With Anthropic’s IPO looking likely for late September or October, we spoke with PitchBook senior analyst Harrison Rolfes, who has been covering the company as it gets ready to go public.

In the first part of this Q&A, we looked at how the artificial intelligence giant makes money, whether its fee structure is durable, and the critical question of its margins. In Part 2, we dive into the risks Anthropic faces from data center opposition and AI models going rogue, as well as how much the company could be worth.

Tom Lauricella: Anthropic is growing fast. What hurdles does the company face on the economic or regulatory fronts?

Harrison Rolfes: I don’t think regulatory is a hurdle, per se. Regulatory just establishes policies that they have to follow so that they fit the standard in whatever country they’re working in. The biggest hurdles have to do with compute (processing power and memory) and energy. They need such a significant tap into data centers and the power grid in order to run a lot of these models.

The second hurdle is the need to maintain a level of trust and accuracy so that people continue to use it. The third hurdle is to become integrated so deeply into these [client] enterprises that it would be almost impossible to switch, or it’d be impossible to continue working without using it.

Lauricella: In other words, Anthropic has to build what Morningstar would call a wide moat, right?

Rolfes: Anthropic needs to win enterprise customers quickly, expand into more of their workflows, and create switching costs by becoming embedded in how those customers operate. That requires sustained access to capital and compute from infrastructure partners such as Amazon and Microsoft.

In regulated industries like financial services and healthcare, the next test is whether Anthropic can turn its existing relationships into large, repeatable deployments. If it can do that while maintaining strong performance, trust, and improving economics, it can build the durable enterprise position that public investors would recognize as a moat.

Lauricella: One worry that that keeps coming up are all the financing deals between the hyperscalers and the big AI Labs like Anthropic. What are you going to be looking for on that front?

Rolfes: We are going to see a new type of bespoke economics come into play where we’re seeing the hyperscalers provide debt financing structures for these AI companies. That benefits the hyperscalers. They’re essentially financing their customer to continue to fund [the hyperscalers’] business. That the money just continues to circulate. They’re also putting up their GPUs as collateral, which is just crazy. Even though GPUs depreciate, they’re still a revenue-generating item.

Similarly, OpenAI and Nvidia are already doing this. Nvidia is essentially backstopping massive debt vehicles for OpenAI to go build whatever they want. In return, OpenAI will purchase Nvidia chips. This kind of situation changes the economics of how these large-scale deals get done. I think we’ll start to see more of that.

However, how many more of those deals can you do as Anthropic? Do you rely on a hyperscalers as a bank and avoid traditional financing? Or do you find additional revenue sources that allow you to build out more data centers? Do you find alternative energy sources to feed the power consumption that’s required?

Lauricella: Meanwhile, Nvidia is also reportedly developing its own open-source AI models to compete directly with Anthropic. How should investors think about that?

Rolfes: Nvidia can offer powerful open-source models at a low cost. That’s because its primary goal is selling more chips, which could make basic AI models cheaper and put pressure on Anthropic’s pricing. For investors, the key question is whether Anthropic (and OpenAI) can remain differentiated through better products, enterprise relationships, and trusted performance. Not simply a better model.

Lauricella: We’re seeing growing opposition to data centers around the country. It’s becoming a big election-year issue. At the same time, reports say that in the next year, there won’t be as much new capacity coming online than had been expected. How much room for error does Anthropic have in terms of data center capacity expansion?

Rolfes: Anthropic has more room for error than a smaller AI lab because it has deliberately diversified its compute across Amazon’s AWS, Google, Microsoft, SpaceX, and other providers.

That said, there’s a distinction between capacity that is contracted and capacity that is actually powered, permitted and online. Delayed projects—whether by local opposition, grid interconnection, or construction bottlenecks—become both a growth problem and a margin problem for Anthropic. It either has to limit usage, delay new customers, or buy scarcer and potentially more expensive compute elsewhere. So I would say they have some cushion, but not unlimited cushion.

At current growth rates, compute can be a bottleneck. If the compute capacity is not there, Anthropic cannot sell the intelligence. In the S-1, I would want to know how much of its 2027 and 2028 capacity is already contracted, how much is actually under construction, when it is expected to be energized, and what happens financially if those projects arrive late.

Lauricella: It seems like not a week goes by without news of an AI model going rogue or new evidence that this is a serious, present-day concern. For investors in Anthropic, what kind of risk does this present, and do you expect the company to address it?

Rolfes: I think this is a real investment risk, but I would frame it less as a science-fiction ‘rogue AI’ problem and more as an operational risk that gets larger as agents receive more autonomy and access to real systems. As these models can write code, use computers, and take actions on users’ behalf, the potential blast radius of a failure increases.

For investors, that can show up as legal liability, tighter regulation, higher security costs, slower deployments, or damage to enterprise trust. Anthropic should have more to say about this than almost any other AI company because safety is central to its brand. It already has a Responsible Scaling Policy, risk reports, and containment controls. In Anthropic’s S-1, investors need to look for disclosure around material AI-safety incidents, governance, testing and monitoring. They also need to keep an eye on whether stronger safeguards could ever slow product releases or limit how the company monetizes its most capable models.

Lauricella: Pulling this all together, what’s your take on what Anthropic is actually worth and what investors will need to see to justify the IPO valuation?

Rolfes: Right now, [based on their most recent funding round], Anthropic is worth $965 billion. It has been reported that they are going to aim for a $2 trillion IPO. If that’s the case, that’s roughly 31 times July’s annualized revenue. That’s over ten times Anthropic’s projected $190 million to $200 billion of 2028 revenue.

Now, to justify that price, first, investors must believe that Anthropic can convert today’s surging consumption into durable recognized revenue, which we discussed before. Second, Anthropic must expand its margins. As I said, I estimate they are at 44%, and they’ve got to get up to 70%.

We also need to see strong user retention. There’s always a back and forth between Anthropic and OpenAI. [Anthropic needs] people to choose one and say, ‘OK, we’re going to stay here, and this is going to be our go-to.’ And then we need to see free cash flow that is not swallowed up by the next generation of compute.

Lauricella: To wrap things up, what will be the key points you will be watching for when Anthropic files its S-1?

Rolfes: On the accounting side, I’m very interested to see how they calculate their true cost of revenue and any revenue-sharing or cloud-partner arrangements, because if you make the comparison between Anthropic and OpenAI, let’s say they both make $100 a month. OpenAI has this 20% revenue share with Microsoft, which means that they really make $80. Anthropic has a similar arrangement, but they’re not counting that 20% off the top off their top line revenue. We don’t know why. I will be interested to see what falls below the top line revenue and why it’s not an apples-to-apples comparison between Anthropic and OpenAI in terms of the overall accounting and finances.

Most of all, I want to see a clear-enough disaggregation of the costs to understand the economics, rather than having major cost categories blended together. At the end of the day, I care about gross margin, how many people are using Claude, how many are paying for Claude, and what it costs Anthropic to serve that usage.