Anthropic’s S-1 Shows a Large Serving Margin and Hundreds of Billions in Future Compute Costs

Anthropic sells access to its Claude models in three ways. Through an API (an application programming interface that lets software send text to the model and receive its answer), customers pay by the token, the small unit of text the model reads and writes. Through subscriptions, they pay by the seat, a fee for each user. And through agents such as Claude Code, the model does multi-step work on a user’s behalf. Revenue was $16.3 billion in the six months to 30 June 2026, against $4.6 billion in all of 2025, according to the S-1’s Prospectus Summary¹. An S-1, formally a registration statement on Form S-1, is the document a company files with the Securities and Exchange Commission (SEC) before selling shares to the public. Until this one, every Anthropic number in circulation came from a leak or a company press release.

The financial statements show that, across all paying customers together, revenue is higher than the costs allocated to serving them. They do not show whether every product or every customer is profitable. The annual statements for 2023 to 2025 are audited. The 2026 interim figures, which cover only part of the year, are unaudited. Several Claude list prices have fallen. Almost half of revenue comes through cloud marketplaces (the AI services through which Amazon, Google and Microsoft sell models to their cloud customers), which also sell rival models. Anthropic orders compute 12 to 18 months ahead, so the cost can arrive before the revenue it is meant to support. The filing warns that margins may shrink when that happens.

There is evidence that customers have adopted Claude Code and other agents and that the agents add value to their work. The costs and returns attributable to them are undisclosed, so their economic rating is provisionally neutral (this report’s grade where the evidence on returns is not yet enough to call a line advantaged or exposed), alongside the API and subscriptions. Revenue that comes through cloud partners is exposed to two pressures: the partners can route customers to other models, and they can bargain over terms. Buying at the valuation set in the initial public offering (IPO) means paying for future models that keep revenue growing faster than the compute and the dilution needed to produce them. Dilution is the shrinking of each existing owner’s stake as new shares are issued.

Six claims are asserted only. Most concern speed and culture: “the fastest enterprise adoption of any technology in history,” “approximately 1% regrettable attrition” (the loss of staff the company wanted to keep), more than 90% of the company’s own code written by Claude, and Claude Cowork outpacing Claude Code’s early adoption. None comes with a definition or a comparison that a reader could check. The asserted claim that matters most is that Anthropic cut customers’ cost per task by about 40% with Opus 5.5 in September 2026 “while preserving our margins.” Opus 5.5 was released after the last reported quarter, so no reported margin can test it.

On the broad reading used here, one claim is contradicted: that the cloud arrangement avoids channel conflict, meaning that a partner who sells Claude has no reason to push competing products. Amazon’s second-quarter release describes Bedrock, Amazon’s service for offering AI models to its cloud customers, adding rival models alongside Claude. Amazon’s 10-Q, its quarterly filing, records about $50 billion invested in OpenAI in 2026. That gives Amazon competing incentives. It does not prove that Amazon treats Claude worse than other models.

The premium-value claim is one of the twelve that are partly supported. Posted token prices are equal, which leaves open whether Claude completes valuable tasks better or completes them with fewer paid tokens. The filing gives no comparison of what customers actually pay per task.

Anthropic’s product is a family of large language models, programs trained on very large amounts of text and code that take written instructions and produce text, code or actions. It sells them in three layers, which the filing calls “Models, Platform, and Applications.” The Business section² describes four model classes, each a trade-off between capability, cost and speed. Mythos is the most capable, and it is sold in two forms. Fable is generally available and runs with safeguards that block or downgrade some sensitive requests. Mythos itself goes only to vetted cybersecurity and life-science organizations. Opus 5.5, released in September 2026, is described as “the enterprise daily driver.” Sonnet is the mid-priced model for everyday and high-volume work, and Haiku is the cheapest and fastest.

The platform sits on top of the models. It is the set of tools a developer uses to turn a model into a working agent, meaning software that carries out a task step by step on its own. It includes tool use (the model calling outside software), computer use (the model operating a desktop by clicking and typing), memory, long context windowsThe amount of text a model can hold in view at once while it works, counted in tokens: the instructions, the documents supplied and the conversation so far. of up to a million tokens (a token is the unit in which models read and write text), code execution and web search. There are also two ways to run an agent: on the customer’s own computers through the Agent SDK (software development kit), or on Anthropic’s through Claude Managed Agents. Models reach company data and tools through the Model Context Protocol (MCP), a connection standard Anthropic introduced in November 2024. The filing notes that Anthropic donated MCP to the Linux FoundationA non-profit that looks after widely used open-source software, such as the Linux operating system, so that no single company controls it. in December 2025, so any model can use it. That helps adoption. It also means that the work of connecting a model to a company’s systems carries over from one vendor to another.

The applications are the products Anthropic sells directly to people. Claude.ai is the chat product. Claude Code is a coding agent that plans, writes, tests and commits code (records the changes in the project’s code history). Claude Cowork applies the same design to general office work, and since September 2026 it has been merging into Claude.ai. Smaller products cover security scanning, design, scientific research and work inside Slack. Seven industry packages bundle connectors and compliance controls for finance, financial advice, healthcare, life sciences, law, education and government.

Pricing follows the layers. Customers who use the API, the interface through which developers’ own software calls the models, pay per token, which the filing calls “pay-for-intelligence.” The amount of text in a token varies by model and language. Individual and Team subscriptions include a usage allowance. The current Enterprise plan charges for access and bills usage separately. Enterprises can also sign annual or multi-year contracts with a minimum spend in exchange for a committed price list, and the revenue from these contracts is still usage-based. Prices are quoted separately for input, the text the model reads, and output, the text it writes. Anthropic’s pricing page³, checked 27 October 2026, lists $10 per million input tokens and $50 per million output tokens for Fable 5.1, $4 and $20 for Opus 5.5, $2 and $10 for Sonnet 5.5, and $0.10 and $0.50 for Haiku 5.5 on prompts up to 100,000 tokens.

The current plan list, checked 27 October 2026, prices Pro at $20 monthly or $200 annually and Max from $100 monthly. Team standard seats cost $20 per month on annual billing or $25 monthly; premium seats cost $100 or $125 respectively. Enterprise lists $20 per seat per month, billed annually, plus usage at API rates. The Enterprise plan description dated 1 September 2026 confirms that its seat fee includes no usage allowance. A Code or Cowork user can therefore bring in both subscription revenue and usage-based revenue. Subscription and usage are two ways of billing for the same products.

One line on the pricing page concerns Sonnet 5. It launched at $2 and $10 as an “introductory” price, with an increase to $3 and $15 scheduled for 1 September 2026. The page now says that increase “will not occur.” A company with room to raise prices would have gone ahead with the increase on its mid-tier model. The filing presents each price cut as a choice made possible by efficiency, and in part it is. Cutting prices is also what a seller does when buyers can switch to another model inside the same cloud console.

Customers buy Claude either directly from Anthropic, through the API or the apps, or through the AI services of the three large cloud providers: Amazon Bedrock, Google’s Gemini Enterprise Agent Platform and Microsoft Foundry. Claude is also built into other companies’ software, including Salesforce’s Agentforce and Microsoft’s Copilot family. It is also sold through systems integrators (consulting firms that build and install technology for clients) such as Accenture, Deloitte, KPMG and PwC. According to Note 16 of the financial statements⁴, Anthropic agreed in May 2026 to invest $325 million in a new services company, launched with private-equity partners as Ode with Anthropic, to sell Claude-based solutions to mid-sized companies.

The price of each unit Anthropic sells falls with each generation. The filing also reports that customers move to the newest model quickly. Model versions released in the six months to 30 September 2026 earned more than 80% of third-quarter revenue, according to MD&A⁵, the management’s discussion and analysis section of the filing. Revenue per unit keeps falling, and the models that earn it are replaced about twice a year.

Four-fifths of Anthropic’s revenue is consumption revenue, meaning fees charged according to how much customers use the models. In the six months to 30 June 2026, consumption-based revenue was $13.1 billion and subscription revenue $3.2 billion, an 80/20 split, according to Note 2 of the financial statements⁴. The split was 75/25 in 2024 and 83/17 in 2025, so subscriptions have held roughly a fifth of the business while revenue went from $386 million for all of 2024 to $16.3 billion for a single half-year. Anthropic books consumption revenue as customers use tokens (the units of text the models read and write). It spreads subscription revenue evenly over the subscription term, which is generally a year or less. Seventy-two percent of H1 2026 revenue came from customers billed in the United States.

The filing does not break revenue down by product. It does not say how much comes from Claude Code, how much from Cowork, or how much from the API, which other software companies use to build their own products. The only product figure on record is from the company’s Series G announcement⁶ in February 2026, which put Claude Code’s run-rate revenue (an annualized figure defined at the end of this section) above $2.5 billion. Instead of a product breakdown, the filing gives one indirect measure. In the third quarter of 2026, Claude Enterprise users who did not use Claude Code spent less than $75 a month on average, while the top 5% of users who did spent about 20 times that, according to the Business section².

Almost half of revenue now comes through the cloud providers: 48% in the first half of 2026, up from 11% in 2023. When a customer buys Claude through Amazon Bedrock or Google’s or Microsoft’s equivalent, the cloud provider bills the customer, collects the money and passes it to Anthropic after deducting its fee. Anthropic keeps the contract with the customer and sets the price, so it books the full amount as revenue and records the fee as an expense.

The filing reports net revenue retention (NRR) of about 900% at 30 September 2026. NRR measures what existing customers spend now against what they spent a year earlier. This version compares one quarter with the same quarter a year earlier, so a few large customers can move it. It counts what customers spend, whatever the price: a customer using twenty times more tokens at half the price counts as tenfold expansion. The filing does not disclose the base the ratio is measured against, which is eligible Q3 2025 revenue excluding Pro and Max subscribers. So the 900% figure cannot be used to work out Q3 2026 revenue, and third-quarter revenue is left blank in this filing.

The filing estimates that 5% to 7% of revenue in the nine months to September 2026 came from customers later found to be breaking Anthropic’s usage policy or commercial terms. Anthropic booked that revenue before it cut off their access. The filing does not give nine-month revenue, so the dollar amount cannot be worked out. Nor does it say who these customers were. Anthropic says enforcement tightened in August and September 2026 and expects the share to fall.

Usage is also seasonal. Total customer spend fell about 20% in the two weeks to 2 January 2026 compared with the two weeks before, because companies froze their code (stopped changing their software) over the holidays. At current growth rates, the dip does not show in the quarter’s revenue. Once growth slows, it will.

The company’s preferred growth metric is “Run-Rate Revenue,” defined as consumption fees in the trailing 28 days multiplied by 13, plus monthly subscription fees multiplied by 12. In effect, it takes recent consumption and subscription fees and scales them up to a full year. Anthropic’s own announcements put it at more than $5 billion in August 2025, $14 billion in February 2026, more than $30 billion in April and more than $47 billion in May 2026 (Series F⁷, Series G⁶, Amazon compute⁸ and Series H⁹ announcements). The S-1 leaves the 31 October 2026 figure blank. For a business growing this fast, annualizing a 28-day window gives a fair snapshot. For a business about to slow, it gives a generous one, because it treats the latest and best month as if it would repeat for a year.

In the first quarter of 2026, Anthropic’s five largest customers paid about 36% of its revenue. By the third quarter that share had fallen to about 21%, and over the nine months it was about 25%, according to MD&A⁵ and the Risk Factors¹⁰. One customer, labelled “Customer C” in Note 1 of the financial statements⁴, provided 11% of revenue in the first half of 2026 and 10% of receivables (money customers owed Anthropic) at 30 June. In 2023 the same label covered 43% of revenue. Two other customers each provided 12% of revenue in 2025. The filing does not name any of them. The Risk Factors add that many top customers “are not subject to long-term contracts” and “are able to reduce or in some cases cease spending at any time.” They also say some large customers have negotiated “flexibility around financial commitments, termination rights.”

The filing’s headline statement, that “no single customer represented more than 10% of our revenue for the nine months ended September 30, 2026”, is accurate, and the fall in concentration is a good sign. But those nine months include the first six, when one customer was above 10%. A business where five buyers paid more than a third of a recent quarter’s revenue, without long contracts, is less diversified than its 11,900 large customers suggest.

The filing does not say what kind of companies the top five are. One event outside the filing is relevant to that question. On 14 August 2026, SpaceX, which owns the Grok models through its xAI subsidiary, completed its purchase of Anysphere, maker of the Cursor coding editor, at an implied equity value of $60 billion, according to SpaceX’s third-quarter 10-Q¹¹. Before the merger SpaceX had a compute agreement with Cursor to “collaborate to improve existing models, including Grok.” A leading AI coding application now belongs to a rival model maker, one that describes its strategy as vertical integration of “compute infrastructure, SI models, and applications.” The SpaceX filing does not say whether Cursor used Claude, and this report does not assume it.

More than half of the Fortune 500 and more than 30% of the Global 2000 paid Anthropic $100,000 or more over the trailing year, and every company in the Fortune 10 did. Each of these companies spends very little with Anthropic relative to its size. For those Fortune 500 customers, Anthropic’s revenue was on average less than 0.1% of their total expenses in the third quarter of 2026, and less than 1% even for the ten largest, according to MD&A. The filing presents this as room to grow. It also means Claude is still a small line in most large buyers’ budgets, small enough to move to another vendor without a board decision.

The case studies in the Business section² are printed as full-page images and are not searchable text. Setting the praise aside, they describe three kinds of work. The first is software engineering: Verizon has more than 33,000 engineers using Claude Code, Salesforce made it the primary coding agent across its engineering organization, and Shopify gives it to more than 7,500 employees. The second is internal knowledge work: Eli Lilly made Claude Enterprise available to more than 50,000 employees, PayPal reports more than 12,000 monthly users, and SAP says its consultants save 25 hours a month. The third is building Claude into a customer’s own product: Walmart’s Sparky shopping agent, Shopify’s Sidekick for merchants, Salesforce’s Slackbot and Agentforce, and SAP’s Joule.

The same pages show how replaceable Claude is. Walmart says Claude “is one of the models that powers” its developer agent. Shopify calls Claude “one of the models the company uses to power Sidekick.” SAP’s Joule “routes to the most suitable model — such as Claude.” PayPal credits “Claude and other tools.” Lilly began “using Claude in 2024 through Amazon Bedrock,” the same Amazon service that now sells OpenAI’s and SpaceX’s models. Salesforce is the exception: it named Claude a preferred model for Agentforce 360 and the first model “fully contained inside its trust boundaryA security term for the line around the systems a company trusts with sensitive data. Data inside it stays under the company’s own controls; data that crosses it goes to systems the company does not control..” Two of the featured customers, SAP and Salesforce, also note that each owns less than 1% of Anthropic’s stock.

The customer accounts support the case that Claude is useful for coding and for long agent tasks. They also show buyers keeping control of their own systems. Anthropic’s Economic Index of 24 March 2026¹⁵ found that computer and mathematical occupations accounted for 35% of sampled Claude.ai conversations in February. That figure measures use. It says nothing about how Claude’s quality compares with rival models or about revenue, and neither source gives coding’s share of revenue. Routing (sending each request to whichever model suits it, as SAP’s Joule does) lets customers make the choice of model again at every new release.

Shopify has described how this substitution works in practice. Its 21 October 2026 engineering account says Sidekick keeps a frontier model, meaning one of the most capable general models, for talking to merchants and for planning. It hands subtasks that process a lot of text to open modelsModels whose weights, the numbers learned in training that make up the model, are published. Anyone can download, adapt and run them on their own computers instead of paying a provider by the token. They are also called open-weight models. that Shopify trains and serves itself. Its specialist model for GraphQL, a language for querying data, costs about 90% less to serve than its frontier-model version on that task. An earlier account, 5 August 2026, describes Shopify’s own evaluations and a learning loop. This setup can keep the most valuable requests on a frontier model while moving the routine work elsewhere. Neither post reports a fall in Claude revenue. Jane Street’s 26 May 2026 account gives a similar reason for building its own AIDE interface: it wants freedom to choose model vendors. When customers build Claude into their workflows, Anthropic has a chance to keep winning those requests. But the interface the customer works in need not belong to Anthropic.

Other announcements from the customers themselves confirm deployments but say less about what those deployments cost or save. SAP, 12 May 2026, describes Claude as a primary reasoning option on a platform that supports any model. Salesforce’s Claudeforce page describes packaged skills and integrations. Walmart, 21 October, describes letting several different agents shop for its customers. Verizon, 15 May, confirms a security collaboration. PayPal, 13 May, confirms training and small-business tools. None of these sources independently confirms the employee counts, spending or savings given in the filing. The Salesforce page was accessed on 27 October; the others carry the dates stated.

The other case studies in the filing have the same limit. DoorDash’s 16 September 2026 account of Vera, its data agent, puts much of the value down to its data and the design of the agent. Rakuten, 24 November 2025, describes a mix of external and internal models. Travelers, 15 January 2026, confirms an engineering deployment. AIG’s 2025 annual report names several AI partners. UST’s technical account explains integration work, and Mozilla, 21 April 2026, reports concrete security findings. AIG and UST were accessed on 27 October. Searches did not independently confirm the Lilly deployment count in the filing or the IRC and Mercado Libre case figures, so those remain examples supplied by Anthropic. None of these cases shows what return Anthropic earns on the cost of training the models that serve the customer.

Anthropic can win the repeated choice of model, and it has been winning it. In the third quarter of 2026 more than half of customers spending $100,000 or more used both the Claude Platform and at least one first-party application (one Anthropic makes itself, such as Claude Code or Cowork), up from about 30% in the fourth quarter of 2025. A customer whose engineers use Claude Code, whose office staff use Cowork and whose product runs on the API is harder to move than one that uses a single model through Bedrock. The case studies show that pattern taking hold at Salesforce, Lilly and Verizon. At Walmart, Shopify and SAP they describe Claude as one model among several inside the customer’s own products.

A Claude model is made in two main stages. The filing describes both in the part of its Business section² titled “Research Lab Identity and the Research-to-Product Pipeline.” The first stage, pre-training, feeds a model very large amounts of text and code so that it learns to predict what comes next. Most of a model’s raw capability comes from this stage, and most of the compute goes into it. Anthropic says each proposed improvement to architecture, data or method is first tested at small scale. It goes into a full training run only if the gain “holds as models increase in size.” Anthropic calls frontier training runs “planned capital deployments with defined inputs” and adds that “not every run performs as our researchers expect.”

Post-training then teaches the trained model to follow instructions, use tools and finish long tasks, by training it on complete, realistic pieces of professional work. The filing calls this collection of training tasks “one of our most durable scientific advantages.” It also says the collection is “difficult to observe from outside the organization.” That is true, and it means a reader has to take the advantage on trust.

Anthropic designs its models and its serving systems together. Its researchers and its inference engineers (who build the systems that run trained models for customers) share “planning cycles, reporting lines, and performance targets,” so that each version is cheaper to run as well as more capable. Anthropic has also chosen not to build image or video generation models, and puts that compute into coding, agents and safety research instead. Those two choices help explain Claude’s position in coding, the use the filing returns to most often.

Release is the third stage, and it comes with checks. Under its Responsible Scaling PolicyAnthropic’s published policy for testing its models for dangerous capabilities, such as in cybersecurity, biology or chemistry, and for deciding what safeguards a model needs before it is released., Anthropic tests each model for dangerous capabilities before release and publishes a “system card” describing the results. It also gives the U.S. Center for Advancing Innovation and Standards for Super Intelligence and the U.K. AI Security Institute access before release. It says it may delay a release or limit access, and it has done so. Mythos went first only to vetted organizations, through Project Glasswing, in April 2026. Fable 5, the general version, shipped in June with restrictions on some cybersecurity, biology and chemistry tasks.

Models released in the six months to 30 September 2026 earned more than 80% of third-quarter revenue, according to MD&A⁵. In the nine months to September, Anthropic shipped 250 model versions, platform features and applications. This is evidence of rapid adoption. It does not show how much revenue a model vintage (the models released in the same period) earns over its life, or when an older version stops earning. The filing does not disclose a model’s useful commercial life. That life matters because training is paid for up front, before a model earns anything, while serving costs are paid as customers use it. The gross margin deducts only the serving costs. A model repays its training only if what it earns above serving and distribution costs, before a newer model takes its sales, adds up to more than its training cost.

Anthropic counts the cost of developing a model as an expense in the period it is spent. Note 1 of the financial statements⁴ says these costs “are generally expensed as incurred due to the significant development uncertainty.” It also says capitalized software costs (software spending recorded as an asset and expensed gradually) “were not significant.” Training expense therefore appears in the income statement when it is incurred, without being spread over an assumed model life. Leased and owned equipment is accounted for separately: finance-lease assetsEquipment held under a finance lease, a long lease that is accounted for much like a purchase made with borrowed money. The equipment is recorded as an asset and the rent still owed as a liability. are recorded, and depreciation (spreading the equipment’s cost over its working life) begins, as the infrastructure enters service.

Dividing gross profit by R&D expense shows how much of the R&D expense gross profit covers. In the first half of 2025 gross profit was $144 million against R&D of $1.85 billion, or 0.08 times. In the second half it was $1.67 billion against $4.49 billion, or 0.37 times. In the first half of 2026 it was $13.906 billion against $9.741 billion, or 1.43 times, so in that half-year gross profit covered R&D expense. The comparison leaves out sales and administration costs. Both sides of the ratio are totals. R&D includes research, internal use and personnel as well as training, and the revenue comes from several vintages. The ratio cannot identify the incremental return on any particular training run, meaning the extra profit that one run produced.

Compute (the computing power used to train and run AI models), people and data each limit how easily what Anthropic has built can be copied, and each does so in a different way. For compute, the chip suppliers control the delivery slots. Anthropic’s orchestration software can move work between chips, but it cannot create capacity or cancel the minimum payments it has signed up to. Research staff carry knowledge with them, and pay and the company’s mission may keep them. The filing reports regrettable attrition (departures of people the company would rather have kept) of 1%. That is Anthropic’s own unaudited measure, and the filing gives no comparison with rivals. Training data mixes public, licensed, commissioned and synthetic (computer-generated) material. Licences and customer-specific evaluations (tests of how a model performs on a particular customer’s work) can be hard to copy, while rivals can also get public and synthetic data, on different terms. The filing does not disclose what the data cost to acquire or what return additional data brings. If the price of a task halved, lower serving costs (the cost of running the models for customers) would help at once. Data purchases, the researchers Anthropic keeps and capacity already under contract would still have to be paid for, unless their scale or terms changed.

Until 2026 Anthropic rented substantially all of its computing capacity from three cloud service providers (CSPs): Amazon Web Services, Google Cloud and Microsoft Azure, according to Note 1 of the financial statements⁴. It runs Claude on four families of AI accelerator chips (chips built specifically for AI work): Amazon’s Trainium, Google’s Tensor Processing Units (TPUs), NVIDIA’s graphics processors (GPUs) and AMD’s GPUs. The filing describes a “proprietary orchestration layer” that lets it treat all of them as “a single, fungible pool” and move work to whichever chips are cheapest for it at the time.

In 2026 Anthropic changed how it gets compute. It began leasing accelerator chips directly and placing them in data centres that other companies build and run on its behalf, an arrangement it calls Anthropic Direct. The Business section² says the first such facility became operational this year. Several partners are involved. There is a colocation agreement with Fluidstack (renting space and power in a data centre for Anthropic’s own equipment). There is a 20-year lease of about 401 megawattsA unit of electrical power; a gigawatt is 1,000 megawatts. Data-centre capacity is quoted by how much power the site can deliver to its equipment, because that sets how many chips it can run. of capacity at a TeraWulf campus in Hawesville, Kentucky, with delivery from late 2027. TeraWulf’s 8-K of 6 July 2026¹⁷, the form a listed company files to report a significant event, says that lease is “supported by an investment-grade credit” (backed by a party with a strong credit rating). There are also leases in Australia and Canada. The MD&A, management’s own discussion of the results, explains why: data centres take longer to build but are “a smaller fraction of total compute costs,” so signing for the buildings early and the chips later leaves the larger decision, the chips, open for longer.

The notes to the financial statements show how fast the commitments have grown. In the first half of 2026 Anthropic signed new and amended agreements with about $417 billion of future minimum payments, plus an estimated $125–150 billion for additional TPU capacity from 2028 to 2031 (Note 16). After June it added about $245 billion more, and about $21 billion of short-term commitments (Note 17). The table of contractual obligations, which would show when all of this falls due year by year, is blank in the first filing. So is the headline total, which the summary describes only as “more than $ .” The only year-by-year schedule is in Note 13. At 31 December 2025, non-cancelable purchase commitments were $54.6 billion, of which $7.1 billion fell due in 2026 and $11.1 billion in 2027. By 30 June 2026 the fixed or minimum commitments under the same kinds of agreements had grown to about $265.6 billion, with terms of one to twenty years, before counting the Broadcom leases. Those commitments roughly quintupled in six months.

Filings and announcements outside the S-1 confirm the scale. Amazon’s second-quarter 2026 release¹⁹ says both Anthropic and OpenAI have made “multi-year, multi-gigawatt commitments” to its Trainium chips. Anthropic’s April 2026 announcement⁸ put its own commitment at up to 5 gigawatts and more than $100 billion over ten years. Broadcom’s 8-K of 6 April 2026²⁰ says Anthropic will access about 3.5 gigawatts of TPU-based capacity through Broadcom from 2027. It also adds a condition: “the consumption of such expanded AI compute capacity by Anthropic is dependent on Anthropic’s continued commercial success.” AMD’s second-quarter release²¹ describes a partnership for up to two gigawatts of its Instinct GPUs. Anthropic’s own announcements put the Google arrangement at up to one million TPUs and more than a gigawatt of capacity online in 2026 (October 2025²²), expanded to multiple gigawatts of next-generation TPUs from 2027 (April 2026²³). Microsoft’s November 2025 announcement²⁴ put Anthropic’s Azure commitment at $30 billion; the S-1 now puts it at $44.9 billion.

The SpaceX contract can also be seen from the supplier’s side. In SpaceX’s 10-Q for the quarter to 30 September 2026¹¹, one customer of its AI (“SI”) segment provided 34.5% of the company’s $14.6 billion of quarterly revenue, about $5.0 billion. The segment’s infrastructure revenue rose from $359 million to $7.0 billion year on year. SpaceX does not name the customer. The S-1 says the SpaceXAI arrangements will cost about $3.1 billion a month once all capacity is delivered, which is expected in December 2026. SpaceX’s figures fit a large ramp-up. They cannot be used to rebuild Anthropic’s quarterly compute expense, though: the customer is not named, and timing and revenue recognition (when SpaceX books the income) get in the way.

Delivery also depends on the suppliers’ financing. Nscale’s 25 September 2026 announcement reports $2.36 billion initially closed, with a further $1 billion expected in November. Lambda, 7 May, reports a $1 billion credit facility. TeraWulf’s Q2 10-Q, filed 5 August, describes development and financing obligations. Fluidstack’s 12 November 2025 announcement describes its role in Anthropic’s planned U.S. buildout. Financing gives a supplier time and pays for capacity. It does not prove the capacity will be delivered. These financing figures overlap the S-1 total above and are not added to it.

Note 8 shows costs falling as a share of revenue. Allocated cost of revenue (the direct cost of delivering what customers pay for) was about 88% of revenue in H1 2025 (the first half of the year), 51% in H2 and 15% in H1 2026. Better chips, serving software, utilization (how fully the chips are kept busy), model mix (which models customers use) and realized prices (what customers actually pay) can all affect that ratio. The filing does not separate their contributions or disclose cost per token (the cost of processing each small unit of text). So the ratio shows that costs improved overall, but not why.

Compute is scarce, and the scarce part is controlled by Anthropic’s suppliers, the chip designers and the clouds: NVIDIA, Google’s TPU programme, Amazon’s Trainium and the data-centre and power developers. Anthropic’s advantage is that it can use all four chip families, which few buyers can. That helps it when buying compute, and rivals can copy it: OpenAI has its own Trainium commitment at Amazon. The Risk Factors¹⁰ section states the long-run risk. If the knowledge needed to build frontier models (the most capable AI models) spreads, “competition could increasingly depend on access to large-scale compute rather than on model development capability.” If that happens, the companies that own compute would have the stronger position. Anthropic has chosen to rent most of its compute from those companies, on long, fixed contracts.

In the same six months Anthropic raised $110.8 billion, almost all from preferred stock, and ended June with $131.7 billion of cash and short-term investments (Consolidated Statements of Cash Flows, in the financial statements⁴). Outside capital is still central to the expansion. The named compute arrangements total about $637 billion, over terms of different lengths, and estimates for the AMD and TPU arrangements come on top of that sum. Those costs will have to be paid from future customer receipts, or investors and lenders will have to supply more capital. The filing leaves the year-by-year payment schedule blank, so future receipts cannot be compared exactly with the amounts due.

Return on invested capital compares after-tax operating earnings with the capital needed to earn them. A version built only from the balance sheet would leave out model development, which is booked as an expense, and contracted computing capacity. Equipment held under finance leases, however, is recorded as an asset. Three ratios built from disclosed figures describe the business, but none of them settles what that return is. Revenue per dollar of compute expense rose from 0.40 to 2.08. From Q2 2025 to Q2 2026, the increase in quarterly revenue was 2.58 times the increase in quarterly compute expense ($10.768 billion ÷ $4.168 billion). In the first half of 2026 (H1 2026), gross profit, which is revenue minus the cost of delivering the service, covered research and development (R&D) spending 1.43 times. All three compare company-wide totals. None of them shows what an extra dollar spent training a model earns.

If models are replaced faster, each training run has less time to earn back its cost. The filing’s figure for the share of revenue from new models shows why that risk matters, but it does not measure how long a model stays useful. Working out the payback properly would need, for each training run, its cost, the later costs of serving and distributing the model, and its net receipts over its life, including any value reused in later models. None of these is disclosed. Contracts running to 2033 and beyond still leave Anthropic paying for hardware across many model generations.

In the flat-revenue case shown, the price halves and usage doubles, so revenue stays the same. If the cost of serving each token stays fixed, serving costs rise from $2.38 billion to $4.76 billion, and the operating loss grows from $0.693 billion to $3.073 billion. If serving cost per token halves as well, the result goes back to where it started, with the other stated assumptions unchanged. Greater usage growth, different training costs or a different product mix would give a different result. Revenue rose after Opus 4.5 and its price cut, but that increase does not show how much of it was customers responding to the lower price, and the filing does not disclose price elasticity (how much usage changes when the price changes).

On expenses the improvement is substantial: compute expense fell from 251% of revenue in Q1 2025 to 48% in Q2 2026. The filing warns that the cost of new compute can arrive before the full revenue it supports. If revenue keeps pace as the 2026 and 2027 capacity arrives, the case that Anthropic can fund its own expansion gets stronger. Cash generated after capital spending and finance-lease principal payments will decide whether that is happening; the revenue-to-compute ratio cannot decide it on its own.

On 27 February 2026 the President directed all federal executive agencies to stop using Anthropic’s technology, and the Secretary of War directed the Department of War to designate Anthropic a “supply-chain risk to national security.” The Secretary stated that “no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity” with it. The S-1’s Business section², under Legal Proceedings, says this followed Anthropic’s “publicly stated position that our technology should not be used for mass domestic surveillance or fully autonomous weapons.” So far, the most consequential commercial effect of Anthropic’s safety stance has been to cost it a government customer and put its contractors on notice.

The courts have split. The district court’s 27 August 2026 order largely granted judgment to Anthropic on two grounds: retaliation, and whether the government had the statutory authority for what it did. The D.C. Circuit’s separate 25 September opinion upheld the Section 4713The federal procurement law under which an agency can bar a supplier it judges a supply-chain risk to national security from its contracts and from the work done under them. supply-chain designation, 2–1. That ruling reaches technology furnished to the Department of War or used to perform its contracts, and stops short of a blanket ban on every contractor’s unrelated business. The S-1 reports a rehearing petitionA request asking an appeals court to reconsider a decision it has already issued. and a stayA court order that pauses a case, or the effect of a ruling, while something else is decided. of the separate Ninth Circuit appeal. Later docket entries, and the reported 13 November response deadline, could not be independently retrieved. The company warns of material revenue losses but does not quantify them.

The second cost came from the Commerce Department. On 12 June 2026 its Bureau of Industry and Security, which administers export controls, required an export licence before any foreign national could use Fable 5 or Mythos 5. That covered foreign nationals inside or outside the United States, including Anthropic’s own staff. Anthropic said the order meant it “must abruptly disable Fable 5 and Mythos 5 for all our customers,” as reported by CSIS on 16 June 2026¹⁶. According to the S-1’s Government Regulation section, the restriction was lifted on 30 June after Anthropic updated its safeguards. For about eighteen days the company’s newest models were unavailable to every customer.

Safety also costs money directly, and the costs are growing. Anthropic expects to invest at least $1 billion over five years to build an external review programme with Accenture, though the terms are not final. Accenture’s 18 September 2026 announcement describes the proposed programme, a team of evaluators embedded at Anthropic, and gives each company’s planned investment as at least $1 billion over five years. These are announced intentions, not completed spending. The evaluation spending pays for confidence in what Anthropic releases. Separately, Anthropic has committed up to $100 million to cyber-defence work through Glasswing.

Anthropic’s public benefit commitments are larger. The donation-matching programme, under which Anthropic matches employees’ charitable gifts of stock with new shares, cost $541 million in 2025 and $2.0 billion in the third quarter of 2026 alone (Note 17 of the financial statements⁴). Its terms are set out in the donation-matching agreement, Exhibit 10.7, filed 26 October. The Anthropic Public Benefit Fund received 5% of the fully diluted shares (the share count including shares that options and similar rights could create) in October 2026. The board may add up to 2% more each time the share price rises 50% above the previous milestone. Both the donation matching and the fund are paid in shares, so they show up as dilution, meaning existing shareholders own a smaller part of the company, as well as expense. The donation matching brings in no customer revenue that the filing discloses.

Enforcing its own rules costs revenue as well. The filing attributes 5–7% of revenue in the first nine months of 2026 to customers who broke its usage policy or commercial terms, and the company is now choosing to give that revenue up. The filing does not say what kinds of violations these were.

The release schedule is the effect hardest to price. The filing says evaluations can delay training, releases or access. Shared industry standards could put the same constraints on competitors and reward compliance work already done. Limits that apply only to Anthropic can leave it bearing those constraints alone. The government exclusion shows a commercial cost, and the private-industry deployments show a possible benefit. The net financial effect of the safety stance is not disclosed.

The S-1’s own Competition section names OpenAI and Google as the direct competitors, followed by “large, well-capitalized technology companies,” open-weight models “released by state actors and international and domestic research organizations,” enterprise software companies where Anthropic’s applications overlap theirs, and “the status quo.” The Risk Factors are more specific: competitors “have attempted in the past and will likely continue to attempt to replicate, distill, extract, or otherwise derive the functionality” of Claude. Distillation means training a cheaper model on a stronger model’s outputs. It is one way cheaper models can copy a frontier model’s lead within months. The filing does not say how often it happens or who does it.

Unusually, Anthropic’s biggest partners are also its competitors, and some of them back its main rival. Amazon’s second-quarter 2026 release¹⁹ says Bedrock, its AI model marketplace, “provides the best selection of leading models” and lists OpenAI’s GPT-5.6, SpaceX’s Grok 4.3 and Google DeepMind’s Gemma 4 as new additions next to Claude Opus 5. Amazon’s 10-Q³¹ records $28.7 billion invested in OpenAI’s preferred stock in the first half of 2026 and a further $21.3 billion afterwards. It also records a $100 billion expansion of OpenAI’s AWS commitment. Google sells Gemini through the same Google Cloud channel that sells Claude. Alphabet’s 10-Q³² lists “the sale of TPU systems” among Google Cloud’s revenues. Microsoft records gains from its investment in Anthropic, $3.2 billion in its June quarter according to its earnings release³³. Its non-GAAP results, the adjusted figures it reports alongside its official accounts, separately adjust for its investment in OpenAI. SpaceX sells Anthropic compute while owning Grok and, since August, Cursor.

The filing only partly covers what happens to Anthropic if one of these partners promotes a rival model or its own. Anthropic keeps a direct contract with each customer who buys through a marketplace, so a change at the cloud provider “would not independently eliminate those contractual relationships.” But the Risk Factors¹⁰ add that customers “may be unwilling or unable to transition to alternative access points on commercially acceptable terms,” and the cloud console is where many enterprises approve spending. With 48% of revenue passing through the three marketplaces, every tenth of that channel lost to a rival model is about 4.8% of Anthropic’s revenue. The partners also see Anthropic’s prices and volumes across their platforms, which the Risk Factors acknowledge gives them “visibility into our pricing and commercial terms.”

Price lists show what competing models cost. Customer evaluations show which model does a particular job best. In Legora’s September 2026 benchmark, Fable 5.1 led overall legal-task quality at 1.06 times the 15-model average, while Gemini Flash and open models produced cheaper results on other parts of the comparison. Legora explicitly keeps its system model-agnostic, meaning it is not tied to any one company’s model. The result supports a Claude quality advantage on some demanding work, with moderate confidence. A customer is still free to buy a cheaper model for simpler work.

Meta competes through both open models and frontier products. Its third-quarter 2026 10-Q³⁴ describes several forms of licensing, and its 21 October results release announces Muse Wonder as its first frontier model. Open-weight alternatives also compete for routine work. DeepSeek’s 13 August V4-Pro release introduced peak and off-peak pricing, and its 17 October V4.1-Flash update extends its coding offer. Qwen’s official Qwen3.8 repository records August open-model releases and coding tools. These are the suppliers’ own claims about their models, and none of them ranks the models against Claude on a common basis. Commercially, they give a customer more models to run itself or send work to, including work that once required a frontier API.

Three of the seven matter most: Claude’s intelligence, compute and adoption. Model quality gives buyers a reason to choose Claude. Contracted computing capacity makes more usage possible, and wider deployment is what can turn that capacity into revenue. International customers produced 28% of H1 (first-half) revenue. In Q3 more than half of large customers used both the platform and an application, up from about 30% in Q4 2025. A customer using several products has more connections to Anthropic, but the filing gives no like-for-like retention comparison showing that those customers switch away less often. The new capacity also still has to be delivered and paid for.

Of the seven, new markets have the least evidence behind them. In April 2026 Anthropic acquired Coefficient Bio, a computational biology company. It has also built a molecular biology lab, sells Claude for Life Sciences, and launched Claude Science, which the filing says Eli Lilly began piloting in October 2026. Through a “Model Hardware Standard,” which is still in research preview, it is extending Claude to laboratory and factory equipment. The filing gives no revenue figure for any of these, and no development cost attributed to them. Reusing the same model in a new field may save work. Domain data, laboratories and safety evaluations still cost money, though, and the filing does not show that these extensions are cheap.

Talent is the priority that shows up in the income statement. Stock-based compensation was $2.2 billion in the first half of 2026. At 30 June, $26.6 billion of compensation for restricted stock units (RSUs, grants of shares that vest over time) had not yet been recorded as an expense, because those units vest only when the company lists its shares; after June the board granted another 91 million RSUs worth $5.8 billion (MD&A, Critical Accounting Estimates; Note 17). Revenue per employee supports the filing’s claim that Anthropic can “grow our business without growing headcount linearly”. But the cost of keeping those employees is rising faster than headcount.

The largest recorded legal settlement is $1.5 billion, in Bartz v. Anthropic, a case in which authors challenged the copying of books. The court approved the settlement in its 20 July 2026 final-approval order; the link goes to a copy of the court document. S-1 Note 13⁴ records the charge in Q3 2025, $300 million paid by year-end, and a $1.2 billion current liability (an amount due within a year) at June 2026. The release is the set of claims the authors gave up. It covers only past conduct involving the covered books as inputs, meaning the material that goes into training. It does not cover claims about what the models output, or anything Anthropic does in future. Settling over the training library therefore leaves Anthropic exposed to future copyright claims.

Judge William Alsup’s 23 June 2025 fair-use order³⁶ treated two things separately: using books to train a model, and acquiring and keeping a library of pirated copies. Fair use is the US legal doctrine that allows some uses of copyrighted work without the owner’s permission. The order held that the training use was fair, and so was digitizing books Anthropic had bought. It left the pirated copies for trial. So the way a company obtains its training data affects what it costs, even though not every item it trains on has to be licensed. Public, licensed, commissioned and synthetic data are different sources, and each has its own cost and legal exposure. Anthropic names those categories but does not say how its spending divides among them. If a dataset is disputed, replacing or licensing it can cost money again when the next generation of models is trained. Paying the settlement does not give Anthropic sole use of its training data.

The S-1’s Legal Proceedings section reports other cases: suits by individual authors who opted outChose not to be covered by the settlement. Authors who opt out give up its payment and keep the right to bring their own claims. of the settlement, claims by Concord and other music publishers, cases in Canada brought by authors and news publishers, and Reddit’s litigation over the use of its data. It does not estimate a loss for any of them. The linked CourtListenerA free public website that republishes US court records and dockets. docket pages showed no readable current entries during this review. So three reported developments are not independently verified here: the 21 October Concord hearing, the latest ABGABG EPE IP v. Anthropic, a separate suit in the federal district court for Northern California, listed in the Evidence Register. filing detail, and the narrowing of Reddit’s claims. The analysis relies on the pending matters Anthropic discloses and on the orders above that could be retrieved, and it assumes no new court ruling. Claims about model outputs can require guardrails or payments. Claims about how data was accessed can add to the cost of acquiring it, even when training itself qualifies as fair use.

Regulation affects how the business works in three places. Export controls (government licence requirements on sensitive technology) can stop a model being used at all: the June 2026 order made Fable 5 and Mythos 5 unusable for anyone until the licence requirement was lifted. The S-1 says similar restrictions “may limit our ability to recruit talent” and serve customers abroad. The Section 4713 designation, upheld on appeal, bars Anthropic from military work and from any contractor work done for the Department of War for as long as it stands. The EU AI Act puts obligations on providers of general-purpose models “with systemic riskThe EU AI Act’s category for the most capable general-purpose models, those the law treats as able to cause harm at large scale. Their providers carry duties beyond those that apply to other models.,” including evaluation, incident reporting and a public summary of training data. The filing lists these under Government Regulation.

The orders establish two things: a $1.5 billion settlement of claims over past use of the covered books as inputs, and a live conflict over government procurement. The settlement does not cover claims over book outputs, future conduct, authors who opted out, music or news. The European Commission’s general-purpose AI guidance and Code of Practice explanation, both accessed 27 October 2026, add a compliance requirement that recurs: transparency and copyright policies, plus safety duties for systemic-risk models. The duties have applied since 2 August 2025, and the Commission’s enforcement powers since 2 August 2026. Models already on the market before August 2025 have a transition period to August 2027. Signing the voluntary code is one way to comply, and signatories are still bound by the law. Evaluation and documentation add to the cost of releasing a model. Customers and regulators can reward the providers able to meet those requirements.

On the facts the filing gives, that is a reasonable accounting judgment, and Ernst & Young audited the 2025 statements prepared on that basis. The judgment has two consequences. First, revenue is larger than the share of customer payments that Anthropic keeps. Second, the cloud providers’ fee is recorded under “sales, marketing, and partnerships,” an expense that is deducted only after gross margin is calculated. Gross margin is therefore measured on a revenue figure that still includes that fee.

The filing does not state the fee rate. MD&A discloses how much the fees paid to channel partners increased: by $341 million in 2025 and by $2.7 billion in H1 2026. The H1 2026 total is that increase plus whatever was paid in the first half of 2025. That earlier amount cannot have been more than $120 million, because $120 million was that half’s entire sales, marketing and partnerships expense. Total channel fees in H1 2026 were therefore approximately $2.70–2.82 billion, or 16.6–17.3% of total revenue. Dividing that by marketplace revenue of $7.82 billion gives 34.5–36.1%, but only if every dollar of channel fees was paid on marketplace sales. The filing does not disclose how the fees are split. For 2025, Reuters’ 30 September 2026 calculation⁴⁰ put total channel fees at about $351 million, or 16 cents per dollar of marketplace revenue. That is a secondary estimate based on a confidential draft.

The figures support one conclusion: channel fees grew faster than total revenue. Any of three things could explain it. Contracted fee rates on the marketplaces may have risen, the mix of channel partners may have changed, or the fees may be tied to the cost of delivering the service. MD&A says the expense can fluctuate with distribution terms and with the cost of delivering the service, but it does not separate those effects. The 34.5–36.1% figure depends on the unconfirmed assumption that all channel fees come from marketplace sales, so it cannot show that AWS or Google doubled its fee rate. Whichever explanation is right, the conflict is unchanged. The cloud provider that distributes Claude owns the console where customers buy it, and it sells rival models alongside Claude.

Subtracting the estimated channel fees from revenue would leave about $13.5 billion, roughly 17% less than reported. A revenue multiple (a company’s value divided by its revenue) calculated on that smaller figure would be about 20% higher. Measured against that smaller revenue figure, gross margin would be about 82%. If reported revenue is kept and the channel fees are subtracted from gross profit instead, the result is about 68%. These are analytical recalculations, not alternative GAAP figures. The reported 85% gross margin leaves out channel fees, training and corporate costs. The 68% measure also deducts the estimated channel fees, without identifying which partners received them.

The other revenue items are conventional. Subscriptions are generally paid in advance and recognized as revenue evenly over the term. That produced deferred revenue (payments received for service not yet delivered) of $1.65 billion at 30 June 2026, up from $269 million at the end of 2025. Remaining performance obligations, contracted revenue not yet recognized under non-cancelable contracts, were $10.8 billion, of which about 75% is expected within twelve months (Note 2). One point from How the Money Comes In also applies here. Revenue from customers later found to have broken the usage policy is recognized as earned and is not reversed, so the 5–7% of nine-month revenue in that category stays in the reported figures.

The quarterly table is blank for the third quarter, the quarter after the last one it reports. Anthropic recorded a $2.0 billion donation-matching charge in that quarter (Note 17), and SpaceXAI capacity was rising toward about $3.1 billion a month by December. SpaceX reports about $5.0 billion from one AI-segment customer in Q3 but does not name the customer. These facts imply cost pressure. They are not enough to reconstruct Anthropic’s quarterly revenue or operating result, because the NRR denominator is missing and nothing confirms that SpaceX’s unnamed customer is Anthropic. The first amendment to the filing should provide the missing comparison.

The filing expects margins to come under pressure as capacity the company has already committed to is delivered, and as it buys more expensive short-term compute. Q2 2026 was its best reported operating quarter. The filing does not say how much of that came from temporarily high utilization and how much from lasting technical gains. Quarterly gross margin is not available either. Note 8 breaks costs down by function for full years and half-years, but not for the six quarters a quarterly series would need. So Figures 17–18 show quarterly revenue, compute and operating margins, and Figure 15 shows gross margin by half-year, as the filing discloses it.

Anthropic’s net loss for the first half of 2026 was $80.7 billion. Almost none of it was cash. The loss includes a $72.4 billion charge for the rise in the fair valueThe price an asset or liability would fetch in an orderly sale between market participants on the measurement date. For instruments with no market price, such as these notes, it is estimated with valuation models. of convertible notes and related instruments, and a further $8.8 billion of net other expense, mostly from first recognizing and then revaluing the note purchase agreements described in the next section; both follow from the rise in Anthropic’s own valuation (MD&A⁵, Results of Operations). Of the $132.9 billion accumulated deficitThe running total of the losses a company has recorded since it began, less any profits. at 30 June 2026, $113.2 billion came from these fair-value changes. When a company’s value rises sharply, the notes it owes, which convert into its shares at fixed prices, become more expensive to settle, and the accounting records that as a loss. The charge stops when the notes convert.

The company reports three adjusted measures to see past this. Adjusted Operating Income excludes stock compensation, non-cash charitable contributions and the Bartz settlement: $1.6 billion for the first half of 2026. Adjusted Net Income also excludes the fair-value charges: $2.7 billion. The fair-value exclusion is fair; those charges are about the capital structure, not the business. The stock-compensation exclusion is the usual one and the usual objection applies. Stock compensation is paid in shares that dilute owners, and with $26.6 billion of RSU cost still to come it will be large for years. Adjusted Net Income also includes $1.1 billion of interest income in the half.

Cash tells a steadier story. H1 2026 operating cash flow was $3.929 billion, against annual outflows of $1.6 billion, $5.7 billion and $2.6 billion in 2023–25. Deducting $482 million of capital spending and $4 million of finance-lease principal leaves $3.443 billion. Working-capital movements added $1.168 billion, and interest income was $1.111 billion. Removing both as a sensitivity leaves $1.164 billion after capital spending and lease principal. The corresponding $1.650 billion figure is before those two outflows. The interest adjustment assumes recognized interest approximates cash received; this is an analytical sensitivity rather than a company-defined cash measure.

Free cash flow measured that way will soon understate the cost of compute. The Broadcom equipment leases, about $161.2 billion of rent over five years, are finance leases. Under finance-lease accounting, the leased equipment goes on the balance sheet as an asset and its rent splits in two: the amortization of the asset and the interest appear as expenses, but the cash paid to reduce the lease liability is classified as a financing outflow, not as capital spending or operating cash. A free-cash-flow figure calculated as operating cash flow less capital spending will therefore leave out most of what Anthropic pays Broadcom for the TPUs. Principal payments on finance leases were only $4 million in the first half of 2026, because most of the leased capacity had not yet started. From 2027 the line will be large, and a reader should add it back when judging cash generation.

The commitment gap is the central fact of the cash flow statement’s future. Customers have committed $10.8 billion under non-cancelable contracts, mostly due within a year. Anthropic’s named cloud minimums and Broadcom lease principal total $430.6 billion across much longer terms; the wider $637 billion inventory also includes ceilings and delivery conditions. The larger inventory is about 59 times customer RPORemaining performance obligations: revenue customers have committed to under non-cancelable contracts that has not yet been delivered and booked. Anthropic’s was $10.8 billion at 30 June 2026., a comparison of different scopes and durations rather than a cash deficit payable today.

Amazon holds the most valuable instrument in Anthropic’s capital structure. Under note purchase agreements signed in 2023 and 2024, Amazon lent Anthropic money in exchange for convertible promissory notes, debt that converts into shares. About $6.2 billion of principal was outstanding at 30 June 2026, all of it non-interest-bearing “rollover notes” left over because earlier conversions hit Amazon’s ownership cap. At the IPO these convert into non-voting Class N common stock at the lower of $2.0928 or $5.0478 a share, depending on the agreement, and the IPO price: 1.84 billion shares in all if the cap allowed (MD&A⁵, Convertible Notes). At the Series H price of about $58.90 per common share, those shares would be worth about $108 billion, 17 times the principal. The balance sheet already carries the notes at $108.6 billion.

The cap is the key term. Conversion cannot take Amazon above 12% of the defined capitalization unless it waives the cap. Unconverted amounts roll into ten-year notes, tested again at later quarter-ends at the lowest applicable fixed, IPO or ten-day average price. Later dilution can create room for further conversions, but only while principal remains: this is a finite claim, not a perpetual right to 12%. The notes can also settle by contractual set-offSettling a debt by cancelling it against money the other party owes, so the netted amount never moves as cash.. Amazon’s 31 July 2026 10-Q³¹ carries the notes at $97.9 billion and reports $62.8 billion of first-half upward preferred-stock adjustmentsIncreases in the value at which Amazon carries its holdings of preferred shares on its books. They count as gains in Amazon’s reported earnings, though no cash changes hands., mainly attributable to Anthropic.

The newer agreements are financing tied to compute. Amazon’s 2026 facility makes $5 billion available each time AWS delivers a compute-capacity milestone, up to $15 billion remaining, settled after the IPO in Class N shares at the 30-day average price. Google’s 2026 agreement has $20 billion remaining through 2030, tied to “certain third-party compute agreements” and revenue targets, with a 14.99% ownership cap and an option for Google to buy up to $10 billion of Class N shares at the IPO price alongside the offering. Broadcom’s agreement provides up to $42 billion of notes “solely to finance quarterly rent obligations” under its own equipment lease, at 9% interest paid in kind (13% if revenue conditions are missed), convertible at 90% of the market price (85% in that case), with 150% of principal payable on a change of control. AMD will buy up to $5 billion of Class N shares as its GPUs are delivered. The agreements are filed as Exhibits 4.3⁴¹, 4.4⁴² and 4.5⁴³, with commercial terms redacted.

The pattern is vendor financing paid in equity. Suppliers can fund purchases by taking Anthropic shares or notes, reducing immediate cash needs while linking their investment returns to its growth. Amazon invests, Anthropic buys AWS capacity, and a later equity revaluation can raise Amazon’s reported earnings. Amazon’s $62.8 billion of H1 preferred-stock adjustments was mainly attributable to Anthropic; the disclosure does not assign the whole amount to it. These arrangements finance part of compute demand. Their investment marks are separate from operating cash earned by either company.

The bank facility provides $15 billion, secured, maturing in September 2031, with a $5 billion minimum-liquidity covenant and release of security after two investment-grade ratings. The filed credit agreement, Exhibit 10.10, amended September 2026, names Morgan Stanley’s affiliate as agentIn a loan, the bank that administers the facility for all the lenders, handling payments, notices and collateral. This is a different use of the word from the principal-versus-agent accounting test in section 12.. Underwriting⁴⁴ identifies the arrangers, including affiliates of Goldman Sachs and J.P. Morgan. The banks pricing the IPO are also lenders with an interest in liquidity and repayment.

The first S-1 leaves out the numbers that define an offering: the price range, the number of shares sold by the company and by selling stockholders, the underwriters’ option to buy more, the proceeds and the fees. EDGARThe SEC’s public online database of company filings. showed no amendment as of 27 October 2026 (EDGAR filing record⁴⁵). The filing does set out how the offering is structured. Anthropic and certain selling stockholders will sell Class A common stock, to be listed on the Nasdaq Global Select Market as “ANT.” Twenty-two banks are named as underwriters, with Goldman Sachs and Morgan Stanley as representatives (the lead banks acting for the group) and J.P. Morgan third on the cover (Underwriting⁴⁴).

The selling stockholders are the seven co-founders. The Principal and Selling Stockholders⁴⁶ table lists Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Christopher Olah. Each is listed with between about 232 million and 244 million Class A shares, about 1.69 billion together, or roughly 10% of the fully diluted share count. The table leaves blank how many shares each will sell. Each founder has pledged 80% of their equity to philanthropy. The founders have already sold shares for cash, in tender offersOffers to buy shares directly from existing holders at a set price for a limited time. In a private company they let founders and employees sell some shares before an IPO. in June 2025, at $5.61 a share, and in April 2026, at $25.91, when buyers took 180 million shares for $4.7 billion (Certain Relationships and Related Person Transactions⁴⁷).

The Use of Proceeds⁴⁸ section is general: working capital, “investments in compute and infrastructure,” hiring and possible acquisitions, with no material acquisition agreed. One specific use is stated. Many employee RSUs vest at once on listing. The company will withhold shares to cover the employees’ taxes and pay the tax in cash, and part of the proceeds will fund that. The filing leaves the amount blank.

Several things happen to the share count just before the shares are sold. All preferred stock converts, the voting series into Class A and the non-voting series into Class N. Amazon’s notes convert into Class N up to its 12% cap. The founders’ LLC receives one share of Class F. RSUs that have met their service condition (the required period of employment) are settled in shares, net of the shares withheld for tax. The Anthropic Public Benefit Fund already holds 846,840,960 Class N shares, 5% of the fully diluted count, issued in October 2026. Those shares are legally outstanding. The fund is a subsidiary, and while it holds them the financial statements do not count them as outstanding.

Several mechanisms can add shares after the IPO. The 2026 equity plan reserves 2.95 billion Class A shares and the purchase plan 160 million, with annual increases (Note 17). The larger reserve is about 17% of the inferred fully diluted count. Amazon’s remaining rollover principal may convert as the 12% cap allows, until that principal is used up or settled. The company can issue more Class N shares through supplier financing and donation matching. The board may also issue up to 2% more outstanding shares to the Public Benefit Fund when the share price is above a milestone on 90 of 120 trading days, with each milestone set 50% above the previous one. These mechanisms add shares under different conditions, so they should not all be treated as issued at listing.

The capitalization and dilution tables are blank in this filing, so the dilution a new buyer suffers cannot be computed. That figure is the gap between the offering price and the net tangible book value per share after the offering (assets minus liabilities and intangible assets, divided by the number of shares). The gap will be large: the company had a stockholders’ deficitShareholders’ equity below zero: liabilities and other claims that rank ahead of common shareholders add up to more than the company’s assets. of $127.6 billion at 30 June 2026, before the preferred stock converts into equity.

Amazon and Google do not appear in the S-1’s section on related-party transactions. That section covers dealings with directors, executive officers and “holders of more than 5% of any class of our outstanding voting securities.” Both companies hold non-voting stock, so they fall outside it. The section lists Spark Capital’s purchases in each funding round, a director’s family company buying $10 million of Series H stock, the founders’ tender-offer sales, the company’s funding of the Long-Term Benefit Trust and the employment of the President’s spouse (Certain Relationships and Related Person Transactions⁴⁷).

Amazon’s and Google’s financial and commercial ties to Anthropic are larger than any other counterparty’s. Amazon has invested $8.0 billion in notes and $10.0 billion in preferred stock, holds notes convertible into 1.84 billion shares, and can provide $15 billion more as AWS delivers capacity. Anthropic has committed to buy at least $110 billion of compute from Amazon through April 2036, and Amazon resells Claude through Bedrock. Google has converted its earlier notes into preferred stock. It put $10 billion into Series G-2 preferred in March 2026 and another $10 billion into Series H in May, can provide $20 billion more, and has an option to buy $10 billion of shares at the IPO price. Anthropic has committed to buy at least $114.5 billion of compute from Google through July 2033. Google also supplies the TPUs that Broadcom leases to Anthropic, resells Claude through Google Cloud and sells its own Gemini models in competition with Claude. All of this is in the S-1, mostly in MD&A⁵ and the Risk Factors¹⁰, which call the relationships “complex dynamics that could give rise to conflicts of interest.”

The S-1 leaves blank the number of Class N shares Amazon and Google will hold after the offering. Neither appears in the Principal and Selling Stockholders⁴⁶ table, which lists holders of more than 5% of voting stock. The notes to the financial statements⁴ give enough to place the stakes. Amazon’s notes convert at listing only until it holds 12% of the company’s capitalization, so the cap sets its stake. Google’s stake can be counted only in part, because the filing does not name the holders of the early non-voting preferred series. Between them the two will own more than 18% of the fully diluted count and none of the vote: Class N shares vote only where the law requires, and the founders’ single Class F share carries 50.1% of the vote.

The commercial contracts themselves are not in the S-1. The exhibit list includes the credit agreement and, with their commercial terms redacted, Amazon’s 2026 financing agreement (Exhibit 4.3), Google’s note purchase agreement (Exhibit 4.5) and Broadcom’s note agreement and master equipment lease (Exhibits 4.4 and 10.11). It leaves out the AWS agreement under which Anthropic committed $110 billion, the Google Cloud agreement under which it committed $114.5 billion, and the Microsoft, SpaceXAI, Nscale and Lambda agreements. The financing agreements refer to “certain commercial agreements” and “compute capacity milestones” that the reader cannot see. An amendment to the S-1 may add them. If it does not, the largest obligations in the business will be known only from the summaries in MD&A.

Amazon earns on AWS usage whichever model wins. Its 10-Q³¹ shows it has hedged its bet: about $50 billion into OpenAI in 2026 and a $100 billion expansion of OpenAI’s AWS commitment. Anthropic’s decade-long commitment to Trainium, Amazon’s own chip, makes that chip a credible product. Amazon’s chief executive, Andy Jassy, said in the April 2026 announcement⁸ that the commitment “reflects the progress we’ve made together.” Google earns on TPU and cloud usage, on gains in the value of its equity stake and on Gemini. Alphabet’s 10-Q³² does not mention Anthropic by name. It lists “the sale of TPU systems” as a Google Cloud revenue source and reports a Google Cloud backlog (revenue already under contract but not yet booked) of $513.9 billion at 30 June 2026. Anthropic has two commitments tied to Google: $114.5 billion directly, and the TPUs behind Broadcom’s $161.2 billion of leases. Together they make Anthropic one of the largest customers of a business that also competes with it.

Neither Amazon nor Google has a vote or a board seat. The seven-member board will have four directors chosen by the Long-Term Benefit Trust and three elected by Class A and Class F holders, and none of them is nominated by Amazon or Google (Management⁴⁹). Their leverage is commercial: the capacity they deliver, the order in which they deliver it, the fees they charge in their marketplaces and how prominently they show Claude next to the alternatives. The ownership caps of 12% and 14.99% stop either company from building up control. They place no limit on what either can do as a supplier or seller.

Anthropic sells Claude through Google’s and Amazon’s clouds, and both also offer competing models. Google’s 8 October 2026 announcement describes agents that route work to Gemini or to Claude depending on the task. AWS’s 26 October roundup adds Claude capabilities to more of its cloud tools. Amazon’s 26 October earnings preview lists rival model families in Bedrock and schedules Q3 results for 29 October. The cloud provider can keep the customer account and the infrastructure revenue while changing which model gets a task. Anthropic can respond by competing on quality, selling directly or offering better terms. The investment relationship does not give Anthropic control of the cloud console, where the model for each task is chosen.

Anthropic is a Delaware public benefit corporation (PBC). Its directors must balance three things: the stockholders’ financial interest, the interests of those “materially affected by our conduct,” and a stated public purpose, “the responsible development and maintenance of advanced AI for the long-term benefit of humanity” (Our Governance Structure and Its Goals⁵⁰). The filing says there is little case law on how courts review that balancing. The company will judge its own performance against qualitative standards set by the board, and will report the results to stockholders at least every two years.

Control rests on five classes of common stock. Class A shares, the ones listed on the exchange, carry one vote each. Class N shares have the same economic rights and no vote. The single Class T share, held by the Long-Term Benefit Trust (LTBT), elects four of the seven directors, a majority. The single Class F share is held by an LLC (limited liability company) whose members are the seven co-founders. It carries 50.1% of the votes on all ordinary matters, including electing the other three directors. Class E shares will be issued to employees (one share per six months of service, plus one per share pledged to the donation programme), and vote only on a few matters concerning the trust. Class T, F and E shares carry no economic rights.

The trust is designed to be independent of both money and management. It has five authorized trustees, with one seat vacant. The current trustees are Neil Buddy Shah (chair), Richard Fontaine, Ben Bernanke and Christopher Stubbs. Trustees are paid $400,000 a year in cash and may not hold Anthropic securities. They choose their own successors by majority vote, after giving notice to the board’s nominating committee. The company funds the trust: $8.4 million in the nine months to September 2026. In August 2026 it also donated $450,000 to the think tank Mr. Fontaine leads. The trustees and company executives have “collaborated closely” on choosing trustees. The trust agreement⁵¹ is filed as Exhibit 99.1.

The founders’ control has a defined end. A founder stops qualifying on any of the specified triggers for departure, death, incapacity, cause or ownership. One ownership trigger is holding less than 2% of the Class A shares held at the IPO. If fewer than three members qualify for 90 consecutive days, a further 450-day transition begins, and after it Class F and Class E disappear. Class F can also be redeemed earlier if its holder asks, or if the LLC is dissolved. Founders who die or become incapacitated in the first ten years may be replaced by successors. Because the ownership threshold is small, a qualifying founder can dispose of most of their economic stake and still keep a role in control. These rules are in the form charter, Exhibit 3.3, and founder LLC agreement, Exhibit 99.2, filed 26 October 2026.

Unlike Class F, the trust has no automatic sunset; it does not end on its own. One way to remove it requires 80% of the board and Class T approval, which itself requires 80% of trustees. The routes that involve the founders require approval from Class F, Class E and Class A. In the first three years, the founders must be unanimous and a majority of Class E votes cast must approve, at a 40% quorum (the minimum share of votes that must be present for the vote to count). Alternatively, while Class F survives, two-thirds of founders and 75% of Class E votes cast must approve, at a majority quorum. Both of these routes also require a majority of Class A votes cast at a majority quorum. Once Class F is gone, holders of 85% of outstanding Class A shares can end Class T. The routes count votes differently. Where approval is measured on votes cast, abstentions still help establish a quorum. Where it is measured against all outstanding shares, as in the 85% route, every share that does not vote counts against. The charter and trust agreement specify these thresholds.

Class F’s vote on ordinary matters also depends on the founders agreeing. Under Exhibit 99.2, a majority of founders normally instructs how Class F votes. If that majority gives no common instruction, Class F votes in the same proportions as the Class A shares, provided a majority of Class A shares are present at the meeting. Special matters have higher thresholds, and the default is against approval. The form bylaws, Exhibit 3.5, and stockholders’ agreement, Exhibit 10.14, set out the rest of the voting rules. Separately, the Daffy and Dechomai proxy agreementsAgreements that hand the right to vote shares to someone other than their holder. Here the holders of shares donated to charity give that right to the CEO or a proxyholder the board designates., Exhibits 10.12–13, give the CEO or a proxyholder designated by the board authority over covered donated shares. Shares donated by the CEO and specified related persons are excluded. All were filed 26 October. Shares held for charity can therefore still be voted by the CEO or the board’s proxyholder.

Class A holders, who own the listed shares, are left with limited control. While Class F exists they hold at most 49.9% of the vote on ordinary matters. In practice the founders decide those matters, including any merger, sale of the company or charter amendment that needs stockholder approval. Class A holders have no vote on the four trust-appointed directors. Because the trust elects a majority of the board, Anthropic is a “controlled company” under Nasdaq rules, a status that exempts it from some of the exchange’s board-independence requirements. It will use the exemption only for its nominating committee, which will not be fully independent. The board is currently majority independent, and the audit and compensation committees are fully independent.

The filing states the commercial purpose of the structure openly: to let management make long-term investments “rather than facing pressure to prioritize short-term results.” In this business that means the compute commitments described in Part I, signed years ahead of the revenue meant to pay for them. Public investors are being asked to fund those commitments with no means of stopping them.

Dario Amodei’s 2025 compensation was $17.96 million. It was made up of a salary of $654,167, RSUs with a grant-date value of $9.26 million, options worth $4.03 million and $4.01 million of other compensation. The other compensation included $2.75 million of personal security and $97,055 for personal use of private aircraft required by his security programme. It also included $1.15 million that was the difference between the price paid for his shares in the June 2025 tender offer and their fair value. Daniela Amodei’s total was $16.41 million, made up the same way. The chief financial officer, Krishna Rao, received $720,250, all cash (Executive Compensation⁵²). Like everyone else’s RSUs, these vest only if the company lists its shares.

The filing discloses no special founder award tied to share-price targets. The founders’ economic stake is the roughly 240 million shares each already holds, most of which they have pledged to give away. The contract terms for grants and for termination benefits are in the 2021 equity plan and award forms, Exhibit 10.4, and the executive severance plan, Exhibit 10.5. The plans were filed on 26 October 2026. Note 17 to the financial statements reports amendments made in October that provide cash severance and accelerated vesting for specified kinds of termination, and says the awards will be revalued afterwards. Voting control comes from Class F, separately from these compensation rights.

On 30 October 2025 Anthropic dismissed Grant Thornton, which had audited the 2023 and 2024 financial statements, and hired Ernst & Young to audit 2025. The Change in Independent Accountant⁵³ section states that Grant Thornton’s reports were unqualified (clean, with no reservations). It also states that there were no “disagreements” on accounting principles, disclosure or audit scope, and no “reportable events,” the SEC’s terms for problems serious enough to disclose. Grant Thornton’s letter to the SEC, dated 26 October 2026 and filed as Exhibit 16.1⁵⁴, says it agrees with those statements. The filing gives no reason for the change.

Ernst & Young’s report on the 2025 statements names two critical audit matters, the areas it found hardest to judge. The first is loss contingencies, meaning the reserves for litigation. The second is the fair value of the convertible notes, $41.7 billion at the end of 2025, which the auditor calls “particularly sensitive to changes in the relative weight” given to two valuation methods. Neither is unusual for a company in Anthropic’s position. They are also the areas a skeptical reader would pick out: what the lawsuits may cost, and how much of the capital structure belongs to Amazon.

A lock-up is an agreement by existing holders not to sell for a period after the IPO, so that the market is not flooded with stock all at once. The shares held back by lock-ups are called the overhang, because holders can sell them once their lock-ups end. Anthropic has four different lock-up schedules. Each one releases stock in small tranches (portions) tied to the company’s earnings reports, instead of in one block at 180 days. The Underwriting⁴⁴ and Shares Eligible for Future Sale⁵⁵ sections of the IPO filing set out the schedules, but every row leaves the share counts blank.

Directors other than the founders, officers, and certain current and former employees may sell 10% of their locked shares from the first day of trading. They may sell a further 35% two trading days after the company releases its results for 2026. They may sell the rest at the earlier of two trading days after the first-quarter 2027 results or day 180. Another group of stockholders may sell 6% on each of days 27 and 48, 7% on each of days 62 and 76, 12.5% after the 2026 results and 9.5% on each of five later dates, which brings them to 86% before day 180. Stockholders who held more than 0.5% of the fully diluted shares (a count that includes shares that options and similar rights could still create) are on a 365-day schedule. Through day 180, each of their steps is 30% of the matching step on that 6% schedule (1.8% where that group gets 6%), so they reach 30% by day 180. After that come steps of 5.6–7% through the third-quarter 2027 results. The founders may sell 29% of their shares from the earlier of the first-quarter 2027 results or day 180, and the rest at the earlier of the third-quarter 2027 results or day 365.

Goldman Sachs and Morgan Stanley, the underwriters’ representatives, may waive lock-ups. If they grant certain specified waivers, other holders can be released pro rata, meaning in the same proportion. Registration rights (a holder’s right to have the company register its shares so they can be sold publicly) appear in the investors’ rights agreement, Exhibit 4.2, and in the Broadcom and Google agreements, Exhibits 4.6–7, filed 26 October 2026. Registration makes a sale possible under securities law, but a holder still has to wait until the lock-up expires. After the IPO, a Form S-8 (the registration form for shares issued under employee stock plans) will cover employee equity. Class N holdings also carry lock-up or standoff terms (a promise made before the IPO not to sell for a period after it), and Amazon’s quarterly report, the 31 July 10-Q³¹, says it expects a customary lock-up.

Shares become eligible for sale in small, frequent tranches, many of them tied to earnings. No IPO pricing date has been verified, so the release dates cannot yet be put on a calendar. Once the share counts are known, the overhang at each release can be measured. For each group, let G be the shares its lock-up covers after IPO sales, and let a(t) be the cumulative fraction released by time t, as shown in Figure 23. Each scheduled release makes another [a(t) − a(previous)] × G shares eligible for sale. Dividing that increase by the shares that were freely tradable just before the release (the float) gives the incremental overhang. Dividing a(t) × G by the same float gives cumulative eligibility. When releases for different groups fall on the same day, add them together. On the first day, tradable supply includes the primary shares (new shares the company sells in the IPO), the secondary shares (existing shares current holders sell in it) and the eligible 10% employee tranche, subject to registration and resale limits. The share counts, the overlap between groups and the actual sales are still blank. That means eligibility cannot yet be turned into a numeric float or a prediction of how much holders will sell. For the founders, the first tranche is 0.29 × (covered holdings minus IPO shares sold), and the 489 million figure above is that amount with nothing subtracted.

Two of these gaps matter more than the rest. The first is realized price per unit of work, which is what customers actually pay, after discounts, for a fixed amount of work. The filing’s central economic claim is that efficiency lets Anthropic cut prices “while preserving our margins.” Without the realized price, a reader cannot separate the effect of efficiency from the effect of selling into a market short of capacity. The second is the year-by-year schedule of compute payments. Without it, an investor cannot see when the fixed costs fall due compared with the revenue they are meant to pay for. That timing is the central question for 2027 and 2028.

A third gap is the missing commercial agreements. The financing agreements with Amazon, Google and Broadcom are filed. They refer to compute milestones and to commercial agreements that are not filed. Together, the AWS and Google Cloud agreements carry at least $224.5 billion of minimum spending. They would show the true-up mechanics (how payments are settled when actual use differs from the committed minimum), the renewal terms, and any most-favoured terms (a promise that no other customer gets a better price) or exclusivity terms. Their absence from the exhibit list is the largest single gap in what an outside reader can check.

Claude Code and agents are provisionally neutral, with low confidence in the economics. Customer deployments and Legora’s task evaluations point to demand for capable agents. Selling directly can keep the customer relationship with Anthropic. The filing’s spending comparison covers the top 5% of Enterprise users who use Code, who spend about twenty times the average non-user. That figure describes the top 5%, not the average Code user, and it does not show that Code caused the extra spending. Shopify’s substitution and Jane Street’s vendor choice put limits on how firmly customers will stay. No figure exists for the cash this line generates after serving and training costs. So the evidence supports a promising product. It does not show a superior return.

The API and platform are neutral, with moderate confidence in the competitive assessment. Customers can send each request to whichever model they choose, while Anthropic still carries its research and serving costs. The 85.4% gross margin and 900% NRR (net revenue retention) are company-wide figures, and neither shows what the API returns. Consumption billing, which charges by usage, is 80% of revenue. Because it also includes app usage, the API’s share of revenue is undisclosed. If Claude keeps a lead in the quality and cost of completed tasks, the profitable requests could stay with Claude. Customers’ own routing systems give them a practical way to move the rest.

Subscriptions and apps are neutral, with low confidence. Subscription billing is about a fifth of revenue, but the apps also bring in usage revenue, so their total share is undisclosed. Individuals can bring Claude into the companies they work for, while competitors bundle their own assistants into existing software. Microsoft reports more than 30 million paid Microsoft 365 Copilot seats in its 29 July 2026 release³³. Claude’s public plan prices show how access is sold. They do not show how many subscribers stay, what it costs to win them, or the cash return on serving heavy users.

The partner channel is exposed, with moderate confidence. It carries 48% of revenue through partners’ consoles, the platforms where their customers choose and run models. The partners who own those consoles can promote competing models and earn money on other workloads. Anthropic keeps the customer contract and sets Claude’s prices, which limits how much control the partners have. The grade rests on those incentives and on the options to switch. The estimated pool of channel fees does not establish the rate set in marketplace contracts, or the return on this line. A direct disclosure that fees are stable, and that customers’ ability to switch favours Anthropic, would weaken this assessment.

The margins in the grid are harder to believe than the growth. Second-quarter operating margin was 8%, and the filing does not disclose quarterly cash flow. Compute took 48% of revenue, personnel 16%, sales, marketing and partnerships 18%, and other costs 9%. The core commitments, $430.6 billion of minimum payments to named cloud providers plus principal owed to Broadcom, will need substantial future cash coming in to meet them. The schedule of when those payments fall due is blank, though, so there is no way to lay out cash in and cash out year by year. A wider list of commitments totals $637 billion and includes arrangements that can be cancelled or that depend on delivery. Neither figure is a revenue forecast from management. Taking either one's average annual size as a guide to revenue would assume what the analysis needs to test: when the payments fall, and how much compute each dollar of future revenue will need.

Finance-lease accounting also affects the free-cash-flow figure. From 2027, much of the cost of the Broadcom TPUs will be recorded in the cash flow statement as financing outflows, repayments of lease principal, rather than as capital spending. A buyer who values the company on operating cash flow less capital spending will then be using a free-cash-flow figure that leaves out a large part of its compute bill. The grid's free cash flow is after all compute costs, however they are classified.

A valuation at these levels would be paying for customers to keep preferring the company's models as each new generation arrives, for revenue to grow into the computing capacity already contracted, and for the cost of serving customers to fall in step with the prices actually realized per task. The filing shows rapid revenue growth and total expenses falling as a share of revenue. The record gives only limited evidence on future task prices, on how long each model stays in service, and on the returns the new capacity earns once it arrives. Those assumptions carry much of the value in the grid.

The strongest bear case (the argument against owning the shares) is that a temporary shortage of compute capacity makes the best reported quarter look better than it otherwise would. Revenue outran delivered compute. From December, SpaceXAI’s fully delivered rate would be about $37 billion a year, compared with Q2’s $22 billion annualized total compute expense. Broadcom leases add costs from 2027, and management expects pressure on margins. There is also limited room to raise the price of a task. Posted prices match rivals’, rival models sit in the same cloud consoles, and customers own the routing that decides which model gets each request. Shopify shows how buyers can move routine subtasks to their own models. If capability spreads faster than Anthropic can recover its training costs, the research programme becomes a recurring expense needed just to keep the calls it already has. The S-1’s warnings about distillation (training a cheaper model on a stronger model’s outputs) and AI-assisted research show how that could happen. They do not show that it is happening.

Anthropic’s named cloud minimums (the least it has committed to spend with specific cloud providers) and its Broadcom lease principal total $430.6 billion. Against that sits $10.8 billion of customer RPO (revenue customers have contracted for but Anthropic has not yet recognised), over a shorter period. That financing is what makes a slowdown in demand matter. The wider $637 billion inventory adds ceilings and conditional capacity on top. The valuation scenarios require years of rapid growth. If receipts slow, the fixed payments could absorb what would otherwise be left for shareholders. As long as founder and trust control survives, outside Class A holders cannot direct spending by themselves. Public-benefit issuance, awards and supplier financing could all dilute existing holders. The procurement designation and the June export restriction show political constraints at work. Customers are also concentrated: five of them paid 36% of Q1 revenue. The bear case does not need the product to fail. Enough of the growth could go to customers and suppliers while cash returns to common holders stay poor.

The evidence backs the bear case on three exposures: the timing of capacity, customers moving work elsewhere, and how little control public shareholders have. It does not show how far margins will shrink, and it does not show a rise in contractual marketplace fee rates. Both unknowns affect the range of possible cash outcomes. Treating every announced capacity amount as a fixed bill does not settle either of them.

The first effect comes from the supplier financing. Amazon’s 2026 facility, Google’s note agreement, Broadcom’s rent-financing notes and AMD’s purchase commitment all settle in Class N shares at or near the market price. Broadcom’s notes settle at a 10–15% discount to it. After the IPO, Anthropic’s share price becomes part of its cost of compute. A higher price means fewer shares issued for each dollar of compute, and a lower price means more. If the shares fell sharply, the same compute would cost existing holders more dilution at exactly the moment the business was under pressure. The arrangement costs holders little while the shares do well, and it adds to the pressure when they fall.

The second effect shows up in the suppliers’ reported earnings. Amazon’s $62.8 billion of H1 preferred-stock adjustments was mainly attributable to Anthropic, and Microsoft reported a $3.2 billion Anthropic gain in its June quarter. Once Anthropic has a listed price, changes in its value will be easier to see. The suppliers’ investment returns will then rise and fall with the equity value of an important customer, and that customer depends on their capacity. The link can make the suppliers’ reported earnings more volatile, though the link itself brings in no operating cash.

The third effect is on prices across the industry. Anthropic reported an 85% gross margin on serving in an SEC filing. That is evidence that frontier inference can be very profitable at current prices, and it will help every lab raise money, including the rivals that match Claude’s prices. The industry has also ordered a very large amount of capacity for 2027 and 2028. When it arrives, the scarcity that has let sellers hold or raise some prices in 2026 will ease. That would most clearly help customers and the companies building products on top of the models. It would most clearly hurt whichever lab has the most fixed cost relative to its revenue.

The fourth effect involves the Public Benefit Fund. Each time the share price rises 50% above the last milestone, the board may issue up to 2% more shares to the fund. A holder whose shares rise 50% gives back up to about 3 points of that gain to dilution (1.50 ÷ 1.02 = 1.47). The cost is modest, and unusual because it grows as the share price climbs. It also means the fund, managed initially by Dario Amodei, Daniela Amodei, Jack Clark and Neil Buddy Shah, could become one of the larger holders of tradable stock over time.