Anthropic: Margins, Competition, and the IPO Bet on AGI

Anthropic's S-1 filing reveals an AI lab with positive past inference margins and rapid growth, but in a market already contested by OpenAI, open source, and integrated giants. Between reliance on third-party clouds and accelerating efficiency gains, the trajectory toward giant revenues hinges on expanding into business use cases far beyond programming.

Sector Growth Shifts: OpenAI and Open Source Gain Ground

In the first half of the year, Anthropic accounted for 90% of the new net annual recurring revenue in the industry. By the third quarter, OpenAI likely dominated this momentum, while open source grew from about $1 billion in ARR in token inference costs in the first quarter to over $10 billion in the third quarter. Several executives prefer to deploy AI on-premises rather than entrust their data to OpenAI or Anthropic; law firms are already using AI in this manner. Low switching costs in programming facilitate transitions between systems. Meanwhile, companies born with AI are customizing open-weight models and conducting inference for both internal and external uses. Google DeepMind and Meta Superintelligence Labs are well-funded and expanding their training capacity, while Elon Musk announces an additional 10 net gigawatts at SpaceX next year, half of which will be dedicated to long-term internal uses. In this context, competition is intensifying, and future growth is expected in verticalized use cases rather than just programming.

Economic Units: From -94% to 85% Inference Margin

Two years ago, Anthropic reported a gross margin of -94% on inference; it is now at 85%. The discussion around gross margin includes revenue per megawatt in inference and efficiency gains observed by inference operators. Anthropic operates a diverse fleet of chips, including Trainium, TPU, and GPU. The token throughput on Trainium has increased by 5 to 10 times in a year, gains attributed to hardware and software improvements, as well as the interaction between models and hardware.

Clouds and Chips: External Commitments and Promises of Cost Reduction

Anthropic relies on third-party clouds, with commitments from SpaceXAI, Google, and Amazon. Google designs its own chips, and Sundar Pichai has stated that the company knows how to reduce inference costs. The contrast highlights that operational efficiency can stem from vertical integrations among certain players, while others optimize compute allocation on partner infrastructures.

Anthropic's Revenues Remain Limited to a Few Major Clients

The two leading labs together report an annual recurring revenue of about $120 billion, but the market remains limited. The top 1% of clients spends between $50,000 and $100,000 per employee, while large companies, despite high computing consumption, invest little in AI. At Anthropic, spending primarily focuses on programming, giving the company an almost exclusively vertical positioning. The fourth-quarter budget season has yet to yield significant interactions with the labs, as ARR accelerated earlier in the year.

Where to Look Next: Verticals, Office, and $1 Trillion Hypotheses

Cybersecurity is identified as closely related to programming, with multi-million dollar contracts and rapid scaling linked to incidents. Biotechnology is also mentioned among the verticals. By 2027, adoption in office work is anticipated, transitioning from AI augmentation to targeted replacements. Microsoft's buyout offers to employees are cited as an example of work dynamics. Dario Amodei discusses a possible trajectory toward $1 trillion in revenue, a target that assumes expansion beyond the current scope.

Anthropic IPO: Numbers, Reactions, and Areas of Uncertainty to Clarify

Anthropic could be traded as early as next month, with a valuation range estimated between $1.5 and $2 trillion, making it one of the most valued companies. The S-1, a substantial document, was followed by stock rebounds among semiconductor and cloud players. Joey Brookhart considers margins, computing margins, and free cash flow in line with expectations and finds the filing difficult to criticize. The Tokenomics practice of SemiAnalysis focuses on the unit economics of AI and also covers hyperscalers and neo-clouds; prior to the S-1, their models relied on public information, with the arrival of audited figures facilitating explanation. Structural differences are highlighted between OpenAI and Anthropic regarding consumption, subscription, and the historical weight of the consumer versus the enterprise. Investors are still questioning compute allocation and the business model of inference. Joey Brookhart also mentions three challenging weeks since the investor day and disappointments regarding September's ARR, while noting that the company quickly reached a valuation exceeding $1 trillion and more than $70 billion in ARR.