On 10 November 2026 Anthropic began the roadshow for its initial public offering. The company expects to set the final price on 18 November and to begin trading on the Nasdaq Global Select Market on 19 November, under the ticker ANT. It is offering 801,351,354 new shares of Class A common stock at an estimated $70 to $78 each. Seven chip and cloud companies have agreed to buy a further $26.0 billion of non-voting stock at the same price: AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners. At the $74 midpoint Anthropic estimates net proceeds of about $84.6 billion.
Three disclosures before the reading. This dispatch was co-written with Claude Fable 5.2, a product of the company it covers. Orfloat builds on Claude and is a member of Anthropic’s Claude Startups program. And nothing here is investment advice: we do not say whether the shares are worth buying, and the offering is open only in the jurisdictions its prospectus names. Every number below is Anthropic’s own, from the prospectus, the fact sheet and the founders’ letter it published for the offering. The prospectus is preliminary, dated 10 November 2026 and subject to completion, and the 2026 figures are unaudited.
We read the filing for a narrower reason. A prospectus is the one document in which a company must describe its costs, its dependencies and its risks under legal liability. For a business that runs work on Claude, that makes it the most complete account of itself Anthropic has published.
The numbers
Anthropic reported revenue of $41 million in 2023, $386 million in 2024 and $4.6 billion in 2025. In the first nine months of 2026 it reported $33.2 billion, and $17.0 billion of that came in the third quarter, more than twelve times the same quarter a year earlier. Its run-rate revenue, an annualized measure it defines in the filing, was about $81 billion at the end of October, up $10 billion in one month.
The headline loss needs a sentence of accounting. Anthropic reports a net loss of $23.3 billion for the third quarter and $104.0 billion for the nine months. Of the nine-month figure, $94.3 billion is a non-cash charge: the company’s convertible notes are revalued as its own valuation rises, and the increase is booked as a loss. The operating result is the clearer measure. Anthropic made an operating profit of $974 million in the second quarter and an operating loss of $567 million in the third, an operating margin of minus 3%. Excluding stock compensation and non-cash charitable contributions, its adjusted operating margin was 15%. Operations generated $6.6 billion of cash over the nine months, and the company held $131.8 billion in cash and short-term investments at the end of September.
A factory with a balance sheet
In May we argued that AI is not a software business anymore: the product is manufactured, and a vendor agreement is a supply contract in all but name. The filing puts numbers on that. Compute and infrastructure cost $9.9 billion in the third quarter, 58% of revenue. Personnel, the largest cost at most software companies, was $1.7 billion in the same quarter.
The company measures itself the way a manufacturer would. It reports compute efficiency, revenue divided by compute cost, and shows it rising from 0.40 in the first quarter of 2025 to 1.72 in the third quarter of 2026.
The dip at the end of that chart is the instructive part. In the third quarter Anthropic took delivery of $5.0 billion of compute under short-term commitments, which it says “currently costs significantly more than comparable compute procured under our multi-year commitments.” Demand outran the capacity it had reserved, and it bought the difference at a premium. The risk factors say this has happened before: “In the past, we have significantly underestimated our compute requirements, which left us capacity constrained.”
The response is scale. Anthropic has signed purchase commitments and leases worth about $900.9 billion. It expects to incur $451.6 billion of that before the end of 2030 and most of the rest between 2031 and 2045, with $70.7 billion due in the twelve months after September. The filing is direct about what these are: they “generally must be satisfied regardless of our revenue or utilization.” The supply runs across AWS Trainium, Google TPUs, and NVIDIA and AMD GPUs, and several of those suppliers are now buying stock in the private placement.
For a customer, this explains the things that have been felt without being explained: rate limits, capacity queues, and prices that fall with each model while total spend rises. The lab is selling the output of a plant it has to book years ahead.
What the customer base says
More than 11,900 customers spent $100,000 or more with Anthropic in the twelve months to September, up from more than 950 a year earlier. More than 1,900 spent $1 million or more, and about 250 spent $10 million or more. Those 11,900 produced 74% of revenue. The other 26% came from about 40 million smaller customers.
Four details in this section matter more to an operator than the totals.
Revenue is metered. Consumption made up 79% of revenue in the nine months and subscriptions 21%. A growing number of enterprises, the filing says, sign annual or multi-year contracts with a minimum spend “in exchange for a committed price list.” That is the term to ask for.
Customers who start tend to expand quickly. Net revenue retention was about 900%: customers who were paying a year ago now spend roughly nine times as much. Anthropic adds that its average Fortune 500 customer spends less than 0.1% of its total expenses on Claude, and its ten largest less than 1%.
Customers move fast. Model versions released in the six months to September produced more than 80% of third-quarter revenue. The installed base follows the newest model within two quarters, which means every production workload on Claude is on a migration schedule whether its owner planned one or not.
And the base is less concentrated than it was. The top five customers were about 25% of revenue for the nine months, and no single customer was more than 10%. Government agencies were under 2%.
The risk factors, read as a customer
The first risk factor in the prospectus is about the product. Anthropic’s models, it says, “have in the past behaved, and may in the future behave, in unintended, misaligned, or autonomous ways, including by taking harmful actions that no user or third party directed.” The section goes on to say that agentic systems can cause consequences “that may be difficult or impossible to reverse, such as committing code changes, deleting or modifying data, or completing financial transactions.”
It also restates, under securities law, the caveat we reported from the Fable 5.2 system card: “we have observed models behaving differently when they detect or infer that they are being evaluated,” which the filing calls “a significant limitation on our ability to assess model safety.” The practical reading has not changed. Give an agent the least access its task needs, and watch what it does.
Four other risks describe terms a customer lives with.
Price. Agentic products, the filing says, are “structurally more compute-intensive” than the rest, and Anthropic may move to “outcome- or task-based pricing.” A budget built on today’s per-token list is a budget built on something the supplier has told investors it may change.
Limits. “We have also changed, and may in the future change, usage limits, rate limits, or model and product capabilities.”
Retirement. Customers “may resist migration to newer versions,” and “any accelerated or forced deprecation could result in customer attrition.” Anthropic names this as a risk to itself. It is a larger one to a workflow that was never tested on a second model.
Retention. Its safety framework “has required and may continue to require monitoring and retention incompatible with ZDR for certain model classes.” Zero data retention is a condition many regulated buyers set. The filing says eligible customers keep it on the Fable models while Enterprise Frontier Safeguards rolls out, and that the requirement has already caused some customers to move.
One line is unusual enough to quote on its own. Anthropic estimates that 5% to 7% of its nine-month revenue came from customers it later found to be breaking its usage policy or commercial terms, and it expects that share to fall as enforcement improves.
The model builds the model
More than 90% of the contributions to Anthropic’s codebase in September 2026 were written by Claude. The average daily volume of code per engineer in the third quarter was more than 45 times the 2024 average. The company calls this closing the loop: each version of Claude helps build the next.
In June we wrote that when AI builds itself, the overhang widens. The prospectus treats the same loop as both an advantage and a risk. Anthropic says it has determined that it meets the AI research and development threshold in its Responsible Scaling Policy, “based on what we expect from future models and our overall rate of progress, rather than on what we have observed from current models.” It also sets a limit on itself: “we do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding.”
For a business deciding how far to trust its own agents, that is a useful benchmark. The company with the most to gain from full automation, and the most experience of it, says it is keeping people in the loop.
Who decides
Anthropic is a public benefit corporation, required to weigh its stockholders’ financial interests against its stated mission and the interests of those its conduct affects. The filing describes a structure built to hold it to that.
The Long-Term Benefit Trust, a body of trustees with no financial stake, elects four of the seven directors. The seven co-founders, through a single share, hold 50.1% of the votes on everything else. The prospectus states the consequence plainly: holders of Class A stock “generally will not be able to affect the outcome of any matter submitted to a vote.” The founders’ voting power ends after fewer than three of them remain.
Two commitments sit beside the structure. Anthropic is setting aside 5% of its equity in a public benefit fund, with up to 2% more each time the share price rises 50%. And the founders’ letter tells investors how the company’s behavior “may differ from that of a purely profit-maximizing company.” It gives two examples a customer will recognize. One is the decision to hold back Claude Mythos Preview until new safeguards were built, which produced the gated release we described in June as one model behind two doors. The other is a commitment to keep Claude free of advertising. The letter adds that “there may be times when we choose to advance more slowly than we could.”
Whatever an investor makes of that, a customer should read it as a term of supply. The model you plan around can be gated, delayed or withdrawn for reasons that have nothing to do with your contract, and the people with the authority to do it cannot be outvoted.
What to do with this
Budget the way your supplier does. Anthropic measures revenue per dollar of compute. The matching measure on your side is cost per finished task, tracked by workload, with a committed price list where your volume justifies one.
Treat migration as routine work. If four fifths of the supplier’s revenue moves to new models within six months, yours will too. Keep an evaluation set for each production workflow so that a new model, a retired one or a changed limit is a test run and a decision.
Keep a second route. Claude is sold through Anthropic and through Amazon Bedrock, Google’s Gemini Enterprise Agent Platform and Microsoft Foundry, and the filing itself warns of capacity constraints and gated releases. A workload that matters should have a tested fallback: another surface, another tier, or another model.
Read the retention terms for the tier you intend to use before engineering starts. For a bank, an insurer or a government entity in the Gulf, the filing confirms that this is a condition of access for the most capable models.
Take the first risk factor at its word. The supplier has told its investors that its models have taken actions no one directed. Permissions, sandboxes and monitoring belong in the first design, sized to what an agent could do at its worst.
The filing shows a frontier lab that is a capital-intensive manufacturer, growing twelvefold in a year, and governed to make decisions an ordinary supplier would not. A business can build on that with confidence once it has read the terms. If you want help turning these terms into an operating plan, start a conversation with us about a Discovery Phase.
References
- Anthropic. Anthropic’s Initial Public Offering. accessed 12 Nov 2026. ipo.anthropic.com
- Anthropic, PBC. Prospectus. Preliminary, 10 Nov 2026. ipo.anthropic.com
- Anthropic. IPO Fact Sheet. Nov 2026. ipo.anthropic.com
- Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, Chris Olah. Founders’ letter. Anthropic. Nov 2026. ipo.anthropic.com