Inderes’ Disclaimer can be found here. Detailed information about each share actively monitored by Inderes is available on the company-specific pages on Inderes’ website. © Inderes Oyj. All rights reserved.
This content is generated by AI. You can give feedback on it in the Inderes forum.
Updated:
AI translation: Original published in Swedish on 11/20/2026. Give feedback here.
Anthropic develops the AI assistant Claude. The company sells access to Claude to businesses and developers, directly or via AWS, Google Cloud, and Microsoft Azure, and also offers subscriptions for businesses and individuals.
Anthropic was incorporated in Delaware in January 2021 as a PBC (Public Benefit Corporation) and is headquartered in San Francisco. As the company is a PBC, its board is legally required to weigh shareholders' financial interests against the company's stated public benefit purpose. The company states that this may lead to decisions that do not maximize profit today, in favor of developing advanced AI responsibly and for the long-term benefit of humanity. As an example, the company notes that the Claude Mythos Preview model was not made generally available due to its capabilities in cybersecurity. The company was founded by seven co-founders, including CEO Dario Amodei and Chair of the Board and President Daniela Amodei. The LTBT (Long-Term Benefit Trust) structure was introduced in 2023, and the first models were commercially launched in Q1'23.
Anthropic has been listed on Nasdaq since November 19. The offering was upsized and priced at USD 83 per share, above the initial range of USD 70–78, and provides the company with ~79 BUSD before issue costs. In addition, seven strategic partners, including Google and Nvidia, are subscribing for non-voting shares for 26 BUSD at the listing price. Several of these partners are also suppliers to the company. At the time of writing, the stock is trading at around USD 95, ~14% above the listing price.
Revenue is primarily derived from consumption, where customers pay for the volume of text and data processed by the models, measured in tokens, with a price that depends on the chosen model. During 9M'26 (January–September 2026), consumption-based revenue accounted for 79% and subscriptions for 21%. The models are available in four classes: Mythos (including Fable), Opus, Sonnet, and Haiku, offering different trade-offs between capability, cost, and speed. On top of these, the company sells its own applications, such as Claude Code for programming and Claude Cowork for knowledge work.
Customers with at least USD 100,000 in revenue over the last twelve months numbered more than 11,900 as of Q3'26 and accounted for 74% of revenue. The remainder came from ~40 million smaller customers. No single customer accounted for more than 10% of revenue in 9M'26. ~44% of revenue in 9M'26 was sold via cloud providers' marketplaces (9M'25: 48%). Anthropic nevertheless has direct agreements with customers, recognizes the full revenue, and pays a fee to the cloud provider that is booked as cost of sales.
The largest cost is compute, meaning the data center capacity required to train new models and run them for customers. During Q1-Q3'26, the cost of running models for customers corresponded to 14% of revenue, while the development of new models accounted for 58%. Consequently, what weighs on the cost base is primarily the development of new models, rather than operations for existing customers.
Capacity is purchased in advance, typically 12–18 months before delivery, through multi-year agreements with Google, Amazon, Microsoft, SpaceXAI, Nscale, and Lambda, among others. The company also leases chips directly and places them in dedicated facilities operated by external parties, a model known as Anthropic Direct. As of Q3'26, including agreements entered into thereafter, commitments amounted to 895.4 BUSD, of which 447.5 BUSD falls due before the end of 2030. A large portion is payable regardless of actual utilization, but ~213.8 BUSD is cancelable or conditional. The commitments are largely absent from the balance sheet, as cloud capacity agreements are expensed only when the capacity is utilized and lease agreements are recognized only when they commence. Following the listing, Amazon and Google each own over 5% of the capital in non-voting shares. They simultaneously serve as compute suppliers and sales channels, but are also competitors.
Revenue increased to 16,956 MUSD, up 1,127% y/y and 46.7% q/q. The driver was new model releases, with models released over the past six months accounting for over 80% of revenue. Price cuts and discounts reduced revenue by ~600 MUSD compared to Q2'26.
EBIT was -567 MUSD, representing an EBIT margin of -3% (Q2'26: 974 MUSD, 8%). The swing is mainly explained by 5 BUSD in costs for 'short-term compute' (extra capacity purchased at short notice at a higher price than under the long-term agreements) and 2 BUSD in non-cash costs. Compute corresponded to 58% of revenue (Q2'26: 48%), and the adjusted EBIT margin, excluding SBC and non-cash items, was 15% (Q2'26: 19%). However, the adjustment for SBC should be interpreted with caution as it constitutes a real cost for shareholders. With the listing, stock-based compensation already earned is expensed at once, amounting to ~8.2 BUSD as of Q3'26, followed by an additional 24.1 BUSD over ~2 years. Reported and adjusted earnings are therefore expected to diverge significantly more going forward.
Net profit was -23.3 BUSD in Q3'26 and -104 BUSD for 9M'26, of which -80.7 BUSD was in H1'26. Almost the entire loss stems from fair value remeasurements. Amazon's convertible notes can be converted into shares at a fixed, low price and are accounted for at fair value. As the company's valuation rises, the liability increases, and this increase is booked as a cost. The same applies to the financing agreements with Google, Amazon, and AMD. The remeasurements do not impact cash, but because Amazon's ownership is subject to a cap, only a minor portion of the notes was converted upon the IPO, while the remainder stays on the books and is now remeasured based on the share price. The company's adjusted profit, excluding these items and SBC, was 3.2 BUSD in Q3'26 and 5.6 BUSD for 9M'26.
The prospectus presents cash flow only for the nine-month period. During 9M'26, operating cash flow was 6.6 BUSD, compared with an adjusted operating profit of 4.1 BUSD, and operating cash flow includes interest income of 2.4 BUSD. Cash and cash equivalents and short-term investments of 131.8 BUSD stem primarily from share issues (110.4 BUSD during 9M'26), and the company also has an undrawn credit facility of 14.8 BUSD.
The company describes its financial model through four drivers: increased usage of Claude, higher revenue per dollar invested in compute, the ability to grow with relatively few employees, and heavy investments in compute. The first three have improved rapidly. Revenue per employee exceeded 6.6 MUSD over the past twelve months, compared with 1.4 MUSD a year earlier, and personnel expenses fell from 40% to 15% of revenue between 9M'25 and 9M'26. Personnel expenses have therefore grown significantly more slowly than revenue, resulting in economies of scale as revenue increases.
The fourth driver works in the opposite direction in the short term. Compute is purchased in advance and arrives in large increments, and the cost is booked before the full revenue is realized. The company notes that margins will therefore come under pressure during periods of large deliveries, sometimes significantly, even if underlying conditions improve. Furthermore, the use of expensive short-term compute is expected to increase in the short term. According to the company, compute efficiency is intended to amplify returns if demand exceeds expectations and provide resilience if it falls short. Over time, Anthropic Direct is designed to lower costs and replace short-term capacity.
Several factors work against consistent and linear margin expansion. The largest cost – the development of new models – cannot be scaled down as long as revenue comes from the latest models, but rather recurs with every new generation. At the same time, the cost base is also largely fixed, and the company itself notes that it therefore has limited ability to quickly adjust costs. As the build-out of its own data centers accelerates, the cash flow profile changes because investments and prepayments are made long before the capacity generates revenue. The company therefore expects a significant net outflow. How well the model scales is most clearly reflected in how many dollars of revenue each dollar of compute generates. The metric was below 1x in 2025, rose to 2.08x in Q2'26, and fell to 1.72x in Q3'26 as new capacity was expensed before revenue had caught up. Because capacity arrives in increments, a single quarter tells us relatively little. The decisive factor is whether the metric improves over several quarters.
Anthropic operates in the market for general-purpose AI models and the agents built on top of them. Developing frontier models is, according to the company, extremely capital-intensive, the manufacturing of chips and memory is concentrated among a few suppliers, and several partners are simultaneously competitors. The company identifies OpenAI and Google as direct competitors, alongside other major tech companies and providers of open models, which the company states primarily compete where customers prioritize cost over capability. The biggest shift in demand is that enterprises are moving from asking AI models questions to letting agents perform work independently, with programming and Claude Code being the clearest example to date. Penetration remains low: among Fortune 500 companies that are major customers, spending with Anthropic averaged under 0.1% of their total costs. At the same time, export controls and U.S. state-level AI laws affect what may be sold, and to whom.
A structural risk is that AI capabilities become a commodity, and the company itself highlights two variations. The first is that the utility of the next improvement diminishes. Customers may conclude that less advanced and cheaper models are sufficient, and that improvements at the top do not justify the additional cost. In that case, they may switch to smaller, cheaper, or open alternatives – which, according to the company, can happen even in a market where overall AI adoption continues to grow. If models become increasingly interchangeable and prices fall faster than usage increases, the fixed commitments will become heavier to bear. This is also the question that is absolutely central to the assessment of the company. The second is that the ability to build highly capable models spreads, partly through distillation and AI-assisted research. If so, the company's technological lead could shrink, and competition may instead be determined by access to compute. At the same time, the company is itself lowering customers' cost per task through both lower prices and more efficient models. Opus 5.5 reduced customers' cost per task by ~40% compared with Opus 5, and the company argues that compute efficiency makes this possible without margins deteriorating.
The buyer gets financial exposure but limited influence. LTBT, a foundation whose five trustees have no financial interest in the company, owns the sole Class T share and appoints four of the seven board members. Special rules apply to one of the seats, including a vote among employees if the company and the foundation disagree. The founders' joint company owns the sole Class F share, which carries 50.1% of the votes on other matters, including the election of the other three board members. Class A shares therefore carry at most 49.9% of the votes combined, and the founders also own Class A shares of their own. Neither the Class T nor the Class F share carries dividend rights. Class F will terminate after a transition period that begins when fewer than three of the founders have remained for 90 consecutive days. The company describes the structure as a safeguard for its long-term mission. Among the risk factors, the company states that the structure can effectively remove Class A shareholders' ability to influence key decisions, including a change of control. Because LTBT appoints the board majority, Anthropic is classified as a controlled company, but the company uses the exemption only for the nominating and corporate governance committee.
The dependence runs in several directions. Around 44% of revenue during 9M'26 was sold via cloud providers' marketplaces (9M'25: 48%), i.e. via companies that also develop their own AI models. The company holds the contract directly with the customer, sets the price, recognizes the full revenue, and pays a fee that is included in selling expenses. The line item for sales, marketing, and partnerships fell from 22% of revenue in Q1'26 to 14% in Q3'26, which the company attributes partly to a smaller share being sold via channel partners and to changes in terms. The company has committed to purchasing compute of at least 121.3 BUSD from Google, 110 BUSD from Amazon, and 44.9 BUSD from Microsoft, and after the IPO, Amazon and Google each own over 5% of the capital. The company's view is that this arrangement provides reach and avoids channel conflict, and that running workloads across four chip families reduces dependence on any single supplier. The negative view is that these same companies account for a large share of both sales and capacity while also competing, which the company itself notes can create conflicts of interest.
According to the company, over 90% of contributions to its codebase in September 2026 were written by Claude. The average amount of code per engineer per day in Q3'26 was over 45 times higher than in 2024. Headcount increased by ~170% in a year to ~5,300, while LTM revenue increased by ~1,200%. Personnel expenses accounted for 15% of revenue in 9M'26, compared with 40% a year earlier. This can be read as a sign that personnel expenses may grow more slowly than the business. At the same time, the company's own use of Claude is included in compute costs, so part of the productivity gain is paid for in compute rather than salaries. Nor does the amount of code in itself say anything about quality, and personnel is a smaller cost item than compute.
Anthropic sells primarily to enterprises that embed the models into their workflows. Growth comes largely from existing customers increasing their usage, and even the largest customers still spend a small share of their costs on Claude.
The cost of serving customers was 14% of revenue in 9M'26. Profitability is determined by development costs, which are spread across a growing revenue base, and ongoing operations generated positive cash flow.
Most of the net loss comes from loan revaluations that track the company's valuation and do not affect cash. Adjusted for these and for SBC, the company was profitable, with a profit of 5.6 BUSD in 9M'26.
Of the 895.4 BUSD in commitments, only ~213.8 BUSD are conditional or cancelable. Costs rise in steps before revenue catches up, as Q3'26 showed.
If models become more interchangeable, prices could fall faster than usage grows. Price cuts reduced revenue by ~600 MUSD in Q3'26, and open models compete on lower prices.
~44% of revenue is sold via cloud providers, which also account for a large share of capacity, while several of them are at the same time developing their own models and financing the company.
This Equity Brief is produced by Inderes. It is intended solely to provide a concise, fact-based presentation of publicly available information about the company at the time of publication. The Equity Brief is intended for informational use only and should not be construed as an offer or advice to buy, sell or subscribe to investment products.
The information has been compiled from several different public sources that Inderes considers reliable. Inderes aims to use reliable and comprehensive information, but Inderes is not responsible for the content or accuracy of the presented information. The information underlying the Equity Brief may change quickly. Inderes makes no commitment to announce any changes to the presented information or to update the material after publication.
The Equity Brief should not be construed as an offer or advice to buy, sell or subscribe to investment products. The reader is responsible for their own investment decisions. The reader should consider their own objectives and financial situation and, where necessary, use advisors when making investment decisions. Inderes and its employees are not responsible for the financial outcome of investment decisions made based on the Equity Brief, or for any other damage (direct or indirect) that may arise from the use of the information.
The Equity Brief may not be edited, copied or made available to others, in whole or in part, without Inderes' written consent. The laws of certain countries may contain restrictions on the distribution of information in this Equity Brief, and persons subject to such restrictions should take these restrictions into account.
Parts of this Equity Brief have been prepared with the assistance of artificial intelligence tools. All content is reviewed and edited by Inderes before publication.
Inderes has made an agreement with the issuer discussed in this Equity Brief, which entails compiling the Equity Brief.
Information about Inderes