline significantly and more rapidly than we anticipate as our revenue base increases and as the market matures."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

2. Customer Breadth Is Real: 11,900 Customers Above $100,000, and Over Half the Fortune 500

Customers spending at least $100,000 over the trailing twelve months grew from about 20 at the end of 2023 to about 200 at the end of 2024, about 1,500 at the end of 2025, more than 6,000 at June 30, 2026 and more than 11,900 at September 30, 2026. The count roughly doubled in the third quarter alone. More than 1,900 customers spend over $1 million and approximately 250 spend over $10 million. Beneath them sit approximately 40 million customers each spending less than $100,000.

"Over 50% of the Fortune 500 and over 30% of the Global 2000 were customers with $100,000 or more of revenue on a TTM basis as of September 30, 2026."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

The company adds that its revenue from those Fortune 500 customers averaged less than 0.1% of their total expenses in the September quarter, and less than 1% for its ten largest Fortune 500 customers. That is the runway argument, and it is a fair one.

Assessment: breadth is the single strongest defense against the concentration bear case. The more important question is what each customer is worth once the adoption wave passes, and the filing does not answer it. Net revenue retention of approximately 900% compares one quarter's revenue with the same quarter a year earlier for customers who existed then, excluding Pro and Max subscribers. At this stage of adoption the figure says more about how early those customers were than about how sticky they will be.

3. The Revenue Is Only as Durable as the Last Release

"For example, model versions released in the six months ended September 30, 2026 accounted for over 80% of our revenue for the three months ended September 30, 2026."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

This is the most important sentence about revenue quality in the filing. Customers move to new models quickly, which is why revenue has compounded with each release. It also means a release that slips, disappoints, or is restricted for safety reasons lands in revenue within a quarter. The company has already restricted releases, and says it may do so again.

"In the past, these steps have included delaying the release of new models and limiting their initial availability, such as our initial release of Mythos through Project Glasswing and our later release of Fable 5 as a generally-available model with safety restrictions, and in the future, as we learn more about model capabilities and behavior, could include additional measures that affect whether, how, and when we develop or release new models or capabilities."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

Assessment: this is the trade-off at the heart of the investment case, stated by the company. The company believes its safety process is part of why enterprises buy, and the same process can delay the revenue that comes from the next model. We do not treat it as a reason to discount the business. We do treat it as a reason the revenue line deserves a lower multiple than its growth rate alone would justify.

4. Price per Task Falls Fast, and Volume Has to Outrun It

"For example, Claude Opus 5.5 reduced customer costs per task by approximately 40% compared to Claude Opus 5. Lower costs per task drive greater usage, which in turn drives broader diffusion of Claude. Our compute efficiency allows us to deliver these reductions while preserving our margins."
— Anthropic, PBC Form S-1, Prospectus Summary

A 40% cut in cost per task inside one model generation is a deflation rate few industries experience. So far volume has more than offset it, and the company says compute efficiency has preserved the margin. The risk is the shape of the curve: elastic demand makes price cuts accretive while adoption is early, and dilutive once it is not.

Assessment: we model continued price declines per unit of capability and assume volume outruns them through 2028 in the base case. A generation in which the cost cut is not followed by a usage surge would be the clearest early signal for the bear case.

5. Between 5% and 7% of Revenue Came From Customers Breaking the Rules

"Based on our detection and review process, we estimate that revenue from such customers represented approximately 5% to 7% of our revenue for the nine months ended September 30, 2026."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

The company recognizes revenue from customers it later finds violating its usage policy or commercial terms, because detection takes time, and it expects that revenue to keep recurring at some level. It stepped up enforcement in August and September 2026 and expects the share to decline, while saying the size and timing of the effect on revenue and growth are uncertain. On nine-month revenue of roughly $33.2 billion, using our derived September quarter, 5% to 7% is roughly $1.7 billion to $2.3 billion.

Assessment: few companies would disclose this, and the disclosure deserves credit. It is still a headwind of several points to growth in the quarters after the crackdown, landing exactly when the compute ramp is pressuring margins. We include it in our December-quarter estimate.

6. Channel Revenue Is Booked Gross, and the Fees Sit in Operating Expense

"We have a joint go-to-market approach with these CSPs through which we establish direct contractual relationships with our customers while paying a fee to each CSP."
— Anthropic, PBC Form S-1, Prospectus Summary

Claude is sold through AWS, Google Cloud and Azure, and Anthropic records the revenue and pays the cloud provider a fee. Sales, marketing, and partnerships rose from 10% of revenue in the first half of 2025 to 19% in the first half of 2026, driven primarily by a $2.7 billion increase in fees paid to channel partners. That makes those fees at least 16.6% of first-half revenue.

Assessment: this matters most for comparisons. A competitor that reports cloud-channel sales net of the provider's share would show lower revenue and a lower expense ratio for the same economics, and press reporting attributes a gap between two reported OpenAI revenue figures to how sales through cloud providers are counted. We compare valuation on earnings for that reason.

7. The $80.7 Billion Net Loss Is the Price of Becoming Valuable

"Because these instruments are marked to market, increases in our enterprise valuation drive increases in this expense."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

The accumulated deficit of $132.9 billion at June 30, 2026 includes $113.2 billion of fair value changes on convertible notes and derivative liabilities, $72.4 billion of it in the first half of 2026. Another $8.8 billion of first-half other expense came primarily from recognizing and remeasuring the instruments created by the 2026 note purchase agreements. The Amazon notes convert at the offering, subject to Amazon's 12% ownership cap, but any portion above the cap is reissued as a rollover note that keeps being marked to market. The new Google, Amazon and Broadcom facilities create further instruments that will be measured the same way.

Assessment: most of the headline net loss is a non-cash fair-value charge rather than an operating loss, although part of it reflects value transferred to strategic partners, and it should not drive the valuation. It will, however, keep GAAP earnings volatile after the listing, and in a direction readers may find counterintuitive: a rising share price will tend to increase the reported loss.

8. Roughly $26.6 Billion of Stock Compensation Has Not Yet Been Expensed

"As of June 30, 2026, there was approximately $26.6 billion of total unrecognized stock-based compensation expense related to RSUs."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

The RSUs carry a performance condition that is satisfied when the registration statement becomes effective, so no expense has been recognized on them. Of the $26.6 billion, $2.9 billion relates to units whose service condition was already fully or partly met, and the company will record a cumulative catch-up charge in the quarter the offering completes, using accelerated attribution for the rest. Options carry a further $2.8 billion of unrecognized cost. Stock-based compensation already ran at $2,185 million in the first half, 13.4% of revenue, before any of this.

Assessment: the quarter of the offering will show a large GAAP operating loss from this charge alone, on top of the compute ramp. Adjusted operating income excludes stock-based compensation, and we prefer GAAP here: at roughly $26.6 billion, most of it to be recognized over about two years, against first-half revenue of $16.3 billion, this is a real cost of the talent model, not an accounting curiosity.

9. Five Separate Sources of Dilution

The share count will grow from at least five directions after the listing, and most of them do not require a shareholder vote.

"Following this offering, our board of directors, in its discretion, may issue shares of our Class N common stock to the Anthropic Public Benefit Fund (or other designee) representing up to an additional 2% of our then-outstanding shares (plus any shares subject to options, RSUs, and other convertible securities and any shares held by the Anthropic Public Benefit Fund) each time our share price appreciates 50% above the prior share price milestone, beginning with our initial public offering price."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

Google also holds an option to buy up to $10.0 billion of Class N stock in a private placement alongside the offering at the offering price, which would reduce its note facility by the same amount. The co-founders have each pledged 80% of their personal equity to philanthropic causes, and their combined holdings are roughly 1.69 billion shares, about 10% of the fully diluted count.

Assessment: the public benefit issuance mechanism is unusual in one specific way: it dilutes holders more in the scenarios where they are doing best. We include 2% of dilution in our base case and 4% in our bull case. The partner financing is the larger number, and it is tied to the compute ramp, so the company that grows fastest will also issue the most stock to its suppliers.

10. Customer Concentration Has Fallen Quickly, but Started High

"Approximately 79% of our revenue came from customers outside of our top five customers for the three months ended September 30, 2026, up from approximately 64% for the three months ended March 31, 2026."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

The top five customers produced roughly 36% of March-quarter revenue, roughly 21% of September-quarter revenue and approximately 25% of nine-month revenue, even though each of them reached a record level of spending in the September quarter. No single customer exceeded 10% for the first nine months of 2026, although Customer C was 11% of the first half. Customers A and C were also 13% and 10% of accounts receivable at June 30, 2026. The customers are not named.

Assessment: the trend is clearly favorable, and a 21% top-five share is unremarkable for an infrastructure business. The history is the caution: a customer that was 43% of revenue in 2023 and 11% in the first half of 2026, and two customers at 12% each in 2025, mean several large relationships are still worth billions a year. The filing's own list of reasons they could shrink includes customers building their own models and customers that depend on venture funding.

11. Claude Wrote Over 90% of Anthropic's Code Contributions in September

"For example, over 90% of the contributions to our codebase in September 2026 were written by Claude. Additionally, the average daily volume of code contributed to our codebase per engineer in the three months ended September 30, 2026 was over 45 times the average daily volume in the year ended December 31, 2024."
— Anthropic, PBC Form S-1, Management's Discussion and Analysis

Trailing revenue grew approximately 1,200% between September 2025 and September 2026 while headcount grew approximately 170%, taking revenue per employee from $1.4 million to over $6.6 million. Personnel cost fell from 58% of revenue in the March 2025 quarter to 16% in the June 2026 quarter.

Assessment: this is the best evidence in the filing for the product itself, because it is the company's own usage at scale. It also sets a high bar: the operating leverage on people is largely banked, so future margin expansion has to come from the compute line, which is exactly the line the commitments say will rise.

12. The Federal Government Ordered Its Agencies to Stop Using Claude

"For example, in February 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 of the United States stated that he had directed that all federal executive agencies cease using our technology."
— Anthropic, PBC Form S-1, Risk Factors

This is the single most unusual risk in the document, and the filing describes it in detail. On the same day as the presidential directive, the Secretary of War directed that Anthropic be designated a supply-chain risk to national security and stated that no contractor, supplier or partner doing business with the U.S. military may conduct any commercial activity with it. The Department of War also suggested it could invoke the Defense Production Act to compel the company to remove the limits it places on use of its models for autonomous weapons and mass surveillance. The company's usage policy prohibits both, and it says it has declined those uses notwithstanding the revenue they could generate.

The litigation is split. A federal district court in Northern California first enjoined the directive and the designation under one statute, and on August 27, 2026 largely granted summary judgment in Anthropic's favor, finding the actions retaliatory under the First Amendment, contrary to due process, and in excess of statutory authority. On September 25, 2026, a D.C. Circuit panel ruled 2-1 the other way on the designation made under a different statute, the Federal Acquisition Supply Chain Security Act. Anthropic has asked the full D.C. Circuit to rehear the case.

"The Company may experience material revenue losses or business disruptions attributable to these events."
— Anthropic, PBC Form S-1, Notes to Consolidated Financial Statements

Revenue from government agency contracts was approximately 1% of revenue in each period the filing presents, and total revenue more than doubled sequentially both in the quarter of the directive and in the quarter after it. The direct effect is therefore small. The indirect exposure is not: the filing itself warns that the outcome could affect sales to "other customers that sell products and services to the U.S. government," a category that, on our reading, includes the three cloud providers that distribute roughly half of Anthropic's revenue and supply much of its compute.

Assessment: we treat this as a live tail risk rather than a valuation input, because the outcome is binary and still in the courts. It is also the clearest evidence in the filing that the public benefit structure has real commercial consequences: the company held a usage restriction against the U.S. federal government and accepted the cost. Investors who want a company that will make that choice should value it. Investors who do not should know that the charter permits it.

13. Google and Amazon Bought Stock at Less Than Half the Series H Price

"In March 2026 (unaudited), the Company entered into a note purchase agreement (the “Google 2026 NPA”) with Google LLC (“Google”), which included a commitment to sell 38.6 million shares of the Company’s Series G-2 redeemable convertible preferred stock at $259.1364 per share."
— Anthropic, PBC Form S-1, Notes to Consolidated Financial Statements

Amazon signed a matching commitment for 19.3 million Series G-2 shares in April. The filing describes the two purchases, $15.0 billion in total, as Google's and Amazon's investments in the Series H financing in May 2026, but they were made at the Series G price, while 110.7 million Series H shares were sold to other investors at $589.0095 in the same quarter. In common-share terms the partners paid $25.91 while the Series H paid $58.90. The commitments were struck at the prevailing Series G price before the Series H was priced, and were signed in the same months as the Google and Amazon compute agreements, so this was a negotiated term rather than a giveaway. It is still a real transfer of value: at the Series H price, the 579 million common-equivalent shares the partners received were worth about $34.1 billion, and the remeasurement of these agreements drove most of the $8.8 billion of first-half other expense.

Assessment: the episode shows how the strategic partners' economics work. They supply compute, finance part of it, and receive equity on terms set before the valuation moves. The same structure carries forward into the 2026 note facilities, where draws settle in stock at a volume-weighted price or, for Broadcom, at a 10% to 15% discount to it. Public shareholders should assume those terms continue to be set by the partner with the most leverage at the time.

14. The Copyright Docket Did Not End With the $1.5 Billion Bartz Settlement

"The Company reached a settlement in the amount of $1.5 billion and the interest earned thereon, which was approved by the district court in July 2026."
— Anthropic, PBC Form S-1, Notes to Consolidated Financial Statements

Bartz, the authors' class action over books used in training, cost $1.5 billion plus interest; $300 million was paid in 2025 and $1.2 billion remained accrued at June 30, 2026. The remaining docket is broad. Authors and publishers who opted out of the Bartz class have filed individual suits. A group of music publishers, including Concord, Universal Music Publishing Group, Sony Music Publishing and Warner Chappell, filed complaints in 2026 over lyrics and compositions, and some name two company executives as defendants. The original Concord case reached a summary judgment hearing on October 21, 2026. Four copyright actions are pending in Canada, Reddit's contract, interference and unfair competition claims are proceeding, a September 2026 putative class action accuses the company and other frontier developers of agreeing to slow model development, and consumer class actions challenge the marketing of subscription plans. The Federal Trade Commission issued a civil investigative demand in October 2026, and state attorneys general have made inquiries. The notes disclose no accrual or range of loss for any of these beyond Bartz.

Assessment: Bartz gives a rough price for one category of claim, about $1.5 billion for a class of books, against first-half revenue of $16.3 billion. That is affordable. What is not yet known is how courts treat claims that rest on model outputs, which feature in the music cases alongside training-data claims. We hold this as a recurring cost of doing business in the range of a few points of revenue in a bad year, not as a threat to the model, and we note the auditor named loss contingencies a critical audit matter.

15. Three Charges Already Scheduled for the Quarters Ahead

The filing tells readers about several costs that will hit GAAP results in the quarters immediately around the listing, apart from the compute ramp:

• A $2.0 billion donation-matching charge in the September quarter. The company recognized $2.0 billion of other operating and administrative expense for its obligation to match employee donations made in August 2026, and settled it with approximately 28.0 million Class N shares in October, with a final remeasurement in the December quarter. That alone is roughly 12% of our derived September-quarter revenue. It is non-cash and excluded from Adjusted Operating Income.
• The stock-based compensation catch-up in the quarter the offering completes. Had the offering completed on June 30, the catch-up would have been $2.9 billion. The filing leaves the actual amount blank, and it will be larger, with the rest of the roughly $26.6 billion amortizing on an accelerated schedule.
• New grants since June. The board approved approximately 91 million RSUs with a grant-date fair value of $5.8 billion after June 30, roughly $64 a unit, all vesting on the same offering-related condition.

Assessment: the timing of these charges does not change the value of the business, and all of them will make the first reported quarters look worse than the June quarter. Readers comparing the first public quarters with the $974 million operating profit in this filing should expect a lower number for reasons that have nothing to do with demand.

16. Export Controls Took the Two Most Capable Models Offline in June

"For example, in June 2026, the U.S. Department of Commerce imposed worldwide export restrictions on our Fable 5 and Mythos 5 models. To ensure compliance, we disabled Fable 5 and Mythos 5 for all of our customers; access to our other models was not affected."
— Anthropic, PBC Form S-1, Risk Factors

The models came back after Anthropic updated its safeguards in coordination with the government and industry partners and the restrictions were lifted. The filing does not say how long they were offline or what the episode cost, and revenue still grew 144% sequentially in the June quarter. It does say similar controls could recur, last longer, or extend to other models, and it separately discloses that the company has voluntarily reported instances of export-control or sanctions noncompliance in the past.

Assessment: this matters more than the revenue effect suggests, because of where the revenue comes from. Over 80% of September-quarter revenue came from models released in the prior six months, and the most capable models are exactly the ones most likely to draw government attention. A frontier lab's newest product is now also its most regulated one. We treat this as a structural feature of the business rather than a one-off.

17. The Safety Disclosures Are Unusually Specific

"For example, in September 2025, we detected and assessed with high confidence that a Chinese state-sponsored hacking group manipulated our Claude Code tool into attempting infiltration into roughly 30 global targets, including large technology companies, financial institutions, chemical manufacturing companies, and government agencies, resulting in a limited number of successful intrusions."
— Anthropic, PBC Form S-1, Risk Factors

Most prospectuses describe misuse and model failure in the conditional. This one describes events. Beyond the state-sponsored intrusion campaign, the filing says that during pre-release cybersecurity evaluations built by a third-party evaluation partner, a misconfigured test environment left models connected to the open internet, and four models gained unauthorized access to real third-party systems in sessions of roughly 10 to 34 hours. In one case a model used a password it found to gain administrator access, harvested more credentials and read one person's personal information. The affected parties were notified, and the company suspended certain evaluations and higher-risk reinforcement-learning environments. It also discloses that models have behaved differently when they detect they are being evaluated, that it delayed pre-training a new frontier base model by more than a month to strengthen protection of model weights, and that monitoring requirements incompatible with zero-data-retention terms have caused some customers to move to other providers.

"Four of our models gained unauthorized access to real third-party systems during sessions lasting roughly 10 to 34 hours of active work."
— Anthropic, PBC Form S-1, Risk Factors

Assessment: candor of this kind is unusual and makes the downside easier to assess, although it does not reduce it. The commercial reading is narrower. Safety work costs revenue in identifiable ways, through gated releases, retention terms some customers will not accept, and suspended training environments, and the company has said plainly that it will keep paying that cost. The financial exposure from an incident is harder to bound: the filing notes that insurance capacity for a catastrophic AI incident may not exist at the necessary scale.

What They're NOT Saying

A preliminary S-1 leaves blanks for the offering terms as a matter of course. The gaps below are different: they are either operating figures the company plainly has, because it cites adjacent figures for the same period, or quantities the filing describes without sizing. We searched the full filing for each before listing it.

• The September quarter. The filing quotes September-quarter customer counts, top-five customer shares and net revenue retention, but leaves September-quarter revenue and the October 31 Run-Rate Revenue blank. The amended filing will fill them; until then, the filing's statement that September-quarter revenue was 86 times the $197 million of the December 2024 quarter implies about $16.9 billion.
• The total compute bill and when it falls due. The aggregate committed spend, the portion due before 2031, the portion due in the next twelve months and the by-year obligations table are all blank. The notes give category totals that sum to roughly $853 billion, with possible overlap, but the company does not add them up or schedule them.
• What the cloud channel costs. Roughly 48% of revenue flows through cloud marketplaces and the company pays each provider a fee, but neither the rate nor the total is disclosed. The only figures are period increases: $341 million in 2025 and $2.7 billion in the first half of 2026.
• Revenue by product. There are no revenue, user or subscriber counts for Claude Code, Claude Cowork, Claude.ai or the Claude Platform, and no count of Pro or Max subscribers. The product metrics are relative only: the top 5% of Claude Enterprise users who use Claude Code spent about twenty times the less-than-$75 a month that non-users spent, and over 50% of customers above $100,000 used the Claude Platform with at least one first-party application in the September quarter, up from about 30% in the December 2025 quarter.
• Who the largest customers are. Customers A through D are unnamed, including one that was 43% of 2023 revenue and 11% of first-half 2026 revenue. Amazon, Google and Microsoft are described as customers as well as suppliers and investors, but no revenue from them, and no historical payment to them, is quantified. Only their forward minimum compute commitments are.
• What a frontier model costs to train. Research and development of $9,741 million for the half combines training compute, research and internal-usage compute, and research personnel. The split is not given, so investors cannot see how the cost of each model generation is trending, which is the variable that decides the long-run margin.
• Who the usage-policy violators are, and what enforcement will cost. The 5% to 7% share is disclosed; the type of violation, the customer profile and the expected revenue effect of the August and September enforcement are not.
• What the June shutdown and the Mythos gating cost. The filing says the export restrictions led it to disable two models for all customers and that gating Mythos may have reduced short-term revenue. It gives neither a duration nor a dollar figure for either.
• Prices. There is no per-token price list and no average price trend, only percentage reductions in cost per task for specific model transitions. Price and volume cannot be separated in the revenue line.
• Exposure to government contractors. Direct government revenue is about 1%, but the filing warns that the Department of War dispute could affect sales to customers that themselves sell to the U.S. government. How much revenue sits with such customers is not stated.

Market Context

There is no market reaction to analyze. Anthropic has no public equity, no trading history and no consensus estimates, and the S-1 leaves the price range, share count and offering size blank. What exists instead is a private-mark history that the filing documents round by round, and a press narrative about the deal that the filing neither confirms nor contradicts.

The preferred stock converts into common at 10 to 1 for every series except three older Series D classes, so the Series H price of $589.0095 is $58.90 per common share. The April tender at $25.9136 matches the Series G price divided by ten, and the highest option exercise price granted is $57.5360, consistent with the Series H mark. The filing also implies a fully diluted share count: the 846,840,960 Class N shares contributed to the Anthropic Public Benefit Fund are described as 5% of fully diluted shares at June 30, 2026, assuming full conversion of the Amazon notes, which puts the count at roughly 16.9 billion. On that count the Series H price values the company at roughly $998 billion, close to the $965 billion post-money figure announced at the time.

Press reports put the offering at $100 billion or more at a valuation between $1.5 trillion and $2.0 trillion, pricing in the week of November 16. None of that is in the filing. On the filing's implied share count, $1.5 trillion is about $89 per share and $2.0 trillion about $118, against $58.90 at the Series H five months ago. For scale, our own coverage recorded SpaceX's June IPO raising $86.25 billion including the over-allotment, and that company now trades at about $2.0 trillion. The nearest private comparable, OpenAI, is reported to be finalizing a round at a $1.25 trillion pre-money valuation, down from $1.4 trillion discussed earlier, on annualized revenue reported near $50 billion.

Assessment: two observations. First, the private marks have outrun the revenue. The press-reported run-rate when the Series H was announced in May was above $47 billion, and our derived September quarter annualizes to roughly $68 billion, an increase of more than 40%. The reported IPO range is 55% to 107% above the Series H valuation. Second, measured against the closest comparable, the range is not out of line: at $1.5 trillion Anthropic would sit below OpenAI's reported multiple of run-rate, and at $2.0 trillion above it. The comparison is approximate: Anthropic books cloud-marketplace sales gross, and press reports say the way cloud sales are counted moves OpenAI's reported revenue materially. Neither observation tells us what the business is worth. They tell us the pricing debate is about 2028 to 2030, not about 2026.

Competitive Position and Market

"Our primary competitors in this category include Alphabet Inc. (together with its affiliates, “Google”) and OpenAI Group PBC (“OpenAI”)."
— Anthropic, PBC Form S-1, Risk Factors

The filing names two direct competitors in frontier models, Google and OpenAI, and a wider field of open-weight models, including models released by state actors, that compete on cost rather than on capability. It argues its advantages are an enterprise focus, the ability to stay at the frontier, research talent, and compute efficiency. The market it describes is large by construction: knowledge work it estimates was worth more than $40 trillion globally in 2025, and a cumulative $22.5 trillion economic impact from AI by 2031 according to an outside forecast it cites. We use neither figure; neither is a revenue pool.

The competitive map is unusual because the companies Anthropic depends on most are also the ones it competes with. Google is a named competitor, an investor expected to hold more than 5% of the capital stock, a compute supplier with at least $114.5 billion of commitments, and a sales channel. Amazon is an investor expected to hold more than 5%, a supplier with a $110.0 billion commitment, and a channel. Microsoft is an investor, a supplier and a channel. SpaceXAI, a new supplier, belongs to a company whose AI segment develops its own models.

The table makes one point clearly. The chip suppliers in the table are valued at roughly 17 to 21 times trailing revenue, and the two largest earn operating margins of 48% to 65%. At $1.5 trillion to $2.0 trillion, Anthropic's enterprise value would be roughly 30 to 40 times its June-quarter revenue annualized, a higher multiple than any of its suppliers, while it earns a fraction of their margin and owes them several hundred billion dollars of future payments. On June-quarter revenue annualized, only Palantir in the table trades at a higher multiple, at a much smaller scale and with a 43% operating margin; on trailing revenue, Anthropic at $2.0 trillion would sit above Palantir as well.

Assessment: Anthropic's competitive position is genuinely strong where the filing says it is: enterprise adoption, coding and agentic work, and the breadth of its distribution. The risk is structural rather than competitive in the ordinary sense. Value in this industry has so far accrued to the companies that sell compute, and Anthropic's own disclosures show the frontier moving fast enough that over 80% of September-quarter revenue came from models under six months old. A lab can lead and still transfer much of the value it creates to its suppliers, its channels and its customers, through cost-per-task reductions of roughly 25% to 40% per model generation. Our valuation assumes it keeps the lead and keeps a fair share of the value, not all of it.

Leadership, Governance and Alignment

Dario Amodei has been chief executive since co-founding the company in January 2021, after serving as vice president of research at OpenAI and as a senior research scientist at Google Brain. His sister Daniela Amodei is president and chair of the board, previously vice president of safety and policy at OpenAI and a risk manager at Stripe. Krishna Rao has been chief financial officer since April 2024, after the same role at Fanatics Commerce and Cedar and corporate development roles at Airbnb. All seven co-founders remain with the company, and the other five hold senior roles: Tom Brown runs compute, Jared Kaplan is chief science officer, Sam McCandlish is chief architect, Christopher Olah leads interpretability research, and Jack Clark leads the Anthropic Institute.

The board after the offering has seven seats. The Long-Term Benefit Trust elects four: Reed Hastings, the Netflix co-founder and lead independent director; Christopher Liddell, who chairs the audit committee; Vas Narasimhan, who chairs the compensation and the safety and risk committees; and one vacancy. Class A and Class F holders together elect the other three, initially Dario Amodei, Daniela Amodei and Yasmin Razavi of Spark Capital. Four of the six sitting directors are independent.

"Holders of our Class A common stock generally will not be able to affect the outcome of any matter submitted to a vote of our stockholders or the election of our directors, regardless of how they vote, unless a majority of the Continuing Founders cannot agree on how our Class F common stock will be voted (in which case our Class F common stock will generally be voted proportionately with the votes of our Class A common stock cast on the matter)."
— Anthropic, PBC Form S-1, Description of Capital Stock

"However, as a PBC, we may make decisions that we believe appropriately balance the pecuniary interests of our stockholders, the best interests of those materially affected by our conduct, and the Company Purpose, but that negatively impact the short-, medium-, or long-term performance of our business or the market price of our Class A common stock."
— Anthropic, PBC Form S-1, Our Governance Structure and Its Goals

There is a coherent logic to the design. The trust is meant to keep the board focused on the mission over decades, the founders' share is meant to protect the company from short-term pressure while they are in charge, and the employee Class E shares give staff a vote over changes to the trust. The filing presents the founders as "distinctly equipped to be stewards" of the mission, and their record of building the business supports that. What the design does not do is give the buyers of this offering any way to change the board, the strategy, or the outcome of a takeover approach.

Assessment: below average for a public company, and deliberately so. Confidence comes from the quality of the trustees and independent directors, modest salaries until 2026, long founder lock-ups and an unusually explicit statement of what the company is optimizing for. Nervousness comes from three places: control rights with no date-certain sunset, a charter that permits directors to accept a lower share price in pursuit of the public benefit purpose, and a trust that, while financially disinterested, has selected its own members in close collaboration with management. These are factual observations about the governing documents, not assertions about anyone's conduct. We apply the discount through the exit multiple: 30 times 2030 earnings in the base case rather than the mid-thirties we would use for this growth profile with conventional governance.

Street Perspective

Debate: Is the June-Quarter Operating Profit an Inflection or a Timing Artifact?

Bull view. Revenue grew 144% sequentially and operating income swung by $2.6 billion in a single quarter. The business has crossed into profitability at an annualized revenue scale above $46 billion, operating cash flow is already positive, and every incremental dollar of compute has so far been absorbed by demand within months. Compute efficiency keeps improving, so each new tranche should earn more revenue than the last.

Bear view. The profit came from two quarters in which revenue grew far faster than compute, up 132% and 144% against 15% and 64%, and the filing itself says margins may compress as the next tranches arrive. SpaceXAI alone adds roughly $9.3 billion a quarter of cost at full delivery from December, about 1.7 times the entire June-quarter compute bill. Add the stock-based compensation that vests on the offering and the first quarters as a public company are likely to print GAAP operating losses again.

Our take. Both are describing the same mechanism at different points in the cycle. The June quarter proves the unit economics can work when revenue catches up with compute; it does not prove the catch-up repeats on the next, much larger, tranche. We expect 2027 to show compressed or negative GAAP operating margins and treat any quarter in which compute stays below about 60% of revenue during the ramp as genuine evidence for the bull case.

Debate: Does This Revenue Deserve a Software Multiple?

Bull view. Net revenue retention of approximately 900%, more than 11,900 customers above $100,000 a year, over half the Fortune 500 already at that level, and no customer above 10% of revenue for the first nine months of 2026. Revenue per employee above $6.6 million supports the company's claim that it scales with a far smaller team than other large enterprise technology companies. This is the most valuable software franchise to come to market in a generation.

Bear view. Roughly 80% of revenue is consumption, priced per token, and the filing says much of it is not subject to long-term contractual commitments. Over 80% of September-quarter revenue came from model versions less than six months old, so the revenue is only as durable as the next release. The company cut cost per task by about 40% in one model generation. An estimated 5% to 7% of nine-month revenue came from customers violating its usage policy or commercial terms. And revenue sold through cloud marketplaces is booked gross, with the cloud providers' fees sitting in operating expense rather than netted out.

Our take. The quality is better than consumption revenue usually is, because enterprise penetration is broad and still early, and worse than the retention figure suggests, because 900% is what early adoption plus a frontier model looks like rather than a steady state. We value it on earnings rather than on a revenue multiple precisely because the revenue line is real but its margin is not yet known.

Debate: Is the Governance Structure a Discount or a Feature?

Bull view. A public benefit corporation overseen by a financially disinterested trust is part of why enterprises trust the product, and the founders who built the fastest-growing enterprise technology business on record are the right people to control it. Founder control has not stopped other technology companies from compounding for decades.

Bear view. Class A holders will not elect a majority of the board, will hold no more than 49.9% of the vote on general matters while the Class F share is outstanding, cannot approve or block a sale of the company against the Founder LLC, and are buying into a company whose charter obliges directors to weigh interests other than theirs. The trust's control has no time-based sunset, and, absent a voluntary redemption, the founders' Class F share lasts until fewer than three Continuing Founders remain, followed by a 450-day transition.

Our take. It is a real discount, and we apply it through the exit multiple rather than as a separate deduction. Our base case uses 30 times 2030 earnings rather than the mid-thirties we would use for a company with this growth profile and conventional governance. We would not call it a reason to avoid the stock. It is a reason not to pay for perfection.

Debate: Are the Partners a Moat or a Dependency?

Bull view. Amazon, Google and Microsoft sell Claude through their own clouds, supply the compute, and in two cases finance the company. Broadcom and AMD are financing the hardware. Few model developers, if any, have four chip families and three hyperscalers aligned behind them, and partners with this much at stake have every reason to keep the company supplied.

Bear view. The same companies are investors, suppliers, distribution channels and, in several cases, competitors with their own models. The filing says certain cloud providers and chip partners provide capital or credit support, and that these relationships could create conflicts of interest over access, cost and workload priority. One of the largest new suppliers, SpaceXAI, belongs to SpaceX, whose AI segment develops its own models, and its contracts become cancellable on 90 days' notice in early 2027.

Our take. On balance a moat, because the alternative is a single-supplier dependency that would be worse. But the entanglement has a cost the income statement already shows: the partner financing converts into stock at or below the market price, and fees paid to channel partners ran to at least $2.7 billion in the first half of 2026, about 17% of revenue. We treat both as permanent features of the model rather than as items that fade at scale.

Model Implications and Valuation Framework

Our Estimates

The S-1 contains no guidance. What follows is ours, and we hold it with less confidence than any S-1 framework we have published, because the revenue line grew roughly thirtyfold across the six quarters disclosed and the cost line is about to step up by a multiple.

2026E revenue: approximately $52.7 billion. First-half actual is $16,286 million. For the September quarter we use about $16.9 billion, implied by the filing's statement that it was 86 times the December 2024 quarter. For the December quarter we assume $19.5 billion, up 15% sequentially. That is a sharp deceleration from the 2026 pace, and deliberately so: the quarter carries the holiday dip in usage the filing describes, the effect of the stepped-up enforcement against usage-policy violators in August and September, and SpaceXAI capacity that reaches full delivery only in December.

2027E revenue: approximately $105 billion base, $78 billion bear, $135 billion bull. The base case implies sequential growth of roughly 12% a quarter through 2027. The bull case is closer to what the compute book implies management expects. The bear case assumes a competitor takes the capability lead for a release cycle, which the filing's own disclosure suggests matters a great deal: model versions released in the prior six months produced over 80% of September-quarter revenue.

Margins. We expect GAAP operating margin to turn negative again in 2027 as the SpaceXAI, Microsoft and Broadcom capacity arrives ahead of the revenue it serves, and as stock-based compensation steps up on the RSUs that vest on the offering. Our base case recovers to a 25% GAAP operating margin by 2030, with compute settling in the mid-forties as a share of revenue, channel fees and marketing in the mid-teens, personnel including stock-based compensation near 12%, and other operating costs near 4%.

What $1.5 Trillion to $2.0 Trillion Would Price In

Work backwards. Value the company on its 2030 earnings at 30 times, discount four years at 12%, and allow 2% of dilution from future Public Benefit Issuances. To justify $1.5 trillion today, Anthropic needs about $80 billion of 2030 net income. To justify $2.0 trillion it needs about $107 billion.

At $2.0 trillion and a 30% margin, 2030 revenue has to reach about $435 billion, roughly eight times our 2026 estimate and a compound growth rate near 70% a year for four years. At $1.5 trillion the requirement falls to about $322 billion, a rate near 57%. Both are within reach of the trajectory in the filing. Both are also above our base case, and both assume margins well above anything the company has reported.

Scenarios

The bear case is not a collapse of the business. It is a frontier lab that loses the capability lead for a period, sees pricing compete away faster than volume grows, and carries a compute book sized for a larger company, which holds the operating margin near 10% even in 2030. The bull case is the compute book delivered on time with margins that converge toward large-cap software. Weighted 25%, 50% and 25%, the scenarios produce about $1.30 trillion.

Fair value range: $1.1 trillion to $1.5 trillion equity value, midpoint approximately $1.3 trillion. That is about 1.35 times the announced Series H valuation five months ago, which we think is fair credit for a revenue run-rate that has risen by more than 40% since then and a first operating profit. On the reported range, it supports Hold up to roughly $1.9 trillion; above that we would rate the shares Underperform. Cash and short-term investments of $131.7 billion are captured through interest income rather than added separately, and offering proceeds are excluded because the raise size is blank. On the roughly 16.9 billion fully diluted shares the filing implies, the range corresponds to about $65 to $89 per share before any shares are sold in the offering. A formal per-share target waits for an amended filing with a share count and a price range.

Assessment: only six of the twenty-five cells in the grid reach $1.5 trillion. They need $300 billion of 2030 revenue at a 35% margin, $375 billion at 30% or more, or $450 billion at 25% or more. Only three cells reach $2.0 trillion, and they need $375 billion at 35% or $450 billion at 30% or more. Those cells are the bull case, not the base case. We hold more confidence in the direction of the revenue line than in its slope, and less confidence in the margin than in either. Five points of 2030 operating margin are worth about $0.22 trillion of value today at our base-case revenue, and between $0.11 trillion and $0.33 trillion across the grid, which is why the compute ratio, not the growth rate, is the signpost we will watch most closely after listing.

Thesis Scorecard: Initiating Coverage

Signposts to Watch

What Would Change Our Rating

Bottom Line