How Anthropic Makes Money—and What Investors Should Know Ahead of Its IPO Filing

It’s not every day that a company at the center of arguably world-changing technology goes public. The initial public offering from artificial intelligence giant Anthropic, expected in late September or October, is one such occurrence. Anthropic appears on track to be the largest stock offering in history, topping SpaceX’s debut in June, and it could play a key role in how AI companies are measured and valued. The challenge for investors is that Anthropic was founded just six years ago (SpaceX is a 20-year-old company), and the business models for making money from AI are untested and changing rapidly—practically by the week. Since it’s a private company, information about Anthropic has been limited. Much more will become available with the Securities and Exchange Commission’s expected release of the firm’s S-1 form, which will formally launch Anthropic’s IPO. Ahead of the filing, we spoke with Harrison Rolfes, senior analyst at PitchBook, who has been covering Anthropic on its road to becoming public. In the first of this two-part Q&A, we look at how Anthropic makes money, whether its current fee structure is durable, and the critical question of the company’s margins. In Part 2, we will dive into the risks the company faces from data center opposition, AI models going rogue, how much it could be valued at, and what needs to happen to have that valuation make sense. Tom Lauricella: First, if you’re an investor looking at the Anthropic IPO, you need to understand the company’s business model. That means knowing its past and expected growth rates, and whether its revenues are repeatable and predictable. What do we know about how Anthropic makes money? Harrison Rolfes: Anthropic builds advanced AI models and assistants and sells access to them. To determine where Anthropic actually makes money, we need to look at its user retention and contract duration. How many people are consistently paying the same amount? If we can get those data points, both of which should be in Anthropic’s S-1 filing, we can determine the recurring payments on average that people spend. Right now, those numbers are all over the place, and we don’t even know how Anthropic tracks them. Lauricella: The big question for investors who have not been on the inside is: What do we know about Anthropic’s financials prior to seeing the S-1? Rolfes: We know that second-quarter revenue exceeded $11.5 billion, up from $4.73 billion in the first quarter and $787 million a year earlier. This means adjusted operating income has turned positive, which is unique for any company and unprecedented for an AI company. Of course, generative AI is new, so everything’s unprecedented, in a sense. But it shows these AI companies have solid enough growth to have positive operating income. The distinction is that the July $65 billion-plus run rate they reported annualizes current sales. That matters because it’s a forward-looking projection of where they stand, and [it’s unclear] whether they’ll hit those projections. If we break it down, we can determine how much Anthropic is making per month. The $65 billion ARR figure puts them at about $5.4 billion per month. Last time, they were at $47 billion, which would have put them at $3.9 billion per month. So in a matter of months, they grew a lot. What’s also interesting about the $65 billion run rate is that it’s a third of their 2028 projections, which means they are likely on track to hit the projections they stated earlier. So until the S-1 becomes public, we still lack the audited GAAP earnings, gross margin, cash flow, and customer concentration data. My big focus is on gross margin, but customer concentration is also important. That means determining how much of Anthropic’s revenue comes from its largest customers. Investors will also want to know how many customers are paying, how much usage is contractually committed, and whether customers consistently expand their spending over time. Lauricella: What do you think is driving Anthropic’s revenue growth? Rolfes: Claude Code is the biggest catalyst. Claude Code exceeded the $2.5 billion run rate back in February. Then you combine that with the enterprise customers, which together generate more than half of Anthropic’s revenue. You have the enterprise mix, and you have Cloud Code, which has both individual and enterprise users paying for it. The [enterprise] business shows that Claude is moving into production-level API workflows; it’s not just being used in isolated experiments. More customers are deploying more agents to complete more steps, and consumers are using more tokens per task. And as tokens per task increase, the associated input and output costs go up, and that flows into Anthropic. Lauricella: What would be your sense then of how recurring this revenue is? Rolfes: It looks repeatable but not necessarily contractual. Enterprise accounted for roughly 80% of revenue. More than 1,000 customers were spending over $1 million on an annualized basis as of April. Lauricella: How much of Anthropic’s revenue model will be similar to software companies, which have predictable revenue streams based on licenses and seats? Will Anthropic be completely different? Rolfes: I think we will see the removal of the subscription service. First, Claude’s model shifted toward usage-based consumption where users ask it to do a bunch of things, it provides some output, and they pay whatever API calls and costs are associated with that. That’s just a usage model. We don’t know how many people are using the model, where this revenue is coming from, or whether that revenue is durable and recurring. Lauricella: That was also a source of some of the cost issues that were coming up. Corporations were saying, ‘Well, wait a minute, we went from like zero to $10 million,’ because it’s just based on whatever the individual developers are doing. Rolfes: Uber is a good example where they supposedly had that crazy bill, and they said, ‘What are we doing? That’s way too much.’ So Anthropic needs to find a way to structure a durable revenue model that’s easy to track. And I think the way to do it is to attach a cost per task requested, and then you have to reduce the focus on the response from the model based on the industry. For example, if you are a financial services company and you’re using Claude Code, you probably don’t need to know medical vocabulary or how to do procedures. You just want financial data. That leads me to believe the usage base is going to go to a completed task model, in which you pay for exactly the task you’re hoping to get out of it. Lauricella: We’re seeing revenue numbers showing growth, but we also know that there’s downward pressure on prices. What do you see in terms of Anthropic’s ability to outrun the impact of falling AI prices? Rolfes: I think Anthropic can outrun falling inference prices if Claude usage grows faster than the price-per-token declines. Vercel, a software company that helps developers make websites, is a good example. Vercel uses Claude internally and makes Claude available to developers through its platform. That relationship expands Anthropic’s distribution and creates more usage. AI may become cheaper, but the market could grow much faster. Vercel’s July production data showed token volume increasing 59% and spending rising 37%, even as the average price per token declined 13.6%. This is called the Jevons paradox: If you lower the price of something and make it more accessible, people actually consume more of it, so you end up spending more. If token prices continue going down, I think it actually ends up balancing or increasing revenue for Anthropic because they still capture all of it at the end. Lauricella: What does it cost Anthropic to generate this revenue? Rolfes: We know what customers pay, but we do not know what it costs Anthropic to provide the service. Claude Opus 5 charges $5 per million input tokens and $25 per million output tokens. Fable 5 charges $10 and $50, respectively. Those are customer prices, not Anthropic’s underlying costs. That’s essentially where the revenue comes from. But then the cost of that revenue is tricky, because you have to account for compute cost, inference costs, failure rates, and all the other mechanics that go into the model searching for your query, as well as how long it takes to generate an output. That is something I think the best minds in the world are still trying to piece together, because you need to go granular to find how much power the GPUs pull, how much time it takes for the model to utilize the GPUs to extract the information, and how much power is consumed during that process. Then you need to spread that out across the entire cohort of people who actually use it. Those numbers will shed light on the gross margin aspect, because we will know that it costs X amount to produce a result. And based on the models I’ve worked with and based on Claude Opus 5’s listed API price, customers pay roughly 0.0033 cents per generated word. Lauricella: What do you think Anthropic’s gross margins look like? Rolfes: I wrote a note on this , and I estimated them to have about a 44% gross margin. That’s accounting for anything and everything I could piece together. That number seems realistic, but it is not enough to compete with SaaS companies. And it isn’t really attractive to public market investors, because we know Anthropic has too large of commitments that extend out to 2030-32. To survive those commitments, you need to bump that gross margin up to like 70%. Lauricella: Putting that in perspective, Microsoft’s gross margins are near 70%, and ServiceNow’s are also right around 70%. But of course, Anthropic is still very much in building-and-spending mode. Rolfes: Yeah, they’re definitely spending, but what’s interesting about these AI frontier models is that to continue to grow the model, you need to spend exorbitant amounts of money. Not only on compute, but on training failures, the employees behind it, operational costs, and everything else tied to it. At the same time, if they don’t continue making money on their current trajectory, there’s no way they’ll be able to afford the additional compute that comes when they try to enter bigger markets like financial services and healthcare and whatnot. If you want to multiply your cohort base by 10, how are you going to get the compute to supply all that? You’ve got to spend a lot of money. The path to margin expansion becomes a major part of the IPO thesis. Coming up in Part 2: A look at the risks that investors will need to consider and how much Anthropic could be worth. The author or authors do not own shares in any securities mentioned in this article. Find out about Morningstar’s editorial policies .