The question came up in the first minutes of the meeting, and Daniela Amodei did not flinch. When will the billions spent on AI models produce returns? Anthropic’s president, pitching investors ahead of the company’s planned listing, answered with the confidence of someone who believes the answer is already visible: enterprises are paying for capability, and capability is compounding. TechCrunch reported that she pushed back on doubts about AI investment returns during the roadshow.
Anthropic filed its draft registration statement with the Securities and Exchange Commission on June 1, making it the first major AI laboratory to formally begin the public-market process. Reports put the expected valuation near $965 billion, built on roughly $47 billion in annualized revenue and what sources describe as extraordinary growth since 2024. The company’s pitch to investors leans on an even larger number: a total addressable market it pegs at more than $30 trillion, based on the full range of work AI models could perform across industries, according to The Wall Street Journal.
The same week, OpenAI’s chief financial officer, Sarah Friar, confirmed that the company’s self-designed AI hardware would ship before the end of the year — a schedule that pulls forward the 2027 timeline earlier filings suggested. The two announcements, taken together, sketch the fork in the road that AI’s biggest labs are now navigating.
Anthropic is betting that models themselves are the business. Its revenue growth has been propelled by enterprise customers who use Claude for coding, customer service and internal automation, and the company has been named by enterprises as the reliability pick over its larger rival. To secure the compute its models need, Anthropic has signed a series of contracts that strain belief: roughly $45 billion over six years for 460 megawatts of capacity at Nscale’s West Virginia data center, running on NVIDIA’s next-generation Vera Rubin systems; $10 billion over six years with Volta for a facility in Norway; and an agreement with SpaceX that provides about $1.25 billion worth of compute capacity each month. The pattern is consistent — Anthropic is buying certainty of supply at any price, because its business model depends on never being without a GPU.
OpenAI is walking a different path. Its own chip program, now acknowledged publicly for the first time on a firm schedule, is an attempt to control the most important input in its business rather than rent it. The company’s CFO describes the device program as part of a strategy to own the experience from silicon to software — an approach that looks less like a chip company’s plan and more like Apple’s: design the hardware, control the platform, and let the ecosystem form around it. OpenAI has also begun its own IPO process, with Chief Executive Sam Altman confirming a filing on national television this month.
The two strategies reflect different readings of the same market. Anthropic’s leadership argues that model capability is the defensible asset: if Claude keeps improving, customers will keep paying, and the hardware underneath is a commodity that can be rented. OpenAI’s leadership appears to believe the opposite: that in a world where several labs can train comparable models, the advantage moves to whoever controls the device, the distribution and the data flows. One company is buying compute; the other is building it.
The IPO stampede is putting both theories to a public test. Three frontier laboratories have filed registration statements with the SEC in roughly forty days, with a combined private-market valuation of about $2.84 trillion, according to one count. That concentration of listings — all from companies that have never reported meaningful profits — is unprecedented in the history of technology finance. The bankers are mostly the same five firms, and the pitch decks share a vocabulary: growth, infrastructure, intelligence as a service.
Investors are being asked to swallow the arithmetic that Amodei defended in her roadshow. Anthropic’s revenue run rate climbed from about $9 billion in late 2025 to more than $47 billion by the spring, a pace that few software companies have matched at any stage of their lives. The company’s costs, however, are locked into contracts of similar scale, and its largest customer relationships depend on partners — Amazon and Google — who are also its rivals in the model race. The same is true, in different form, for OpenAI, whose relationship with Microsoft has cooled into open competition.
The divergence the two companies embody will be decided by the markets that now have to price them. Anthropic is asking investors to value a pure model business on the assumption that intelligence is a product. OpenAI is asking them to value a hardware-plus-software business on the assumption that intelligence is a platform. Both claims cannot be right at the current scale; both might be.
What is clear is that the capital markets have become the final arena of the AI race. The labs that could not go public a year ago because they had no revenue large enough to describe are now filing because their revenue is large enough to count — and their bills are too. Amodei’s roadshow answer to the returns question was a promise that the model business pays. Friar’s hardware timeline was a promise that the platform business pays. The public market will collect on both promises, quarter by quarter, starting this year.
