In AGNT Podcast Episode 16, Gemma Allen and Raphaëlle d’Ornano dig into Anthropic’s IPO filing and what it reveals about building frontier AI. They discuss why access to compute is central to Anthropic’s strategy, how the company makes money, and why controlling the customer relationship could become as important as building the most capable models.
D’Ornano also walks through the assumptions behind her valuation of Anthropic—and the risks that could change the calculation.
Recorded at the New York Stock Exchange with Gemma Allen, principal analyst at theCUBE Research.
Why is Anthropic committing so much to compute?
For d’Ornano, the most striking part of Anthropic’s filing is the scale of its compute commitments. Allen cites $895.4 billion in committed spending, raising a question that extends across the AI industry: will the infrastructure companies have promised actually be available when they need it?
D’Ornano has previously described compute as Anthropic’s biggest potential single point of failure. Frontier models require enormous resources to train and run, and she argues that staying at the frontier is essential to the company’s revenue. Anthropic has responded by securing capacity from multiple providers, including hyperscalers and neoclouds, while exploring a longer-term path toward more control over its infrastructure.
“If at any point they don’t have enough compute, it’s all over.”
Raphaëlle d’Ornano · 05:28
Allen points out that making commitments is one thing; getting access to the underlying power and computing capacity is another. Building more of that capacity itself could give Anthropic greater control, but d’Ornano notes that it would take time and substantial capital.
What kind of business is Anthropic building?
The filing also raises a question about the quality and durability of Anthropic’s revenue. D’Ornano argues that it should not be treated like conventional software-as-a-service revenue: much of the business depends on usage and the sale of AI tokens.
Those tokens reach customers through several channels, including cloud platforms and Anthropic’s own products, such as Claude Code and Cowork. The distinction matters because distribution through cloud providers comes at a cost, while first-party products give Anthropic more control over how it packages and sells its technology.
D’Ornano says token prices have fallen sharply, although she argues that pricing for the most advanced models has held up differently from the broader market. Anthropic’s challenge is to maintain pricing power while continuing to pay for increasingly demanding training and inference.
Can Anthropic turn a vast market into a durable business?
Allen distinguishes between the size of AI’s potential market and any one company’s ability to capture it. Even if the opportunity is enormous, frontier labs still face competition, potential commoditization, and uncertainty over who will own customer relationships.
For d’Ornano, expanding Anthropic’s first-party agent products is central to the answer. In her view, selling access to models through APIs alone presents a different—and more difficult—economic proposition from building products that customers use directly.
“We do have the conviction that this is a company that is going to continue growing and compounding its growth, notably by expanding its first-party agent products.”
Raphaëlle d’Ornano · 14:45
How does Raphaëlle value Anthropic?
D’Ornano describes building a discounted cash flow model rather than valuing Anthropic using a multiple of annual recurring revenue. Her model separates revenue channels and incorporates assumptions about compute, distribution costs, sales and marketing, and eventual operating margins.
In the scenario she discusses, Anthropic reaches $939 billion in annual revenue in 2036, with training and inference compute accounting for more than 40% of revenue through the end of her projections. Those assumptions produce a valuation of approximately $1.7 trillion before additional risk adjustments.
After accounting for risks—including insufficient compute, cybersecurity problems and harmful agent behavior—she says she would be more comfortable investing at a valuation closer to $1.1 trillion to $1.2 trillion. She describes the $1.5 trillion to $2 trillion range circulating at the time of recording as the high end.
These are d’Ornano’s modeled scenarios and judgments, not Anthropic forecasts or guaranteed outcomes. The valuation depends in particular on the company expanding its first-party products and sustaining growth.
What happens to margins as agents scale?
Training is only one part of the compute equation. D’Ornano expects inference—the computing required to run models—to become a larger consideration as agentic workloads expand.
She also points to costs beyond compute. Anthropic has benefited from developer enthusiasm and word of mouth, she says, but reaching a much larger enterprise market could require a bigger sales organization. Her model ultimately assumes an operating margin of around 30%.
The broader question is whether frontier labs can grow revenue fast enough to support both the infrastructure needed to stay competitive and the commercial operations needed to reach customers.
What to watch
Compute delivery: Whether Anthropic and its partners can turn long-term commitments into available computing capacity.
First-party products: Whether Claude Code, Cowork and future agent products strengthen customer relationships and pricing power.
Token economics: How quickly prices fall, particularly for frontier models, as training and inference demands grow.
The valuation: Which assumptions about revenue, compute costs, distribution and operating margins investors are willing to accept.
Agentic workloads: How the next generation of AI agents affects demand for inference—and the risks companies must manage.
Timestamps
00:52 Anthropic’s IPO and first impressions of the S-1
02:20 Compute commitments and the question of available capacity
03:06 Why compute is a potential single point of failure
05:34 Who controls the power and infrastructure?
06:12 Renting capacity and the longer-term case for building compute
08:42 Anthropic’s revenue mix and distribution channels
09:54 Token price deflation and first-party products
11:03 Market opportunity, competition and customer relationships
12:01 D’Ornano’s discounted cash flow approach
13:13 Revenue assumptions, valuation and risk
15:18 Compute spending and the cost of generating revenue
16:02 Inference, sales costs and operating margins
18:44 The bigger picture: the AI market and agentic swarms