What should investors glean from Anthropic’s prospectus?

Anthropic, the frontier AI lab, filed its most-anticipated IPO prospectus last week. It aims to go public in mid-November with the ticker ‘ANT’ on Nasdaq. While the issue size is not made public yet, the AI lab is reportedly seeking a valuation of $1.5-2 trillion or 75-100x trailing revenue.

AI, in its current form is a novel technology to the world — like the internal combustion engines, computers and the internet of the past — and the technology is evolving day by day. At this point in time, a layperson’s estimate of the extent of disruption that AI could cause is probably as good as a scholar’s. Therefore, it is really challenging to define an addressable market for Anthropic and the market share it could command in the future. However, the risk factors detailed here introduce enough uncertainty to cloud the investment case for the time being. Also, with interest rates elevated, the valuation leaves no margin for error.

Nevertheless, here’s a summary of Anthropic’s business, things going for it and those that are not.

Business

Anthropic, founded in 2021, is an AI research lab that develops frontier AI models and lets customers use them for a consideration. Its products include the Claude family of large language models (LLMs) — Mythos, Fable, Opus, Sonnet and Haiku, and agentic AI platforms Claude Code for software developers and the more generic Claude Cowork meant for automating workflows. Anthropic’s customers range from individual users to large corporations.

Anthropic makes revenue in two ways depending on the customer’s usage pattern—consumption-based and subscription-based. Under the former model, the customer is billed according to their consumption of tokens. For example, assume a software developer builds a translation app that uses Claude. Every time someone uses the app, tokens are consumed for translation. Anthropic will then bill the software developer based on the number of tokens the app has consumed. Agentic use-cases, including Claude Code for coding, are also billed based on the number of tokens consumed. The latter revenue model is rather simple and more familiar. Users pay subscriptions to use Claude models for an agreed period of time. Table 1 contains the split between the two models, which is roughly 80:20.

Besides, Anthropic also works directly with companies requiring guidance on automating complex workflows. Anthropic engineers working with a pharma company on drug discovery, can be an example. Their recent offering is called ‘Ode with Anthropic’, which is an AI services company aimed at taking Claude’s capabilities to mid-sized companies and providing them support.

That said, what are Anthropic’s costs? Refer Table 1 for this part.

The first is compute and infrastructure cost — the most important cost-head. Note that Anthropic’s models are hosted on hyperscalers or large data centres operated by companies such as Google, Amazon and Microsoft. These data centres are equipped with powerful chips, including the likes of Nvidia’s GPUs (graphic processing units) that Anthropic uses — both for model development (training) and inference. Inference is the process by which an already developed/ trained AI model works on the user’s input to generate an output. Anthropic leases server/ compute capacity from these cloud companies, whom Anthropic calls as CSPs or cloud service providers, over long-periods of time. The annual lease expense is a direct cost and is captured under this cost-head. This was 55 per cent of revenue in H1 2026.

The second cost item is personnel expenses, which is self-explanatory. These accounted for 19 per cent of H1 2026 revenue. Interestingly, Anthropic appears to compensate their employees more using share-based payments (like stock options). Two-thirds of personnel cost in H1 2026 were share-based payments. Share-based payments do not entail cash outflow as in the case of salaries, but the trade-off lies in a future dilution for shareholders.

Next up is sales, marketing and partnerships. The sales, marketing part should be self-explanatory. We’ll talk more about partnerships. Anthropic refers to the CSPs as partners here. Here’s how it works. With the advent of cloud, enterprise software is now hosted on cloud platforms such as Amazon AWS, Microsoft Azure and Google Cloud. Anthropic uses them as distribution to reach the thousands of companies whose workflows take place on these platforms. In doing so, it promises to pay CSPs a sum, which is included in this cost-head. Overall along with marketing, these expenses were 19 per cent of H1 2026 revenue.

Finally, other operating and administrative costs. These work out to 11 per cent of H1 2026 revenue. However, this item generally may not be this high. It spikes in quarters where there are expenses linked to the company’s ‘donation matching program’. Under this program, if eligible employees pledge to donate to eligible non-profit organisations, the company will match that donation, by issuing shares (non-voting Class N common stock, to be precise). We’ll talk more on this in a moment.

Deducting the above costs from revenue gives operating income. Adjusting for the non-cash share-based payments and donation matching expense gives adjusted operating income of $1.6 billion, which is 10 per cent of H1 2026 revenue.

Between operating income and net income are interest income and taxes. However, there is also another item. Anthropic has issued convertible notes (or convertible bonds) and redeemable convertible preferred stock (RCPS) over various funding rounds. These are essentially equity waiting to be converted. Anthropic carries them at fair value and so, changes in their fair value must be recognised in P&L. If the value of these instruments increases between quarters, there will be a marked-to-market expense and vice-versa. This will continue until they are eventually converted to equity. Of Anthropic’s $81 billion of reported loss in H1 2026, almost all of that $81 billion had to do with fair value changes and thus are considered as non-cash adjustments to arrive at adjusted net income. Adjusted net income in H1 2026 works out to $2.7 billion which is roughly the sum of adjusted operating income of $1.6 billion and interest income of $1.1 billion.

Public benefit corporation

Anthropic is constituted as a PBC or public benefit corporation. Though it may seem contradictory, this enables the company to balance three things at the same time: interest of the shareholders, interest of the stakeholders (who are affected by the company’s conduct; includes customers and the broader society because of AI’s pervasive nature) and its public benefit purpose stated in its charter — ‘the responsible development and maintenance of advanced AI for the long-term benefit of humanity’. The AI lab has a wholly-owned subsidiary called ‘Anthropic Public Benefit Fund LLC’ (fund) to carry out its public benefit initiatives.

Why is it contradictory? Take Meta Platforms’ case. Had it been a PBC, it may not have set up algorithms to keep engagement elevated among kids, keeping the interest of its users in mind. However, this may not be in the best interest of shareholders, as it would mean lower ad revenue. Back to Anthropic. Because AI is a generational technology and is too powerful to be taken lightly, perhaps it calls for such a structure. Its success and whether Anthropic is using it only as a façade to build goodwill or whether it is serious about it, remains to be seen.

Here are a few facts. Anthropic has already contributed 5 per cent of fully-diluted shares, as of June 30, 2026, to the fund. The ‘donation matching program’ also is meant for this purpose. The seven founders have separately pledged 80 per cent of their personal equity stake. Further, the company plans to issue about 2 per cent of fully-diluted shares every time the stock price grows 50 per cent, after going public.

What’s to like

While it is a fact that Anthropic is yet to become profitable, it cannot be denied that it has shown incredible growth in revenue — the kind tech start-ups can only dream of. Founded only in January 2021, its revenue has reached $19.7 billion on a trailing 12 months (TTM) basis. This is quite impressive, considering the time to develop the models from scratch and go-to-market. Alphabet reached $20 billion in revenue 10 years after inception, Microsoft in 24 years and Meta Platforms in 13 years.

The flywheel behind Anthropic’s growth is simple: Investment in more, better compute > Model improvement > Efficiencies for customers > Revenue growth. The number of large customers from whom Anthropic generated $100,000 or more in TTM revenue has grown from about 20, as of December 2023, 200 (in December 2024), 1,500 (in December 2025) to 11,900, as of September 2026. Customers from whom Anthropic generated $1 million and $10 million in TTM revenue were about 1,900 and 250, as of September 2026. Growth has not only come from customer additions but also in business with existing customers. This is captured in the net revenue retention of around 900 per cent, as of September 2026. This means that if Anthropic generated $1 of revenue from a customer in July-September 2025, it generated $9 from the same customer in July-September 2026.

Such rapid growth has resulted in operating leverage, with the AI lab turning profitable both at the operating level and net level, in Q2 2026, on an adjusted basis (see Chart 1). Revenue to compute and infrastructure cost ratio is on the rise — from $0.4 of revenue per $1 of compute cost in Q1 2025 to $2.1 of revenue per $1 spent on compute in Q2 2026. Operating cash flows, too, have turned positive in H1 2026.

The balance sheet is also flush with cash of about $131 billion, from recent funding rounds. For perspective, the entire balance-sheet size is only $147 billion.

What’s not to like

If you are impressed with the company so far, hang on. The factors that investors need to be concerned about, outnumber the positives. We’ll deal with them one by one. But before proceeding, do bear in mind the company’s asking valuation of $1.5-2 trillion (per media reports), and evaluate the factors discussed.

Growth rate tapering?

First, the financials disclosed in the prospectus stop with H1 2026. This, while the company was comfortable with sharing data as of September 2026 for a select items—some of which have been discussed above. This adds fuel to the sceptic voices in the market that suggest this to be the company’s deliberate choice. Stopping financials as of H1 2026 may, in fact, be advantageous to Anthropic — because Q2 2026 was the quarter that broadly saw the end of token-maxxing. Token-maxxing was a trend among corporates that incentivised employees based on the number of tokens consumed rather than on the effectiveness of the outcomes. The trend even saw companies such as Ford fall back to hiring employees, after AI investments failed to live up to expectations. Per media reports, sequential revenue growth in Q3 2026 appears to have declined from 147 per cent in Q2 2026 to 45 per cent (see Chart 1). In fact, this could be the reason why the rumoured asking price may have come down to $1.5 trillion, which was maintained at $2 trillion sometime ago.

A 45-per cent sequential growth is still good, but perhaps in isolation. When building on a low base (TTM revenue of $19.7 billion), as compared to the prospectus indicated addressable market of $22.5 trillion by 2031 and given an asking multiple of up to 100x trailing revenue, 45 per cent growth does not make the cut. Also, makes one question, ‘Is AI adoption slowing?’

A problem only growth can solve

A slow growth rate comes across as a bigger problem considering this. Look at Table 2. Forecasting a demand surge for its products, Anthropic has struck massive long-term deals (some extending to 2036) with CSPs for leasing compute, that work out to a total of $827 billion. The leases are yet to commence, and so, the liabilities in relation to them are yet to show up on the balance sheet. While payments for some of them are based on actual consumption of compute, at least 62 per cent or $512 billion is on non-cancellable basis. Meaning, the lease obligations in these deals must be honoured, irrespective of whether Anthropic manages to garner demand for its products — a problem only sustained growth can solve. In this context, a weakening growth rate does not help Anthropic’s case.

The prospectus also suggests that Anthropic may pivot from leasing to building its own compute infrastructure. Even in that case, the problem does not go away, as the lease payments morph into upfront capital commitments. How Anthropic might source the cash for this is in the dark.

Can decline in training cost help?

Unlikely. Anthropic’s compute and infrastructure cost includes both cost of model development and inference. Model development costs are not capitalised. Given the AI lab operates in the frontier of AI technology and is a strong believer in scaling laws, it is unlikely to stop training new, more powerful and more autonomous models. Scaling laws suggest that when a model’s size, training data and compute are scaled up, there is a predictable rise in the model’s performance. Training cost can thus be reasonably expected to continue.

This is in contrast to the business models of traditional software and internet companies, where the spend on developing intellectual property is largely upfront, leaving the company to reap the benefits of operating leverage later.

Compute up, tokens down

Anthropic’s token prices are now just one-fourth of what they were in January 2025 (see Chart 3). This is as per Ramp’s AI index, which tracks the AI spending of its clients. Ramp is a fintech company that sells corporate expense management software. Token prices have now fallen to $0.62 per million tokens versus $2.56 per million tokens in early January 2025.

This raises a question: If Anthropic’s moat is real and demand for its models is so strong, why should token prices fall so drastically? Token prices of OpenAI are also on the decline. With a clutch of cheaper Chinese open-weight models having entered the fray, it raises concerns whether AI models are getting commoditised. Customers, too, are getting smarter and using tools that choose the model according to the complexity of the user’s prompt. Basic prompts need not be sent to the most sophisticated models, thus saving tokens.

On the other hand, data from BCA Research indicate that compute costs are rising (Chart 2). Constraints in the supply of power to data centres under construction make it difficult to assume a drop in compute costs in the near term.

Concentration risk

While no customer represented more than 10 per cent of 9M 2026 revenue, top five customers account for about 25 per cent of 9M 2026 revenue. The company is also becoming increasingly dependent on the distribution of CSPs. Share of revenue from CSP marketplaces is up from 11 per cent in 2023, 32 per cent in 2024, 47 per cent in 2025 to 48 per cent in H1 2026. This is a risk in the sense that the same CSPs also distribute their own AI offerings, such as Gemini and Copilot, as well as rival models like ChatGPT, whose developers could pay more for promotion than Anthropic.

Circular financing, dilution

Over the years, Anthropic has frequently raised funds using convertible instruments such as convertible notes and RCPS. These are often circular financing deals where CSPs such as Google and Amazon subscribe to these instruments, with the apparent intent that some of the money will flow back to them when Anthropic pays for compute. They also have the optionality of capital appreciation when they are converted to equity or when Anthropic undergoes a successful listing. For instance, look at the Broadcom deal in Table 2. Out of the $161 billion in total lease obligations, Broadcom allows Anthropic to settle up to $42 billion by issuing convertible notes (notes yet to be issued).

While these arrangements may be convenient, they might result in significant dilution for equity holders. These instruments are carried at their fair values (their values had they been converted to equity) as said above, and together they add up to $266 billion, as of June 30, 2026. For context, balance-sheet total on this date was $147 billion, after considering $133 billion of accumulated losses.

The dilution does not stop there. Employees are largely paid in stock options and RSUs. About 64 per cent of personnel cost over 2023-25 and in H1 2026 was met using share-based compensations. Equity holders also need to factor in the dilution from issue of shares in relation with the ‘donation matching program’ and other contributions to the public benefit fund.

Legal, regulatory hurdles

Given AI’s powerful capabilities and the potential misuse through deepfakes, cybercrime and hacking, investors should expect greater regulatory scrutiny of the AI ecosystem, which could constrain growth or may even result in liability for companies like Anthropic, if something goes wrong.

Further, while Anthropic apparently harbours noble intentions with its public benefit initiatives, it faces allegations of copyright infringement in multiple lawsuits. There is a high likelihood that Anthropic may not be able to defend itself in these lawsuits, forcing it to settle off court. In August 2024, a group of authors sued Anthropic, alleging that it had infringed their copyrights by using their literary works to train LLMs. The company later agreed to a $1.5-billion settlement (known as Bartz settlement).

Bottomline

Today, much of the US equity market has become one AI trade, betting on the AI investment cycle, from chips and cloud infrastructure to software and power companies. Therefore, Anthropic’s IPO is more than just a listing. For AI bulls, it is a referendum on whether the market believes AI labs can evolve into high-quality businesses. For AI bears, it’s about vindication if the bubble breaks. What unfolds in the first few months to a year after Anthropic’s listing could go down as one of the most defining episodes in the history of stock markets.

Published on October 31, 2026