A man asks a chatbot for help talking to his mum. It gives him a bit of reassuring advice, then recommends a dating app for meeting older women. That was the joke in one of Anthropic’s Super Bowl campaign spots in February. There was a good joke outside the frame, too. An AI company had bought space in America’s most famous ad break to tell people that some places ought to be free of advertising.
Anthropic promised to keep Claude, its AI assistant, ad-free. Nine months later, it’s preparing to sell something else. In a preliminary prospectus dated 10 November, the company set out a proposed offering of 825 million shares at $70 to $78 apiece and said it had applied to list on Nasdaq. Investors are being invited to put a price on a business that’s made a virtue of drawing lines around how it makes money and what it will put into the world.
The people most likely to put a lasting value on Anthropic’s promises may never buy a share. They’re the people deciding how much of their company’s work Claude gets to do. Their decisions will show up in the numbers. The reasons for those decisions, especially when another model can do the job for less, will be much harder for Wall Street to read.
The money is in what customers do after the trial
For most of the AI boom, companies have competed to show that their model can do something the others can’t. The demos can be genuinely impressive. But a company buying AI for its staff may have a narrower question. If several tools can already take an hour out of the weekly reporting job, which one is the head of operations prepared to let read the underlying documents?
Dylan Stuart, Senior Partner, Strategy at Lippincott, sees an investment case in that change of emphasis.
“Silicon Valley is spending billions making AI smarter, faster and cheaper. As Anthropic prepares for its potential IPO, investors face a rather more interesting question. What if the most valuable AI company isn't the one with the smartest technology, but the one businesses are most comfortable trusting? There's a serious economic argument here as AI capabilities converge, intelligence risks becoming a commodity. And commodities, however miraculous, rarely command premium valuations for long.
For investors, the question is whether that trust translates into something tangible: greater enterprise adoption, deeper integration, stronger retention and ultimately more predictable earnings. If it does, Anthropic could justify being valued differently from competitors fighting over model performance and price.”
Stuart’s argument depends on which jobs are becoming interchangeable. A modest difference between models might be immaterial when you’re asking for ten headline options. A substantial difference in how well a model handles a company’s codebase may keep buyers paying for performance. In between lies a great deal of ordinary work where the choice increasingly comes down to whether a company can get a tool approved, introduce it without a revolt and still have someone to call when it behaves strangely.
Think of a studio that starts with a handful of Claude accounts to summarise meeting notes. A few months later, producers want to connect it to their project folders so it can find decisions buried in six months of correspondence. That’s when the studio asks about permissions, how long data is kept and what happens when someone leaves the team. A reputation that people are willing to act on can help the person who backed the trial make the case for giving it more work.
The prospectus puts a number on some of that expansion. Anthropic reported net revenue retention of roughly 900% as of 30 September. On its measure, a qualifying group of existing customers brought in about nine times as much revenue over three months as the same group had in the equivalent period a year earlier. A handful of big customers spending more can do plenty of lifting in that figure. So can a coding tool that suddenly becomes indispensable. Over time, the stronger evidence for Stuart’s argument would come from customers who keep expanding their use even when they have a credible alternative and the option to move.
A studio might happily pay more per seat if a common set of controls means it can bring another client account onto the system without a month of paperwork. Anthropic has its own costs to cover. The filing puts its computing and infrastructure expenses at $18.8 billion for the first nine months of 2026. A lasting premium would depend on how much customers will pay for the work Claude does and how much it costs Anthropic to deliver it under the conditions that won their business.
A promise customers can check for themselves
The Super Bowl campaign made one of Anthropic’s promises unusually easy to check. A sponsored card turning up in a Claude conversation would be hard to explain away. Subscriptions and usage fees give Anthropic every reason to persuade a business to give Claude more work. The pledge sets a limit on the kind of business it’s prepared to do inside the conversation itself. You can see why that’s appealing in a product people use to think through decisions they haven’t yet discussed with their colleagues, let alone anyone trying to sell them something.
Andy Bacon, Founder & Partner at Manifesto Growth Architects, puts the emphasis on whose interests the company serves when the stakes rise.
“Anthropic has developed incredible technology, but the element at the heart of their market value is the question of trust. Customer trust has always been fundamental to successful brands, but it becomes even more critical when the stakes are this high – both in terms of financial value and the impacts of AI on society.
Trust is built on three factors: Credibility - can people trust what you say; Reliability - do you keep your promises and; Intimacy - do people feel safe with you. These are divided by the key thing that erodes trust – Self-Orientation - are you focused on personal gain or the greater good. Harvard call this the ‘Trust Equation’.
So far, Anthropic has done the right things to build trust. They made safety their positioning from day one and talk openly about what could go wrong with their own product.
Becoming a listed company will bring significant new tests. Public markets reward quarterly growth. Anthropic's brand is built on sometimes slowing down. Every safety decision will be read through the share price, and every commercial decision through safety promises.
Anthropic has delivered on the trust equation so far. The big question is can they stick to it.”
The Trust Equation first appeared in The Trusted Advisor, by David H. Maister, Charles H. Green and Robert M. Galford. In an adviser’s relationship with a client, intimacy means feeling safe enough to share something sensitive. In an agency, it might mean knowing what happens to the unreleased campaign deck someone just dropped into a chat window. Anthropic excludes commercial customers’ inputs and outputs from model training by default. If the answer you give depends on which account a freelancer happened to use last Tuesday, though, it could be a long afternoon for the producer who has to explain the agency’s use of AI to the client.
Safety gets interesting when it holds up a launch
A promise about what happens inside a research lab is harder for customers to verify than a promise about advertising. In February, Anthropic revised its Responsible Scaling Policy, the framework it uses to manage risks from increasingly capable AI. It moved away from an earlier commitment to pause training beyond certain capability thresholds until it could put the required safeguards in place. The new version added a public safety roadmap and regular risk reporting. Anthropic’s argument was that some safeguards would need to be adopted across the industry to work and that it needed commitments it could realistically meet on its own.
The change gave the public more to inspect while giving Anthropic greater discretion over when to stop. How you weigh those two things depends in part on whether you think the company will still be willing to delay work when competitors are pressing ahead. In its prospectus, Anthropic says it recently postponed the start of training a new model by more than a month to reduce the risk of the model being stolen. That gives some substance to Bacon’s point about slowing down. A delay will be harder to explain once a launch date has made its way into someone’s revenue forecast.
There’s also a governance mechanism behind the branding. Anthropic is a Delaware public benefit corporation, and its prospectus says its Long-Term Benefit Trust generally has the power to elect and remove a majority of board members. That gives the trust a formal say in who oversees the company as new shareholders come in. The consequential decisions will still come one at a time, often with imperfect information and an attractive reason to ship on Friday.
Whose trust counts?
There’s another constituency that tends to disappear when an AI company talks about earning trust. In July, a US federal court gave final approval to a $1.5 billion settlement in Bartz v. Anthropic, a class action over books the company downloaded from pirate sites. In an earlier ruling in the case, the court found that using books to train models qualified as fair use on the facts before it. It treated building and keeping a library of pirated copies as a separate use and rejected Anthropic’s fair use argument for that practice.
For people who make a living licensing their work, that distinction is worth keeping. An agency can feel confident about the protections for its own client files while an author has good reason to ask how their work got into the model’s training material in the first place. Both have a stake in what AI companies do with other people’s work. Only one of them gets to ask questions in a procurement meeting.
The debate over ownership and creative credit puts pressure on Bacon’s question about whose interests come first. The businesses buying the software can reward a vendor that looks after them. The people whose work helped make it possible have to rely on different means to be heard. A brand that puts responsibility at the centre of its identity invites scrutiny from both.
What Wall Street will get to see
There’s a generous reading of what going public could do for Anthropic. Fresh capital could pay for the people and computing power needed to do more safety work. Regular disclosures could make it easier to see whether customers keep giving Claude bigger jobs after the initial enthusiasm wears off. Investors who take a long view may even welcome a decision that costs the company a month if it protects a relationship that could last years.
The tension is likely to show up in much smaller decisions than the IPO itself. How much detail does Anthropic give customers when it changes a policy? How quickly does it tell them when a safeguard fails? A team that used Claude for brainstorming can probably take a day off from it. A team that’s put Claude inside its production workflow needs to know who’s dealing with the problem by the time the morning status meeting starts.
Creative businesses face their own version of this. As clients ask what they’re paying an agency for when AI can make the work, part of the answer is having someone prepared to call and say the work needs another day, or that the cheapest way of making it puts a client asset somewhere it shouldn’t go. You can promise that sort of judgement in a pitch. Clients find out what it’s worth on a Thursday afternoon when a deadline is slipping and the person on the phone could very easily have let it slide.
The first day of trading would give Anthropic a share price anyone could look up. I’d be more interested in an earnings call where the company had to explain that a launch had slipped because a safeguard wasn’t ready. The investor who’d pencilled the revenue into that quarter and the agency deciding whether to connect its client archives would be listening to the same answer for very different reasons.
