Dario Amodei, Anthropic’s chief executive, introduced his artificial intelligence company to potential stock market investors this week with a sublime vision. He predicted the end of cancer and mental illness within a decade, and a transformation of cognitive work. “If we do this right, we can create an age of miracles,” he said in a marketing video for Anthropic’s initial public offering.
Wall Street is grappling with how much it may cost to get to utopia.
Stock analysts at Goldman Sachs, the investment bank leading Anthropic’s I.P.O., have told investors privately this week that they forecast the company will burn through more than $65 billion next year, and another $41 billion in 2028, said two people familiar with those discussions. Those totals would make Anthropic’s spending among the most aggressive corporate bets on funding the giant A.I. infrastructure build-out.
“We are in aggressive investment mode,” Krishna Rao, Anthropic’s chief financial officer, said in the video, without discussing specific spending forecasts.
That spending would create a drastically different financial picture than Anthropic has shown recently. The company generated cash from its business in the first nine months of this year, its I.P.O. prospectus shows, a feat in the expensive business of A.I. model development. It has barely spent the $141 billion it raised privately from investors.
The listing on Nasdaq, expected next Thursday, could value Anthropic at $1.4 trillion. It would not only cement its status as one of the world’s most prominent companies but hand it additional firepower to help it pull further ahead of well-heeled rivals like OpenAI and Alphabet. The I.P.O. would deliver Anthropic an additional $100 billion to spend, and make it easier to raise even more money to fund those huge commitments.
But Anthropic’s public debut will lead to costs that don’t appear today. It said in the prospectus it anticipated about $40 billion of expenses in the coming years tied to preexisting stock grants it has given employees that haven’t yet hit its accounting books.
The company’s investor prospectus pegged the size of the future bill that Anthropic has run up for chips, power and other computing infrastructure to help it develop more advanced A.I. models at $900 billion over the next two decades.
Stock pickers and analysts said the prospectus is missing more granular details about the spending, like how much is going toward A.I. training and research, serving customers and the company’s own internal uses of its A.I. service Claude.
“There’s definitely a desire to keep the focus on the massive, broad huge opportunity ahead of us, rather than go into the specifics,” said Ashley MacNeill, managing director at the private equity firm Vista Equity Partners.
Ms. MacNeill, a former investment banker who worked on high-profile tech I.P.O.s, added that she didn’t think Anthropic was “trying to hide the ball” but said the lack of disclosures “creates a vagueness on how they will run this company from a cost perspective going forward.”
Investors are bracing for Anthropic to run up the bill regardless. The size of the existing financial commitments the company laid out “is likely going to be meaningfully short of what Anthropic ultimately needs,” analysts at the investment bank William Blair wrote to investors this week.
The report characterized Anthropic’s pace of constantly releasing new versions of Claude as a “model treadmill.” The analysts wrote that the company’s potential valuation was attractive but added that “we empathize with investors that call out the risk of a company in a 4-year-old industry.”
Part of investors’ challenge is determining exactly what kind of business they’re signing up to support. Anthropic’s ascent has been dizzying. In one year, the company grew quarterly sales from about $1 billion, roughly the size of the dating app Tinder’s parent company Match Group, to $17 billion, the size of more-than-century-old stalwarts like Pfizer and IBM.
Anthropic executives have stressed that they have a lot more room to grow. Its focus on corporate customers, rather than consumers, has helped it leapfrog the ChatGPT maker OpenAI in sales this year. Its success has largely been thanks to heavy adoption by software developers. Next, they’re going after a wider range of white-collar employees to automate more of a variety of work tasks.
“To most workers, Claude is still just a chatbot,” Nicholas Lin, an Anthropic employee, said in an interview on Monday. The company is trying to change that. On Tuesday, it introduced new tools for people to include Claude as a participant in a work meeting or a spreadsheet.
The company will need those products to catch on. Goldman Sachs, in its private forecast to investors, predicted that Anthropic would surpass $440 billion in annual revenue by 2029, more sales than Apple generated last year. That surge would help Anthropic generate enough cash to cover its costs by that year, according to the forecast. The Financial Times earlier reported on the investment bank’s forecast.
But Anthropic’s customers aren’t locked into long-term contracts, and the vast majority of the company’s revenue is generated by versions of its A.I. models it released in the last six months, according to the prospectus.
Shivaram Rajgopal, a professor of accounting and auditing at Columbia Business School, said that could be a problem if Anthropic can’t gain enough customer loyalty and keep prices high.
In a worst-case scenario, the company could have similar business dynamics to Netflix, which has to continue spending heavily on producing shows to keep subscribers. “That’s a business model risk,” Dr. Rajgopal said. “These are slow-moving disasters.”
Anthropic, for its part, said the computing bills to develop new A.I. models will pay off. The spending, it argued, creates a reinforcing cycle that will help it keep its lead in the A.I. race, and eventually generate profits. Executives say customers have kept spending on Claude because its new model versions keep getting better, and the company can also keep costs down for customers and itself as models improve.
The company considers itself the ultimate beneficiary of its own creation, as it uses Claude to automate and speed up its model development.
Anthropic published a report last month that suggested its A.I. agents, or tools that can complete work tasks like writing code or creating spreadsheets, worked six times as many hours in a week as its human employees. “Our pace of discovery is speeding up, and a lot of that acceleration comes from Claude itself,” Sam McCandlish, an Anthropic co-founder and chief architect, said in the investor video.
Many investors weighing the deal haven’t seen Claude transform their work in the same way. “There’s an anxiousness and disconnect you see within A.I. believers and nonbelievers,” Ms. MacNeill said. “We’ve only sat with this technology for a few years.”