Weekly Wrap · IPO filing deep dive
Everything important in the S-1, explained from first principles: how the company sees AI, what it sells, the numbers, the compute behind it, the people, and who controls it after the IPO.
The plan
How Anthropic sees AI
Powerful AI, scaling laws, closing the loop, virtual collaborators
Frontier AI from first principles
Tokens, training, inference, agents, compute, the value chain, competition
Products and customers
Models, platform, Claude Code, Cowork, verticals, who pays and how much
The numbers
Revenue, mix, costs, margins, the $80.7bn loss, quarterly trends
Compute and capital
Chips, data centres, the Amazon, Google, Broadcom and AMD deals, funding, debt
People, research and safety
Founders, board, talent, scaling policy, constitution, incidents, regulation
The IPO and who controls it
Offering, public benefit corporation, five share classes, the trust, ownership
Every fact comes from the S-1 itself. Where Anthropic states a belief or a plan, the slide says so. Where the draft leaves a figure blank, the slide says it is not yet disclosed.
FY = calendar year to 31 Dec. H1 = six months to 30 Jun. 9M = nine months to 30 Sep. “Adjusted” = the company’s own non-GAAP measures. $bn = US$ billion.
Source: Anthropic S-1 (26 Oct 2026). Structure by TheWrap.
Anthropic at a glance
Anthropic builds Claude, a family of frontier AI models, and sells access to it through an API, a developer platform and applications such as Claude Code and Claude Cowork. It was founded in 2021 and is a Delaware public benefit corporation.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary".
What this document is
The document a company files with the US securities regulator before it sells shares to the public for the first time.
The cover states it is incomplete and may change. The price range, share counts, offering size, Q3 2026 revenue and several other figures are blank in this version.
Source: Anthropic S-1 (26 Oct 2026), cover page and "The Offering". IPO process description by TheWrap.
Timeline
Amber dots: safety, security and regulatory events the S-1 discloses.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary", "Business", "Risk Factors", "Management's Discussion and Analysis".
Part 1
The worldview the S-1 sets out: what Anthropic believes AI is becoming, how it explains progress, where it sees the opportunity and why it says safety sells. Every view in this part is the company's own.
Part 1 · How Anthropic sees AI
The S-1 defines it as AI systems at the current leading edge of capability and safety, performing at or near the highest level of any existing system across a broad range of tasks.
The five "core advantages" Anthropic lists. It says the mission (1) directs the other four.
One mission guides every decision and keeps talent, research and compute on the highest-value work.
Anthropic says safety and capability must advance together.
The mission attracts researchers; a low-ego, truth-seeking culture keeps them.
Founding team originated Scaling Laws, GPT-2/GPT-3, RLHF and mechanistic interpretability.
Capability and safety research are integrated and shape design, training and deployment.
Each Claude version helps build the next: "closing the loop".
Compute from several partners is run as one fungible pool and sent to the most productive workload.
Four chip families: AWS Trainium, Google TPUs, NVIDIA GPUs, AMD GPUs.
Research, roadmap, sales and Applied AI teams are organised around enterprise use cases.
Anthropic says enterprises are where impact is greatest and feedback richest.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary"; "Business"; "Select Defined Terms".
Part 1 · How Anthropic sees AI
The five attributes of "Powerful AI", as the S-1 lists them
Expert-level thinking. Cognitive capabilities matching or exceeding top human experts across most professional and scientific disciplines.
Every virtual interface. Can navigate text, audio, video and direct computer control, like a human working virtually.
Long autonomous work. Reasons through complex tasks over hours, days or weeks, and asks for clarification when needed.
Works in teams. Coordinates across multiple copies of itself and collaborates with humans and other AI systems.
Reaches the physical world. Digitally controls laboratory equipment, robotic systems and other machines.
Anthropic's term for "a highly capable and impactful form of frontier AI" with the five attributes on the left.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary"; "Business".
Part 1 · How Anthropic sees AI
Illustrative shape only, drawn by TheWrap. The S-1 gives no chart or numbers for this relationship.
The S-1 describes them as the empirical finding that AI model performance improves predictably as model size, training data and compute increase.
The extent, efficiency and durability of these gains are uncertain. The relationships "may plateau, diminish, or fail to materialize at larger scales".
Source: Anthropic S-1 (26 Oct 2026), "Business"; "Risk Factors".
Part 1 · How Anthropic sees AI
Algorithmic and method improvements in training and inference that get more useful output from the same amount of compute (S-1 definition).
Using AI to build more capable AI could pose "a considerable safety risk". Its Responsible Scaling Policy covers what has been termed "recursive self-improvement".
Source: Anthropic S-1 (26 Oct 2026), "Business"; "Select Defined Terms".
Part 1 · How Anthropic sees AI
Anthropic says its "scale and growth create self-reinforcing loops" and that it believes these "create durable competitive advantages". The first loop uses customer feedback to improve Claude; the second turns revenue into the compute, research and talent that pay for the next model.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary"; "Business".
Part 1 · How Anthropic sees AI
Anthropic's name for AI systems that work alongside humans on complex, multi-step work at professional level. They carry the access rights, system connections and context their organisation grants them.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary"; "Business".
Part 1 · How Anthropic sees AI
How Anthropic says it built the $40T+ knowledge-work figure ("Industry, Market, and Other Data")
The S-1 calls the result "the annual economic value represented by global knowledge workers". It says it has not independently verified the third-party data behind its market figures. The four numbers measure different things and are not additive.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary"; "Business"; "Industry, Market, and Other Data".
Part 1 · How Anthropic sees AI
An open standard Anthropic introduced in November 2024 that gives AI tools one common way to connect to data sources, tools and workflows.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary"; "Business".
Part 1 · How Anthropic sees AI
Risks the S-1 describes
Bad actors have already run high-profile cyberattacks using AI to find vulnerabilities, write malware and evade detection. Misuse in biology "could result in even greater harm".
AI at scale could concentrate power and wealth within and between countries, eroding trust and destabilising the geopolitical order.
Misaligned AI could corrupt institutional decisions or cause large-scale harm through autonomous actions that users and operators cannot see or understand.
Malicious instructions hidden in content a model reads, such as a website or document, meant to hijack its behaviour.
Why Anthropic says safety is also commercial
Alignment, interpretability, safeguards, red-teaming and external reviewers.
Anthropic says well-aligned models follow instructions, avoid destructive actions without understanding intent, and resist prompt injection.
"Rigorous safeguards, predictable model behavior, and demonstrable accountability" are, in its view, "decisive criteria" when choosing an AI partner for mission-critical work.
Trust lets enterprises expand the work and systems they hand to Claude.
Example it gives: in April 2026 its frontier model Mythos was first released only as a gated research preview (Project Glasswing) because of its cyber-vulnerability skills.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary"; "Business"; "Select Defined Terms".
Part 2
How a large language model is built and run, what a token is, how agents work, what compute means, who sits where in the value chain, how the money flows and who Anthropic competes with. Definitions follow the S-1's own glossary.
Part 2 · Frontier AI from first principles
An AI model trained mainly on large datasets to understand and generate natural language and computer code.
Reinforcement learning from human feedback: people rate the model's answers and those ratings become the reward it learns to maximise.
Training against written principles (Claude's "constitution"): be broadly safe, broadly ethical, compliant with guidelines, genuinely helpful. The model critiques and revises its own answers.
Anthropic says post-training has grown "from a minor adjustment step into one of the largest areas of our research investment", because models moved from chat assistants to agents that act on their own.
Source: Anthropic S-1 (26 Oct 2026), "Select Defined Terms"; "Business".
Part 2 · Frontier AI from first principles
The most a model can handle in one request, measured in tokens. It covers instructions, earlier conversation, user input and the model's output. (In the token row, · marks a space.)
The small units a model turns its input and output into: words, parts of words, characters or bytes. The count depends on the model and its tokenizer.
Illustrative token bill: one task, two models Illustrative, not Anthropic prices
| Model | Price per 1M tokens | Tokens used | Bill |
|---|---|---|---|
| Larger model | $10.00 | 40,000 | $0.40 |
| Smaller model | $1.00 | 40,000 | $0.04 |
Source: Anthropic S-1 (26 Oct 2026), "Select Defined Terms"; "MD&A"; "Business".
Part 2 · Frontier AI from first principles
The S-1 says agentic work uses "significantly more compute per user interaction" than a single chat reply.
Agentic AI. AI that plans, sequences and executes actions with some autonomy, using external tools, services and data on a user's behalf.
Tool use. The model calls outside functions and APIs during a task, several in parallel if needed.
Computer use. The model sees a screen and clicks, types and navigates apps the way a person would.
Interleaved thinking. The model reasons between tool calls, not only once at the start.
Multi-agent orchestration: one agent coordinates, others run subtasks in parallel
"A model alone is not an agent." The harness supplies the secure environment, credentials, memory and the loop itself. Anthropic offers it as the Agent SDK or as Claude Managed Agents.
Source: Anthropic S-1 (26 Oct 2026), "Select Defined Terms"; "Business"; "Risk Factors".
Part 2 · Frontier AI from first principles
What "compute" is made of (S-1 definition)
What Anthropic uses it for, and where the cost lands in its accounts
Training builds a model; inference runs it for users. Compute that gives free users access to Claude is booked in sales and marketing. The S-1 says training and serving Claude needs "enormous and increasing quantities" of compute.
Source: Anthropic S-1 (26 Oct 2026), "Select Defined Terms"; "Business"; "Risk Factors"; "Notes to Consolidated Financial Statements".
Part 2 · Frontier AI from first principles
Supply flows left to right; money flows right to left
Accelerator designers named: NVIDIA, AMD, Google (TPU), AWS (Trainium).
Most leading-edge chip making is in Taiwan, mainly at TSMC.
High bandwidth memory is a key part of AI accelerators.
Largely made by SK hynix, Samsung, Micron.
Grid or on-site power, transformers, cooling. Developers build and run sites.
Example named: Fluidstack.
Rent out compute; also resell Claude in their marketplaces.
AWS, GCP, Azure.
Train frontier models and sell access to them.
Anthropic. Rivals named: Google, OpenAI.
Claude Code, Cowork and products others build on Claude.
Examples: Legora, Salesforce. SIs: Accenture, Deloitte, KPMG, PwC.
Enterprises and prosumers (individuals who pay for AI tools).
11,900+ customers paying $100k+ (TTM, 30 Sep 2026).
A company that runs large-scale cloud infrastructure and rents out compute, storage and networking. The largest are called hyperscalers.
Accelerators and memory come from "a limited number of manufacturers". TSMC has at times been unable to meet demand, and Anthropic has not always obtained enough chips on time.
Source: Anthropic S-1 (26 Oct 2026), "Risk Factors"; "Business"; "Prospectus Summary"; "Select Defined Terms".
Part 2 · Frontier AI from first principles
Pay per use, metered in tokens, priced by model. The "substantial majority" of revenue.
A fee for a set level of usage, terms of 12 months or less. Heavy users can buy more.
Share of revenue by pricing model
Two ways to buy, one contract
Share of revenue sold through CSP marketplaces
Anthropic is the principal in marketplace sales and books the revenue gross.
Source: Anthropic S-1 (26 Oct 2026), "Select Defined Terms"; "MD&A"; "Notes to Consolidated Financial Statements". Pricing-model shares as reported in the S-1.
Part 2 · Frontier AI from first principles
Named as primary competitors: Google and OpenAI. Also large, well-capitalized tech companies investing to reach the frontier.
Large companies and specialised AI labs at or below the frontier, competing for work that values cost over intelligence.
Released by state actors and international and domestic research groups. May take some use cases and pressure prices.
AT THE CENTRE
How it says it competes:
Some large rivals, the S-1 notes, have more money, compute and brand recognition.
Bundling AI with other products, building it into tools customers already use, closed ecosystems, zero or negative margins.
As Anthropic builds its own apps, it may compete with enterprise software companies in select categories.
Legacy systems, entrenched workflows and change-management cost. Risk-averse buyers may simply not adopt.
Source: Anthropic S-1 (26 Oct 2026), "Business"; "Risk Factors".
Part 2 · Frontier AI from first principles
| Closed model (e.g. Claude) | Open-weight | Open-source | |
|---|---|---|---|
| Model weights | Held by the developer. The S-1 calls weights its "primary intellectual asset" | Publicly released | Publicly released |
| Training data and code | Kept in-house; S-1 cites proprietary training methods | Not necessarily shared | Shared |
| Who runs it | The developer, or its cloud partners | Anyone, under the licence | Anyone, under the licence |
| How used | Through an API, platform or apps | Download, run, modify | Download, run, modify |
| S-1's view | Anthropic says it competes on the most capable models for the hardest problems | "prioritize cost efficiency over maximum intelligence"; may pressure prices | |
The numbers that encode what a trained model has learned. The S-1 says they represent the accumulated investment in compute, data and research.
Distillation and extraction, as the S-1 describes the risk
Consequences the S-1 lists
As AI helps build AI, frontier know-how may spread through distillation, weight theft or AI-assisted research.
Source: Anthropic S-1 (26 Oct 2026), "Select Defined Terms"; "Risk Factors"; "Business".
Part 3
What Anthropic sells, how it is built in layers, who buys it, how fast accounts grow, and the routes Claude takes to reach customers.
Part 3 · Products and customers
A basic building block, like a Lego brick. Developers snap primitives together to make their own AI product.
Source: Anthropic S-1 (26 Oct 2026), "Business".
Part 3 · Products and customers
| Class | What the S-1 says it is for | Latest named in S-1 |
|---|---|---|
| Mythos (incl. Fable) | Most capable class. Frontier performance across all domains, especially coding and long-running agentic tasks. | Mythos 5.1 and Fable 5.1 (Sep 2026) |
| Opus | "Powerful workhorse" for reasoning-heavy tasks. Opus 5.5 is designed as the "enterprise daily driver". | Opus 5.5 (Sep 2026) |
| Sonnet | Versatile everyday class. Balances performance, cost and speed; used for high-volume sub-agents. | No version named |
| Haiku | Lowest cost and fastest. For high-volume work such as customer-service agents and bulk document processing. | No version named |
Mythos and Fable are one underlying model in two set-ups. Fable is generally available, with safeguards that block or limit sensitive topics. Mythos goes only to vetted cyber-security and life-science users through trusted-access programmes such as Project Glasswing (first released there in Apr 2026).
The capability-cost-speed curve (schematic, not to scale)
Source: Anthropic S-1 (26 Oct 2026), "Business" and "Prospectus Summary". Curve is a TheWrap schematic; the S-1 gives no scores or prices.
Part 3 · Products and customers
Tune how Claude thinks and acts
Plug in data, know-how and devices
Anthropic runs the serving
Turn a model into a working agent
Harness, in plain words: the safe workspace, passwords, memory and work loop a model needs to plan, use tools, check its work and finish a job alone. "A model alone is not an agent" (S-1).
Source: Anthropic S-1 (26 Oct 2026), "Business".
Part 3 · Products and customers
Software that does more than suggest the next line. It plans, writes, tests and fixes code across a large codebase, with the developer reviewing.
Released as a research preview (an early public test version).
Individual developers picked it up before company buyers had looked at the category.
Developers became internal champions. The S-1 says it is "becoming the default choice" in many enterprise engineering teams.
Non-technical teams are increasingly using Claude Code and Cowork for their own work, the S-1 says.
Source: Anthropic S-1 (26 Oct 2026), "Business" and "Prospectus Summary".
Part 3 · Products and customers
The user describes a task in plain language. No code or terminal.
Claude breaks the task into steps.
It works across local files, connected apps, Chrome and the desktop, clicking and typing where no integration exists.
It returns finished documents, spreadsheets, presentations and organised folders.
The model looks at the screen and clicks, types and moves between apps the way a person would.
The S-1 gives no Cowork user, usage or revenue figures, and does not say how it measures "adoption" in the Cowork vs Claude Code comparison.
Source: Anthropic S-1 (26 Oct 2026), "Business" and "Prospectus Summary"; "Select Defined Terms".
Part 3 · Products and customers
Flagship chat product on desktop, web and mobile. Files, projects, memory, web search, artifacts. The S-1 calls it "top-of-funnel" demand for enterprise sales.
Coding agent in the terminal, desktop, web and mobile, with GitHub and code-editor links.
Knowledge-work agent on desktop, web and mobile. Being merged into Claude.ai from Sep 2026.
Public beta for Claude Enterprise customers. Finds code vulnerabilities, rates them and proposes patches.
Tag Claude into a Slack channel or thread to delegate work. Replies with answers, drafts, analyses or code changes in the thread.
Designs, prototypes and presentations made by conversation and direct editing.
Research workbench. Queries 60+ scientific databases and returns publication-ready figures and manuscripts.
Claude for Microsoft 365 and Claude in Chrome sit inside tools knowledge workers already use.
Gives access to each product above with company controls: single sign-on and automatic user set-up (SAML, SCIM), role-based access, audit logs, a Compliance API, settings designed to support US health-data rules (HIPAA), and one admin console for IT oversight and cost visibility.
Source: Anthropic S-1 (26 Oct 2026), "Business".
Part 3 · Products and customers
Vertical solution: a version of Claude pre-loaded with one industry's knowledge, data connections and compliance controls, for fields where Anthropic says general tools are not enough.
Pre-built agents for pitch books, earnings reviews, know-your-customer checks and financial analysis. Runs inside Excel, PowerPoint, Word and Outlook.
Client-meeting prep, follow-ups, portfolio rebalance reviews, estate and tax briefs. Advisor review built into each workflow.
For providers, payers and health-tech firms. Clinical reasoning and medical coding; links to Medicare coverage data, ICD-10 codes and the NPI registry.
From early discovery to clinical trials and regulatory filings. Connects to lab notebooks, data stores and literature.
Pre-built legal agents linked to e-discovery (Relativity), contracting (Ironclad), research databases and deal rooms, with privilege controls.
Higher-education version with a student learning mode. Claude for Teachers gives verified US K-12 teachers free premium access.
US agencies use it in a FedRAMP High authorised setting. Non-US governments and classified work go through Amazon Bedrock.
Enterprise and Team plans; connectors and plugins on individual plans; the Claude Platform; and, for Education, institution-wide agreements.
Source: Anthropic S-1 (26 Oct 2026), "Business".
Part 3 · Products and customers
From AI-native start-ups to the world's largest organisations. They need single sign-on, audit logs, data-retention rules, role-based access and the option to run Claude inside their own cloud.
Software developers are the main adoption force: users, builders and internal champions at once.
Individual knowledge workers (researchers, developers, analysts, writers) who pay for Claude Pro or Claude Max subscriptions. The S-1 does not give plan prices.
Anthropic calls them a "top-of-funnel engine": people who try Claude themselves and then push it inside their employers.
A "customer" is a prosumer or an organisation. Anthropic does not track an "enterprise" category. It uses customers paying $100,000 or more over the trailing twelve months (TTM) as its measure of enterprise adoption.
Source: Anthropic S-1 (26 Oct 2026), "Business", "Select Defined Terms" and "Management's Discussion and Analysis".
Part 3 · Products and customers
Customers paying $100K+ over the trailing twelve months
Dates are unevenly spaced; each bar is a count the S-1 gives. Sep 2025 to Sep 2026: about 12.5 times.
Customers by annual spend, 30 Sep 2026
Revenue from $100K+ customers
| Period | Revenue | % of total |
|---|---|---|
| FY2023 | ~$37M | ~90% |
| FY2024 | $220M | ~57% |
| FY2025 | $3.5bn | 76% |
| TTM to 30 Sep 2026 | Not yet disclosed | Not yet disclosed |
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis" and "Business". Multiple computed by TheWrap from reported figures.
Part 3 · Products and customers
Pick the group. Every customer with at least $1 of revenue in the base period, excluding Claude Pro and Max subscribers (fixed-price plans).
Base period: the three months ending one year before the date. For ~900%, that is Jul to Sep 2025.
Current period: the three months ending on the date, Jul to Sep 2026.
Divide the group's current-period revenue by its base-period revenue. Growth, cut-backs and departures all count.
Customers already paying in Jul to Sep 2025 together paid about 9 times as much in Jul to Sep 2026, after netting off those who spent less or left. New customers are not in it.
Illustrative cohort (made-up round numbers, not Anthropic's)
Illustrative NRR = $500 ÷ $100 = 500%, even though customer A left.
Source: Anthropic S-1 (26 Oct 2026), "Select Defined Terms", "Prospectus Summary" and "Business". Illustration by TheWrap.
Part 3 · Products and customers
Share of $100K+ customers using the platform and at least one app
Accounts spread from one layer to two.
Average monthly spend per Claude Enterprise user, Jul–Sep 2026
Non-Code users spent under $75 a month on average; the top 5% of Code users about 20 times that.
Quarter-on-quarter revenue growth after Opus 4.5 (Nov 2025)
Opus 4.5 came with more capability and a lower price; Anthropic says several factors drove the jump.
Revenue for Jul–Sep 2026, by age of model version used
Anthropic says that with each release most customers move to the latest version.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis" and "Business".
Part 3 · Products and customers
Where adoption usually starts. One person sees hours of work done in minutes, then the team and the company follow.
Claude Code, Claude Cowork, Claude Enterprise
Claude moves from something staff consult to something the organisation runs on, as workflows, data and conventions are encoded into it.
Claude can suggest the next workflow to automate
The deepest level. Product teams build Claude into what they sell: customer-facing agents, AI features, new product lines.
Claude Platform
Named organisations
Production uses the S-1 cites
Inside the Fortune 500, $100K+ customers include
Source: Anthropic S-1 (26 Oct 2026), "Business" and "Prospectus Summary".
Part 3 · Products and customers
How the cloud joint go-to-market works
Anthropic says this lets it set prices, avoid channel conflict and keep the gains from cheaper compute.
Source: Anthropic S-1 (26 Oct 2026), "Business", "Prospectus Summary", "Management's Discussion and Analysis" and "Select Defined Terms".
Part 4
Revenue, costs, margins and the very large non-cash items that drive the net loss. Every figure is from the S-1 tables; the few calculations TheWrap made are labelled.
Part 4 · The numbers
H1 2026 covers six months only, yet it is 3.6 times full-year 2025. Revenue was $0 at commercial launch in Q1 2023.
Source: Anthropic S-1 (26 Oct 2026), "Summary Consolidated Financial and Other Data". H1 2026 vs FY2025 multiple computed by TheWrap.
Part 4 · The numbers
Q3 2026 revenue (three months to 30 September 2026) is left blank in this draft of the S-1.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis". Quarter-on-quarter growth (QoQ) computed by TheWrap.
Part 4 · The numbers
Customers pay per use. Usage is metered in tokens (small chunks of text or data) and priced by the model chosen.
A fixed fee for a set level of usage, generally for 12 months or less, booked evenly over the term.
| FY2023 | FY2024 | FY2025 | H1 2025 | H1 2026 | |
|---|---|---|---|---|---|
| Consumption | $39M | $288M | $3.8bn | $983M | $13.1bn |
| Subscription | $2M | $98M | $789M | $180M | $3.2bn |
Anthropic expects consumption to remain the substantial majority of revenue.
Source: Anthropic S-1 (26 Oct 2026), "Notes to Consolidated Financial Statements", Note 2 "Revenue"; "Management's Discussion and Analysis". Shares computed by TheWrap.
Part 4 · The numbers
| FY2023 | FY2024 | FY2025 | H1 2025 | H1 2026 | |
|---|---|---|---|---|---|
| United States | $34M | $263M | $3.4bn | $820M | $11.8bn |
| International | $7M | $123M | $1.2bn | $343M | $4.5bn |
By the billing address of the customer. Anthropic says it expects international expansion to be a meaningful driver of revenue growth.
Source: Anthropic S-1 (26 Oct 2026), "Notes to Consolidated Financial Statements", Note 2 "Revenue"; "Management's Discussion and Analysis". Shares computed by TheWrap.
Part 4 · The numbers
Customers above 10% of revenue (names not disclosed)
| Share of revenue | FY23 | FY24 | FY25 | H1 25 | H1 26 |
|---|---|---|---|---|---|
| Customer A | <10% | <10% | 12% | <10% | <10% |
| Customer B | <10% | <10% | 12% | 18% | <10% |
| Customer C | 43% | <10% | <10% | <10% | 11% |
| Customer D | 11% | <10% | <10% | <10% | <10% |
Share of revenue sold through CSP marketplaces
The online stores of cloud service providers (AWS, Google Cloud, Azure). They bill the customer and collect on Anthropic's behalf.
Source: Anthropic S-1 (26 Oct 2026), "Notes to Consolidated Financial Statements", Note 1 "Concentrations of Risk"; "Risk Factors"; "Management's Discussion and Analysis".
Part 4 · The numbers
| US$ millions | FY2023 | FY2024 | FY2025 | H1 2025 | H1 2026 |
|---|---|---|---|---|---|
| Revenue | 41 | 386 | 4,585 | 1,163 | 16,286 |
| Compute and infrastructure | 433 | 2,522 | 7,326 | 2,334 | 8,945 |
| Personnel | 95 | 454 | 1,839 | 538 | 3,152 |
| Bartz settlement | – | – | 1,500 | – | – |
| Sales, marketing, and partnerships | 2 | 43 | 568 | 120 | 3,168 |
| Other operating and administrative | 62 | 347 | 1,412 | 306 | 1,714 |
| Total operating expenses | 592 | 3,366 | 12,645 | 3,298 | 16,979 |
| Operating loss | (551) | (2,980) | (8,060) | (2,135) | (693) |
| Interest income | 69 | 106 | 431 | 115 | 1,111 |
| Change in fair value of convertible notes, derivatives, SAFEs | (1,045) | (5,434) | (34,308) | (10,942) | (72,388) |
| Other income (expense), net | (2) | – | (8) | – | (8,757) |
| Loss before income taxes | (1,529) | (8,308) | (41,945) | (12,962) | (80,727) |
| Provision for income taxes | 2 | 1 | 20 | 7 | 21 |
| Net loss | (1,531) | (8,309) | (41,965) | (12,969) | (80,748) |
Brackets are losses or expenses. H1 = six months to 30 June. Figures for H1 are unaudited.
Source: Anthropic S-1 (26 Oct 2026), "Summary Consolidated Financial and Other Data"; "Management's Discussion and Analysis", "Results of Operations".
Part 4 · The numbers
Compute and infrastructure
Personnel
Sales, marketing, and partnerships
Other operating and administrative
Each panel has its own scale. Total operating expenses were 1,444% of revenue in FY2023 and 104% in H1 2026.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis". Percentages as printed in the S-1.
Part 4 · The numbers
Compute and infrastructure expense by quarter, and as % of revenue
What the line contains
Most compute is bought 12 to 18 months ahead under multi-year deals. The cost arrives as capacity comes online, not as demand rises, so faster revenue growth shrinks the ratio.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis", "Components of Results of Operations" and "Quarterly Results of Operations".
Part 4 · The numbers
Revenue per $1 of compute and infrastructure expense, by quarter
Calculated by TheWrap: quarterly revenue ÷ quarterly compute and infrastructure expense.
Three drivers Anthropic names
Running a trained model to answer customer requests, as opposed to training it.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis", "Compute Efficiency". Ratios computed by TheWrap from reported figures.
Part 4 · The numbers
Pay in the form of shares or options. It is a real expense in the accounts but uses no cash when granted.
| US$ millions | FY23 | FY24 | FY25 | H1 25 | H1 26 |
|---|---|---|---|---|---|
| Personnel | 95 | 454 | 1,839 | 538 | 3,152 |
| of which SBC | 29 | 192 | 1,165 | 286 | 2,185 |
| SBC share | 31% | 42% | 63% | 53% | 69% |
H1 2026 personnel rose $2.6bn: $1.9bn more SBC and $510M more cash pay, both from headcount growth. Anthropic expects extra SBC at the IPO for RSUs whose performance condition is met then.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis", "Results of Operations" and "Quarterly Results of Operations". Ratios computed by TheWrap from reported figures.
Part 4 · The numbers
Sales, marketing, and partnerships expense by quarter, and as % of revenue
What the line contains
| Increase drivers | FY2025 | H1 2026 |
|---|---|---|
| Channel partner fees | +$341M | +$2.7bn |
| Marketing and advertising | +$170M | +$273M |
| Total line, growth | +1,221% | +2,540% |
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis".
Part 4 · The numbers
Other operating and administrative expense by quarter
Covers donation-matching charges, data-quality and evaluation vendors, IT and cloud software, payment processing, offices, professional services. Q4 2025 included $541M of non-cash donation-matching charges.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis"; "Notes to Consolidated Financial Statements", Note 13 "Commitments and Contingencies".
Part 4 · The numbers
TTM revenue per employee
2026 figure is "over $6.6 million".
The total for the last twelve months up to a date, which smooths out single-quarter swings.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis", "Operating Efficiency"; "Risk Factors".
Part 4 · The numbers
Operating income or loss divided by revenue. Q2 2026 is the only quarter shown with operating income: $974M.
It says margins may compress in periods when large new compute capacity arrives, and expects more short-term compute to compress them in the near term.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis", "Quarterly Adjusted Operating Income (Loss) and Adjusted Operating Margin".
Part 4 · The numbers
H1 2026 bridge, US$
Operating loss with three items added back: stock-based compensation, non-cash charitable contributions (share-based donation matching) and the Bartz settlement. Non-GAAP means it is not defined by US accounting rules.
| US$ millions | Adj. op. income | Adj. margin |
|---|---|---|
| FY2023 | (522) | (1,273)% |
| FY2024 | (2,788) | (722)% |
| FY2025 | (4,854) | (106)% |
| H1 2025 | (1,849) | (159)% |
| H1 2026 | 1,618 | 10% |
FY2025 bridge: −$8.1bn + $1.2bn SBC + $541M charitable + $1.5bn Bartz = −$4.9bn.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis", "Non-GAAP Financial Measures".
Part 4 · The numbers
H1 2026, from operating loss to net loss
Anthropic borrowed through convertible notes (loans that can turn into shares) and earlier used SAFEs (agreements to receive shares later). They are carried at fair value and re-valued every quarter. When Anthropic's own valuation rises, the shares they convert into are worth more, so the liability rises and the increase is booked as an expense. It is a non-cash item: the cash flow statement adds it back.
The −$8.8bn of other expense is mainly the fair value of 2026 note purchase agreements with Google and Amazon, broken down two slides on.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis"; "Notes to Consolidated Financial Statements", Note 4 "Fair Value Measurements"; "Consolidated Statements of Cash Flows".
Part 4 · The numbers
H1 2026 bridge, US$
Net loss excluding SBC, non-cash charitable contributions, the Bartz settlement, the fair-value change of notes, derivatives and SAFEs, and the fair-value change of certain mark-to-market instruments.
| US$ millions | Net loss | Adj. net income |
|---|---|---|
| FY2023 | (1,531) | (457) |
| FY2024 | (8,309) | (2,683) |
| FY2025 | (41,965) | (4,448) |
| H1 2025 | (12,969) | (1,741) |
| H1 2026 | (80,748) | 2,671 |
The $8.7bn mark-to-market item in H1 2026 largely offsets the $8.8bn other expense.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis", "Non-GAAP Financial Measures".
Part 4 · The numbers
Interest income by quarter
Cash, cash equivalents and short-term investments
What the −$8.8bn of other income (expense) is
| H1 2026, US$ millions | Amount |
|---|---|
| Recognition and change in fair value of note purchase agreements | (9,405) |
| Change in fair value of warrant rights | 724 |
| Change in fair value of holdback liability | (39) |
| Other | (37) |
| Other income (expense), net | (8,757) |
Deals signed with Google (March 2026) and Amazon (April 2026) that included commitments to sell them Series G-2 preferred shares at $259.14 each. They are re-valued every period and the change goes through this line.
Interest income comes from cash, Treasuries, corporate debt and commercial paper. $112.4bn was cash and equivalents and $19.3bn short-term investments at 30 Jun 2026.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis"; "Consolidated Balance Sheets"; "Notes to Consolidated Financial Statements", Notes 4 and 8.
Part 4 · The numbers
| US$ millions, three months to | Q1 2025 | Q2 2025 | Q3 2025 | Q4 2025 | Q1 2026 | Q2 2026 |
|---|---|---|---|---|---|---|
| Revenue | 376 | 787 | 1,382 | 2,040 | 4,731 | 11,555 |
| Compute and infrastructure | 944 | 1,390 | 2,047 | 2,945 | 3,387 | 5,558 |
| Personnel | 218 | 320 | 489 | 812 | 1,295 | 1,857 |
| Bartz settlement | – | – | 1,500 | – | – | – |
| Sales, marketing, and partnerships | 24 | 96 | 202 | 246 | 1,033 | 2,135 |
| Other operating and administrative | 125 | 181 | 233 | 873 | 683 | 1,031 |
| Operating income (loss) | (935) | (1,200) | (3,089) | (2,836) | (1,667) | 974 |
| Interest income | 51 | 64 | 124 | 192 | 349 | 762 |
| Fair-value change, notes, derivatives, SAFEs | (5,701) | (5,241) | (9,165) | (14,201) | (12,791) | (59,597) |
| Other income (expense), net | – | – | (3) | (5) | (309) | (8,448) |
| Net loss | (6,592) | (6,377) | (12,138) | (16,858) | (14,439) | (66,309) |
| Adjusted operating income (loss) | (830) | (1,019) | (1,284) | (1,721) | (617) | 2,235 |
| Adjusted operating margin | (221)% | (129)% | (93)% | (84)% | (13)% | 19% |
| Adjusted net income (loss) | (786) | (955) | (1,168) | (1,539) | (310) | 2,981 |
Q3 2026 figures are not yet disclosed in this draft. Quarterly figures are unaudited.
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis", "Quarterly Results of Operations".
Part 4 · The numbers
Illustrative calculation
Consumption fees in the trailing 28 days × 13, plus aggregate monthly subscription fees on the measurement date × 12. It annualises a short window of activity.
The S-1 states Run-Rate Revenue "as of October 31, 2026" but the value is blank in this draft.
The S-1's own limits on the metric
Source: Anthropic S-1 (26 Oct 2026), "Select Defined Terms"; "Risk Factors"; "Business"; "Management's Discussion and Analysis".
Part 4 · The numbers
Revenue in the latest three months from customers who were already paying a year earlier, divided by what those same customers paid in the same three months a year earlier. Claude Pro and Max subscribers are excluded. ~900% means they now spend about nine times as much.
Source: Anthropic S-1 (26 Oct 2026), "Summary Consolidated Financial and Other Data"; "Management's Discussion and Analysis"; "Select Defined Terms".
Part 5
How Anthropic buys compute, who it buys it from, what it has committed to pay, and how it has funded all of it.
Part 5 · Compute and capital
Interchangeable. Any chip in the pool can be switched between training, research, customers and internal use, like cash that can pay any bill.
A research finding that gets more model capability out of the same amount of compute. Anthropic says its research produces proprietary ones.
Source: Anthropic S-1 (26 Oct 2026), "Business: Our Approach to Compute" and "Our Infrastructure: Multi-Chip Compute Approach".
Part 5 · Compute and capital
Compute and infrastructure expense by quarter, US$
Source: Anthropic S-1 (26 Oct 2026), "Management's Discussion and Analysis": Compute Investments, Results of Operations, Quarterly Results, Cash Flows. Ratios computed by TheWrap from reported figures.
Part 5 · Compute and capital
Anthropic says data centres take longest to build but are a small part of total compute cost, so it books sites early and commits to the costlier chips later.
Power generated at the site itself, bypassing the public grid.
Source: Anthropic S-1 (26 Oct 2026), "Business: Our Data Center Strategy", "Facilities", "Risk Factors" (power), Notes 13 and 16 to the financial statements.
Part 5 · Compute and capital
| Agreement | What it is | Size |
|---|---|---|
| Compute deal, Apr 2026 | Must spend at least this, May 2026 to Apr 2036, regardless of usage | $110.0bn |
| Amazon 2023 NPA | Notes, 2 tranches: $1.2bn (Sep 2023), $3.4bn (Mar 2024, for $2.8bn cash); 8% | up to $4.0bn |
| Amazon 2024 NPA | Notes, 3 tranches: $1.3bn (Nov 2024), $1.3bn (May 2025), $1.4bn (Nov 2025) | up to $4.0bn |
| Amazon 2026 NPA | $5.0bn of preferred bought in Q2 2026; rest drawable in $5.0bn steps at compute milestones, until 30 months after the IPO | $20bn, $15bn left |
A contract where an investor agrees to lend by buying notes that later turn into shares instead of being repaid in cash. Notes are issued in instalments called tranches.
Amazon notes outstanding at 30 Jun 2026
| Note | Principal | Fixed price | Class N shares |
|---|---|---|---|
| Amazon 2023 NPA | $2,186M | $2.0928 | 1,044.7M |
| Amazon 2024 NPA | $4,017M | $5.0478 | 795.7M |
| Total | $6,203M | 1,840.4M |
The notes convert into non-voting Class N shares at the lower of the fixed price and the IPO price, up to the 12% cap. Any excess is reissued as a new rollover note and converts at each quarter end at the lowest of fixed price, IPO price and 10-day average market price.
Source: Anthropic S-1 (26 Oct 2026), "MD&A: Liquidity and Capital Resources: Convertible Notes (Amazon, Outstanding Notes)", "Infrastructure Service Arrangements", Notes 9 and 16.
Part 5 · Compute and capital
| Facility | Terms | Outcome |
|---|---|---|
| Google 2023 NPA, up to $2.0bn | 3 tranches; 10.0% interest; 48-month maturity | All converted: Series E-3, E-4 (2025), G-4 (Jan 2026, at $233.22/share) |
| Google 2024 NPA, ~$1.3bn | 8.5% interest; due Aug 2030 | Converted 2025 into 24.2M Series E-5 shares at $53.28 |
| Google 2026 NPA, $20.0bn left | Draws tied to third-party compute milestones and revenue targets; each funded over 5 quarterly closings; 14.99% ownership cap | Nothing drawn at 30 Jun 2026. Post-IPO, Google takes notes on agreed terms or Class N shares at 30-day average price |
| IPO option, $10.0bn | Google may buy Class N shares at the IPO price in a concurrent private placement | If used, the 2026 NPA shrinks by the same amount |
Source: Anthropic S-1 (26 Oct 2026), "MD&A: Convertible Notes (Google)", "Infrastructure Service Arrangements", Notes 9 and 16.
Part 5 · Compute and capital
A lease treated like buying on credit: the asset and the debt both go on the balance sheet.
Source: Anthropic S-1 (26 Oct 2026), "MD&A: Infrastructure Equipment Leases", "Convertible Notes (Broadcom)", Notes 7 and 9.
Part 5 · Compute and capital
Source: Anthropic S-1 (26 Oct 2026), "MD&A: AMD Agreement", "Infrastructure Service Arrangements", Note 17.
Part 5 · Compute and capital
Non-cancellable purchase commitments by year, as of 31 Dec 2025 (total $54.6bn)
Disclosed aggregates, US$ (overlapping snapshots, not additive)
Plus an estimated $125bn to $150bn for extra TPU compute in 2028 to 2031, and $21bn of short-term deals after 30 Jun 2026. Lease liabilities on the 30 Jun 2026 balance sheet: $1.5bn.
The MD&A table of obligations by year (2026 to 2030 and later), its total, the amount due in the next 12 months and the range for variable-price deals are all left blank.
Source: Anthropic S-1 (26 Oct 2026), "MD&A: Contractual Obligations", Notes 7, 13, 16 and 17 to the financial statements. Lease liability sum computed by TheWrap.
Part 5 · Compute and capital
A top-up payment owed if actual spend falls below the agreed minimum.
Largest named arrangements, total over the contract, US$
Spans differ: SpaceXAI into 2029, Microsoft and Google to 2033, Lambda to 2034, Amazon to 2036. Several can be cancelled after initial periods or if a provider misses delivery. Not additive.
Source: Anthropic S-1 (26 Oct 2026), "MD&A: Contractual Obligations: Infrastructure Service Arrangements and Infrastructure Equipment Leases".
Part 5 · Compute and capital
Preferred stock sold for cash, by round, US$
Series A and B dates are not given in the S-1; amounts are carrying values. Series G includes $10.0bn from Google and $5.0bn from Amazon.
| Other instruments | Amount |
|---|---|
| Amazon convertible notes, 2023 to 2025 | $8.6bn |
| Google convertible notes, 2023 to 2025 | $3.3bn |
| SAFEs, 2023 (all now preferred) | $122M |
Simple agreement for future equity: cash now for shares at the next priced round, with no interest or maturity.
All Google notes and part of the Amazon 2023 notes later turned into non-voting preferred series (E-2 to E-5, F-2, G-3, G-4). $6.2bn of Amazon notes were still outstanding at 30 Jun 2026. Principal at face value.
Source: Anthropic S-1 (26 Oct 2026), "Prospectus Summary", Notes 9, 10 and 16, "Recent Sales of Unregistered Securities".
Part 5 · Compute and capital
Revolving credit commitments, US$
A bank credit line the company can draw, repay and draw again, like a corporate credit card.
| Original facility | A&R facility (Sep 2026) | |
|---|---|---|
| Size | $2.5bn, raised to $3.25bn | $15.0bn |
| Letter-of-credit limit | $400M, raised to $1.0bn | Full $15.0bn |
| Matures | May 2030 | 4 Sep 2031 |
| Minimum liquidity | $500M | $5.0bn from Q4 2026 |
| Drawn, 30 Jun 2026 | Nil; $78M letters of credit | n/a |
Senior secured on substantially all assets; security falls away if two of Moody's, S&P and Fitch rate it investment grade.
Other financing facts to the filing date
Source: Anthropic S-1 (26 Oct 2026), "MD&A: Revolving Credit Agreement", Notes 9 and 17, "Recent Sales of Unregistered Securities".
Part 5 · Compute and capital
Operating cash flow, US$
Turned positive in H1 2026 ($3.9bn), on a smaller loss after non-cash items and timing of vendor payments and customer receipts. The FY2024 outflow reflected prepaid hosting capacity.
Investing cash flow, US$
Mostly buying short-term investments ($19.0bn in H1 2026). Servers, networking and other equipment: $482M in H1 2026.
Financing cash flow, US$
Preferred stock sales: $110.4bn in H1 2026, $16.5bn in FY2025. Convertible notes for cash: $1.75bn, $6.05bn, $3.45bn in 2023 to 2025.
Source: Anthropic S-1 (26 Oct 2026), "MD&A: Cash Flows", Consolidated Statements of Cash Flows. Each chart has its own scale.
Part 5 · Compute and capital
| US$ millions | 31 Dec 24 | 31 Dec 25 | 30 Jun 26 |
|---|---|---|---|
| Cash and short-term investments | 3,141 | 20,277 | 131,651 |
| Total assets | 7,752 | 24,661 | 147,358 |
| Convertible notes (fair value) | 14,116 | 41,677 | 108,566 |
| Total liabilities | 14,679 | 45,397 | 117,467 |
| Redeemable convertible preferred | 2,874 | 29,769 | 157,510 |
| Accumulated deficit | (10,040) | (52,154) | (132,902) |
| Total stockholders' deficit | (9,801) | (50,505) | (127,619) |
All preferred stock converts into common stock, and the Amazon notes convert into Class N shares up to the 12% cap. Both stop being liabilities or preferred claims and become common equity. The pro forma figures are blank in this draft. Of the $132.9bn accumulated deficit, $113.2bn relates to fair-value changes in convertible notes and derivative liabilities as the valuation rose.
30 Jun 2026, to scale
Source: Anthropic S-1 (26 Oct 2026), Consolidated Balance Sheets, "Summary Consolidated Financial Data", "Capitalization", "MD&A: Liquidity and Capital Resources".
Part 6
Who runs Anthropic, how its research becomes product, the safety rules it says it follows, and the incidents, regulation and lawsuits the S-1 discloses.
Part 6 · People, research and safety
All seven co-founded Anthropic in January 2021. The S-1 says all seven remain active leaders, and each has pledged 80% of personal equity in Anthropic to philanthropy.
RLHF trains a model using human ratings of its answers. Scaling laws describe how performance improves predictably with model size, data and compute.
The seven together form the Founder LLC that votes the Class F shares (Part 7). Benjamin Mann is often called a co-founder, but the S-1 says he is not a member of the Founder LLC.
Source: Anthropic S-1 (26 Oct 2026), "Business: Our Culture, Values, and Talent" and "Our Governance Structure and Its Goals".
Part 6 · People, research and safety
Elected by Class T (the Long-Term Benefit Trust): 4 seats
Elected by Class A and Class F voting together: 3 seats
Not a director
A financially disinterested trust that holds the Class T share and generally elects a majority of the board.
Board after the IPO: seven authorised seats. Hastings, Liddell, Narasimhan and Razavi are independent under Nasdaq rules. Because the trust can elect a majority, Anthropic is a Nasdaq "controlled company". Dario and Daniela Amodei are siblings. Ages as of 30 Sep 2026.
Source: Anthropic S-1 (26 Oct 2026), "Management".
Part 6 · People, research and safety
Employees, approximate
A majority of staff work in research, engineering and product. Headquarters are in San Francisco, with offices in Europe and Asia-Pacific. Staff are called "Ants". The S-1 says it hires from fields such as philosophy, journalism and education, and calls its approach "talent density".
The seven values the S-1 lists
Trailing twelve months: the total for the latest 12-month period.
Source: Anthropic S-1 (26 Oct 2026), "Business: Our Culture, Values, and Talent", "Facilities", "Management’s Discussion and Analysis" and "Risk Factors".
Part 6 · People, research and safety
The S-1 gives three reasons: large economic impact, safety problems show up there, and outputs are largely verifiable.
Anthropic says it does not currently focus on them. That compute goes to coding, agentic skills and safety research.
Pre-training teaches a model general patterns from huge amounts of data. Post-training shapes that model into a useful, safe assistant.
Source: Anthropic S-1 (26 Oct 2026), "Business: Research Lab Identity and the Research-to-Product Pipeline".
Part 6 · People, research and safety
Anthropic’s written rules for catastrophic risk. Safety tests are tied to capability thresholds, so more capable models face stronger checks.
What the policy requires, per the S-1
First published Sep 2023. The S-1 says two other leading labs adopted similar frameworks within months. This draft does not list numbered safety levels.
Responses the S-1 describes, from lightest to strongest
Source: Anthropic S-1 (26 Oct 2026), "Business: Safety Research is Core to Our Business", "Our Approach to Model Releases", "Race to the Top"; "Risk Factors".
Part 6 · People, research and safety
The four priorities, in the order the S-1 lists them
A written document describing how Claude should think and behave. Anthropic uses it as training material.
How the S-1 says it is used
Constitutional AI is one of the advances the S-1 says its researchers originated. Co-founder Jared Kaplan helped pioneer it.
Source: Anthropic S-1 (26 Oct 2026), "Business: Claude’s Constitution", "Our Talent"; "Our Governance Structure and Its Goals".
Part 6 · People, research and safety
Does the model do what we intend?
What is happening inside the model?
How is misuse stopped in use?
The science of how large language models work inside, studied the way a neuroscientist studies a brain.
Every model also goes through multi-stage evaluations documented in public system cards. Anthropic holds ISO/IEC 42001 certification and gives pre-release access to the US CAISSI and the UK AI Security Institute.
Source: Anthropic S-1 (26 Oct 2026), "Business: Alignment", "Interpretability", "Safeguards", "Model Validation and Training Data".
Part 6 · People, research and safety
A Frontier Red Team checks whether the model gives attackers meaningful new capabilities. A system card reports evaluations and known weaknesses.
The US CAISSI and the UK AI Security Institute evaluate certain models before release. Some government partners get expanded early access.
Anthropic says it wants labs to compete on safety. It shares findings through the Frontier Model Forum and calls for "pacing the frontier" through industry standards.
The learned numbers that make up a trained model. Whoever holds them can run the model.
Source: Anthropic S-1 (26 Oct 2026), "Business: Our Approach to Model Releases", "Race to the Top"; "Risk Factors".
Part 6 · People, research and safety
The S-1 states: Anthropic believes this is the first documented large-scale cyberattack executed without substantial human intervention.
The S-1 states: Affected parties were notified. No evidence of coordination, goals beyond the task, or evading oversight.
The S-1 states: This awareness "creates a significant limitation on our ability to assess model safety."
A model acting differently because it can tell it is in a test, much like a driver who slows down only when a police car is visible.
Source: Anthropic S-1 (26 Oct 2026), "Risk Factors".
Part 6 · People, research and safety
Export controls
A legal order from a regulator requiring a company to hand over documents or answers during an investigation.
Source: Anthropic S-1 (26 Oct 2026), "Business: Government Regulation", "Export Controls and Trade Regulations"; "Risk Factors".
Part 6 · People, research and safety
| Case | Filed | Claim | Status at filing of the S-1 | |
|---|---|---|---|---|
| Bartz v. Anthropic (N.D. Cal.) | Aug 2024 | Authors: copyright in books used for training (class action) | Settled for $1.5bn (agreed Sep 2025, approved Jul 2026). $300M paid in 2025; $1.2bn owed at 30 Jun 2026 | Settled |
| Bartz opt-out suits | After settlement | Authors and publishers who left the class sue separately | Various stages | Open |
| Concord I (now N.D. Cal.) | Oct 2023 | Music publishers: copying of song lyrics; removal of copyright information | Both sides sought summary judgment; hearing 21 Oct 2026 | Open |
| Music Publisher Matters | Jan-Oct 2026 | Universal, Sony, Warner Chappell, BMG and others: lyrics and compositions; some name two executives | Preliminary stages | Open |
| Four Canadian actions | May 2025 to Oct 2026 | Author class actions (BC, Quebec), La Presse, news publishers | Quebec author case stayed pending the BC case | Open |
| Reddit v. Anthropic (Cal. state) | Jun 2025 | Use of Reddit data; breach of contract | Narrowed in Sep 2026 to contract, interference and unfair competition claims | Open |
| Consumer class actions | Not given | Alleged deceptive advertising of Claude subscription plans | Early stages | Open |
| Anthropic v. Department of War | Not given (actions Feb-Mar 2026) | Federal ban and "supply-chain risk" label after Anthropic opposed use for mass domestic surveillance and fully autonomous weapons | N.D. Cal. ruled largely for Anthropic (27 Aug 2026). D.C. Cir. upheld the Section 4713 label 2-1 (25 Sep 2026); rehearing sought | Split |
Anthropic says it disputes the allegations and has meritorious defences in each pending matter.
Source: Anthropic S-1 (26 Oct 2026), "Business: Legal Proceedings", "Other Matters"; Note 13 to the consolidated financial statements.
Part 7
What is being sold and to whom, the public benefit corporation, five share classes, the Long-Term Benefit Trust, founder voting control, philanthropy and executive pay.
Part 7 · The IPO and who controls Anthropic
| Listing | Nasdaq Global Select Market, symbol "ANT" |
| Security sold | Class A common stock (one vote per share) |
| Who sells | Anthropic (new shares) and selling stockholders (existing shares) |
| Price range | Blank in this draft |
| Shares offered | Blank in this draft |
| Net proceeds | Blank in this draft |
| Extra-share option | Underwriters may buy more Class A shares from Anthropic for 30 days. Size: blank |
| Stock split | 10-for-1 forward split of Class A and Class N on 17 Sep 2026 |
Anthropic keeps the cash only from the new shares it sells. Cash for existing shares sold by selling stockholders goes to those holders.
22 underwriters; Goldman Sachs and Morgan Stanley are the representatives
General corporate purposes: working capital, operating costs, capital spending including compute and infrastructure, and hiring. Part pays tax withholding on RSUs settled at the IPO (amount blank). No material acquisition is agreed.
Co-founders: up to 365 days (or the Q3 2027 results release, if earlier). Non-founder directors, officers and some holders: up to 180 days (or the Q1 2027 results release). Some holders: 365 days. Shares unlock in stages tied to results releases; share counts are blank.
Source: Anthropic S-1 (26 Oct 2026), cover page, "The Offering", "Use of Proceeds", "Shares Eligible for Future Sale", "Underwriting".
Part 7 · The IPO and who controls Anthropic
A company form under Delaware law whose directors must weigh a stated public benefit and affected people alongside shareholders' financial returns.
Source: Anthropic S-1 (26 Oct 2026), cover page, "Our Governance Structure and Its Goals: Public Benefit Corporation".
Part 7 · The IPO and who controls Anthropic
| Class | Held by | Votes | Economic rights | Transfer / conversion | Redeemable for nothing |
|---|---|---|---|---|---|
| Class A | New public buyers and existing holders, incl. founders and employees | 1 vote per share All Class A together capped at 49.9% of votes while Class F exists | Yes dividends and liquidation, pro rata | Freely transferable (after lock-up); not convertible | No |
| Class N | Mainly strategic business partners; also the Public Benefit Fund and donation matches | No vote except where the law requires | Yes same as Class A | Transferable; becomes Class A when sold publicly or to an unaffiliated buyer | No |
| Class T | Long-Term Benefit Trust (1 share) | Elects the Class T directors a majority of the board; no general vote | None par value only | Only with Trust Amendment Threshold approval | Yes by that threshold |
| Class F | Founder LLC of the 7 co-founders (1 share) | Fixed 50.1% of all votes on general matters, incl. non-Class T directors | None par value only | Not transferable | Yes on sunset events |
| Class E | Employees, excluding co-founders (0 shares at IPO) | 1 vote per share only on Specified Class T Director disputes and LTBT changes | None par value only | Not transferable | Yes on leaving, or when Class F ends |
Only Class A and Class N share in dividends and sale proceeds. Classes T, F and E carry control rights but receive only par value ($0.00001 per share) in a liquidation.
Source: Anthropic S-1 (26 Oct 2026), "Description of Capital Stock", "The Offering".
Part 7 · The IPO and who controls Anthropic
For one Class T seat, LTBT nomination or removal without cause needs the Nominating and Governance Committee (NGC) to agree. If it does not within 15 days, Class E employees vote.
With 50.1% of votes, the Founder LLC can decide general stockholder votes, including the 3 non-Class T seats, mergers and a sale of the company. Its vote is not diluted by new share issues.
Because the LTBT elects a board majority, Nasdaq treats Anthropic as a "controlled company". Anthropic says it will use this exemption only for its NGC, which will not be fully independent.
Source: Anthropic S-1 (26 Oct 2026), "Our Governance Structure and Its Goals", "Management: Board of Directors", "Description of Capital Stock".
Part 7 · The IPO and who controls Anthropic
Trustees (selected for AI safety, security, policy, social enterprise expertise)
A Delaware "noncharitable purpose trust" that holds the single Class T share. It picks directors but has no role in day-to-day management.
Anthropic's payments to the LTBT ($ millions)
Source: Anthropic S-1 (26 Oct 2026), "Our Governance Structure and Its Goals: Long-Term Benefit Trust", "Certain Relationships and Related Person Transactions".
Part 7 · The IPO and who controls Anthropic
When does founder control (Class F) end?
resign, die or become incapacitated, hold under 2% of their IPO-date Class A stake, are removed for cause (incl. competing), or take an approved leave.
A supermajority of Continuing Founders redeems it, or the Founder LLC dissolves. Once triggered, a sunset cannot be extended or waived.
1 share per six months worked, plus 1 per Class A share pledged via donation matching. Half redeemed each year after leaving; all when Class F ends.
Never automatic. Ends only if the Trust Amendment Threshold approves redeeming or eliminating Class T.
The LTBT still elects the Class T majority. Class A elects the other directors and decides other votes.
Class A and F elect all directors, so the Founder LLC's 50.1% decides. If both end, Class A decides everything.
Source: Anthropic S-1 (26 Oct 2026), "Our Governance Structure and Its Goals: Founder Voting Structure, Class E Common Stock, Sunset Provisions".
Part 7 · The IPO and who controls Anthropic
Class A shares beneficially owned at 30 Sep 2026 (millions, incl. options and RSUs vesting within 60 days)
Total shares outstanding, every ownership percentage, pre- and post-IPO voting power of each holder, and shares each seller will sell. Each Class A holder's voting power is cut to 49.9% of its pre-IPO percentage.
Source: Anthropic S-1 (26 Oct 2026), "Principal and Selling Stockholders". Combined co-founder total summed by TheWrap from reported figures.
Part 7 · The IPO and who controls Anthropic
The fund may donate or sell the shares for public benefit work. The board decides allocations for now. Initial managers: Dario Amodei, Daniela Amodei, Jack Clark and LTBT Chair Neil Shah.
Each 50% share-price gain may bring another issue of up to 2%
Illustrative: invented $100 IPO price, not Anthropic's
Board discretion. A milestone counts when the close is above it on 90 of the last 120 trading days, and the 10-day average before issue is also above it.
Each co-founder pledged 80% of personal Anthropic equity to philanthropy. The S-1 calls these voluntary and revocable, not legally binding, and says they do not affect Class F voting.
Employees pledge equity to charity and Anthropic matches in Class N shares: 30.6M shares issued in Q1 2026 and about 28.0M in Oct 2026. Q3 2026 expense for the August match: $2.0bn.
Source: Anthropic S-1 (26 Oct 2026), "MD&A" and "Business: Our Commitment to Public Benefit", "Risk Factors", Notes to financial statements.
Part 7 · The IPO and who controls Anthropic
Named executive officer pay, FY2025 (grant-date value of equity)
| FY2025 | Salary | Stock | Options | Other | Total |
|---|---|---|---|---|---|
| Dario Amodei CEO, co-founder | $0.65M | $9.26M | $4.03M | $4.01M | $17.96M |
| Daniela Amodei President, co-founder | $0.65M | $9.26M | $4.03M | $2.46M | $16.41M |
| Krishna Rao CFO | $0.71M | None | None | $0.01M | $0.72M |
Main related-party dealings since 2023
| Spark Capital | Fund of director Yasmin Razavi bought $777M of Series C to H preferred stock, 2023 to 2026. |
| Bodhi LLC | Owned by director Vas Narasimhan and his wife: $10.0M of Series H (2026). |
| Tender offers | Amodeis sold shares in Jun 2025 ($5.61/share) and, with Rao and Holden Karnofsky, in Apr 2026 ($25.9136/share). |
| Family employment | Holden Karnofsky, Daniela Amodei's spouse, employee since Jan 2025: $430,213 cash in 2025. |
| Voting proxies | Dario Amodei votes shares held by two donor-advised funds. |
| Stockholders' Agreement | LTBT and Founder LLC agree to vote for the board setup. Anthropic must give the LTBT advance notice of a liquidation, a sale of 50%+ of votes or economics, or a transfer of control of an AI model above board-set capability thresholds. |
| LTBT and Founder LLC | Anthropic funds the LTBT ($8.4M in 9M 2026), reimburses both entities' costs and pays Founder LLC legal fees (amount blank). |
Source: Anthropic S-1 (26 Oct 2026), "Executive Compensation", "Certain Relationships and Related Person Transactions". Spark total summed by TheWrap from reported figures.
Summary
Anthropic says “Powerful AI” will be widely available within a few years, and that each Claude model now helps build the next.
Claude models, the Claude Platform and applications led by Claude Code and Claude Cowork, mostly paid per token used.
Revenue went from $41M in 2023 to $4.6bn in 2025 and $16.3bn in H1 2026. 11,900+ customers pay $100K+ a year.
Adjusted operating margin reached +10% in H1 2026. The $80.7bn GAAP net loss is mostly a non-cash fair-value charge.
Four chip families, multi-year take-or-pay contracts, $141.1bn raised and $131.7bn of cash and short-term investments.
A public benefit corporation. After the IPO the co-founders’ Founder LLC holds a fixed 50.1% of votes; a financially disinterested trust elects most of the board.
Source: Anthropic S-1 (26 Oct 2026). Summary by TheWrap of the preceding slides.
Not yet disclosed
| Area | Left blank in this draft |
|---|---|
| Offering | Price range per share, number of Class A shares offered by the company and by selling stockholders, underwriters’ option, gross and net proceeds, underwriting discount |
| Share count | Class A and Class N shares outstanding after the offering, shares issued in the Amazon Notes conversion and the RSU settlement |
| Recent revenue | Revenue for the three months to 30 Sep 2026 and the Run-Rate Revenue value |
| Commitments | Aggregate committed spend on compute and infrastructure, the amount expected before 31 Dec 2030 and the amount due in the next 12 months |
| Pro forma | Pro forma balance sheet and pro forma net loss per share after the preferred stock and note conversions |
A first S-1 is filed before the price is set. Later amendments fill in these numbers before the shares are priced.
Source: Anthropic S-1 (26 Oct 2026), cover page, "The Offering", "Prospectus Summary", "Capitalization".
Glossary
| Token | The unit a model reads and writes; usage is metered and billed in tokens. |
| Context window | How many tokens a model can process in one request, including its own output. |
| Pre-training | The first training stage: the model learns patterns by predicting the next token. |
| Post-training | Refining a pre-trained model with reinforcement learning and alignment methods. |
| Constitutional AI | Training against written principles (Claude’s constitution) the model uses to critique itself. |
| Inference | Running a trained model to produce an output, every time someone uses it. |
| Compute | Chips, memory, networking, storage and power used to train and run models. |
| Compute multiplier | An algorithmic improvement that gets more output from the same compute. |
| Agentic AI | AI that plans and carries out multi-step tasks, using tools, with some autonomy. |
| Net Revenue Retention | Last three months’ revenue from customers who paid a year earlier (excl. Pro and Max), divided by their revenue then. |
| Run-Rate Revenue | 28-day consumption fees x13 plus monthly subscription fees x12, at a date. |
| CSP | Cloud services provider such as AWS, Google Cloud or Azure. |
| Take-or-pay | A contract where the buyer pays for agreed capacity whether or not it uses it. |
| Convertible note | A loan that can turn into shares; marked to fair value as the company’s value changes. |
| SAFE | Simple agreement for future equity: cash now for shares at a later financing. |
| PBC | Public benefit corporation: directors balance shareholders, affected parties and a stated purpose. |
| LTBT | Long-Term Benefit Trust: financially disinterested trustees (five seats) who elect most of the board. |
| Non-GAAP | Company-defined measures that adjust standard accounting figures, e.g. adjusted operating income. |
Source: Anthropic S-1 (26 Oct 2026), "Select Defined Terms" and notes to the financial statements. Wording simplified by TheWrap.
Sources
Single source: Anthropic, PBC, Form S-1 Registration Statement, filed with the US Securities and Exchange Commission on 26 October 2026 (preliminary prospectus, subject to completion), 614 pages.
| S-1 section | Used for |
|---|---|
| Cover page, The Offering, Use of Proceeds | Offering terms, ticker, underwriters, use of proceeds |
| Prospectus Summary, Select Defined Terms | Headline metrics, definitions |
| Business | Thesis, products, customers, compute, talent, research, safety, regulation, legal proceedings |
| Management’s Discussion and Analysis | Results, quarterly data, non-GAAP measures, liquidity, agreements, commitments |
| Risk Factors | Disclosed incidents and supply facts only |
| Our Governance Structure and Its Goals | PBC, LTBT, share classes, sunset provisions, philanthropy |
| Management, Executive Compensation, Related Person Transactions | People, pay, related-party dealings |
| Principal and Selling Stockholders, Description of Capital Stock | Ownership and class terms |
| Consolidated financial statements and notes | Revenue disaggregation, concentrations, notes, preferred stock, subsequent events |
Source: Anthropic S-1 (26 Oct 2026). Ratios, shares and multiples computed by TheWrap from reported figures.