What Anthropic's IPO Investors Want to Know

We’re getting ready for Anthropic to make its initial public offering prospectus public sometime after Labor Day. Though the company’s private investors have had a window into its torrid growth and projections, this will be the first time public investors are getting a full financial picture of the Claude maker.

Just how complete a picture remains a big question, though. We’re hearing that investors want a lot more from Anthropic than the financial statements that securities regulators generally require companies to report.

In a series of meetings, executives at Anthropic have been going back and forth with potential large investors on whether the company can provide other business- or performance-related metrics particular to AI. These include AI tokenomics, including what it makes off every AI token it processes, how much revenue it’s generating per gigawatt of compute, and net revenue retention, a measure of a tech vendor’s ability to keep customers and increase their spending. There’s no guarantee Anthropic will provide any of these.

The investors’ questions reflect the anxieties hovering over the AI boom. While revenue growth has been off the charts and Anthropic is even turning an operating profit by some measures, investors are worried it won’t be able to sustain that growth—particularly as businesses tap open-source models to cut costs. Investors are also in the dark about how much in total Anthropic is spending on its compute and what a steady drip of deals with SpaceX, Google and others will do to its margins.

Of course, Anthropic doesn’t need to legally disclose such metrics as token prices or usage, according to IPO lawyers. SpaceX, which had transformed itself into an AI and cloud computing company in the months before its June IPO, didn’t disclose them, either.

But answering these questions could dispel some of the concerns lingering over what could be a record-busting IPO. And more significantly, what Anthropic says now will set a benchmark for what OpenAI and other AI companies disclose.

Anthropic and OpenAI spokespeople didn’t have comments.

In fact, advisers to OpenAI’s IPO, now expected for next year, have been talking about how much they should reveal about similar metrics, according to a person with knowledge of the discussion.

High on the wish list for some investors is an understanding of how many tokens—or bits of code or language—Anthropic is processing and, more specifically, how much revenue it is generating per millions of tokens processed, net of discounts. They also are eager to hear about its costs for serving tokens.

Investors say these numbers can help them measure the unit cost of Anthropic’s business better than its gross margin (sometimes called inference margin), which reflects how much it’s spending to run—but not train—the models. Revealing how much Anthropic is spending or making for every unit of AI could help investors understand its return on every dollar of investment in compute.

(Some other public companies like Google have been sharing with their investors the monthly gross tokens processed.)

Similarly, if Anthropic discloses a net revenue retention number, investors would be able to compare it to the enterprise software incumbents its business is threatening (though there are plenty of ways to cherry-pick the inputs of this metric to make it look strong.)

However, it may be hard for Anthropic to come up with a representative revenue retention metric. Anthropic does sell subscriptions for its Claude chatbot, but a lot of its revenue comes from sales of Claude through its application programming interface, which don’t generate the type of renewals that show up as recurring. And Claude’s API use may be uneven month to month, as developers switch models for specific tasks and use Claude less or more in a month.

Anthropic could provide investors with other breadcrumbs designed to showcase its enterprise heft. It’s likely to tell investors how revenue is divided between different cohorts of enterprise customers that spend from $100,000 to millions of dollars, according to people familiar with the matter. Such disclosures are standard for large enterprise companies, like Snowflake and Datadog.

Investors also want to see how much of its revenue is coming from its latest models rather than older ones, as they can gauge whether the cost of research and development is paying off for the billions Anthropic has spent to train the next best models.

Then there are Anthropic’s compute costs. As the company moves to control its own servers for developing and running AI, and to eventually make its own custom chips, investors are hoping it will share how much revenue it generates per gigawatt or megawatt of power it has to pay for.

It’s not clear Anthropic will go to such lengths. Instead, executives have been touting how the company is extracting efficiency from its compute, especially compared with two years ago. Anthropic hasn’t been discussing much about its custom chip efforts or how it plans to split up its compute spending among big cloud providers such as Google or multitenant data center developers like Hut 8, investors said.

If they don’t get these answers from the prospectus, they can ask more questions in an investor day Anthropic expects to hold in mid-September, though the timing could still change.

But investors may want to temper their expectations. One large public investor summed up Anthropic’s approach to the markets as: “Don’t think too hard. Just look at the revenue growth rate. That’s all you need to know.”