《经济学人》分析：AI安全恐慌下 Anthropic 约2万亿美元 IPO 或需修订 S-1、下调估值甚至推迟，Altman 称 OpenAI 今年不上市

Should Anthropic press pause on a potential $2trn IPO?
Amid a safety freakout, it wants to slow down AI development
THROUGHOUT ITS five-and-a-half-year history, Anthropic, a safety obsessed artificial-intelligence lab, has grappled with the tension of how to build AI safely while making enough money to push forward the frontier of AI research. On September 12th that tension came to a head. Dario Amodei, its boss, called on the industry to “slow the pace” of AI development, just as his firm was believed to be on the cusp of launching what may be the biggest initial public offering (IPO) of all time.
His call for a slowdown, supported by other bosses of AI labs, such as OpenAI’s Sam Altman and Elon Musk of SpaceXAI, followed a week in which angst about potentially catastrophic risks from AI broke out from the echo chambers of Silicon Valley to evening news broadcasts and dinner-table conversations. They were triggered by warnings from researchers at Anthropic that there were credible fears within the company that AI could wipe out humanity in a matter of years.
It is unusual, to say the least, for people within a firm to publicly air such doomsday warnings about their most cutting-edge products. It is doubly so when seeking to entice investors to back an IPO that reportedly aims to value the company at about $2trn—about as much as the ten highest previous Silicon Valley IPOs added together. If the warnings come true (and assuming a few lawyers survive the calamity), Anthropic could face what David Sacks, Donald Trump’s former chief AI adviser, has called the “mother of all product-liability lawsuits”.
That has led to speculation that, at a minimum, Anthropic will have to amend the confidential S-1 paperwork it has filed with the Securities and Exchange Commission (SEC) ahead of the IPO in order to address the fallout. More seriously, its underwriters could reduce the valuation they hope to achieve, or postpone the IPO. Mr Altman raised the stakes on September 12th by stating in an interview with Fortune magazine that OpenAI, which has also filed paperwork for an IPO, no longer plans to list this year. “Given everything happening with safety, this would right now be an ill-advised moment to go public,” he said.
Mr Altman’s remarks were doubtless a cheeky dig at his arch-rival; OpenAI had previously let it be known that it was unlikely to be ready for an IPO until next year. And some Anthropic investors, such as Brad Gerstner of Altimeter Capital, which co-led the AI lab’s latest funding round, were quick to throw their weight behind Mr Amodei and his firm. “Anthropic will IPO. The market knows how to price risk,” he posted on X.
How quickly the firm goes public, and at what price, depends on various factors, however. John Coffee, a securities-law expert at Columbia Law School, says it is not unusual for companies to amend SEC filings prior to an IPO. This can be done quickly. He suggests that Anthropic, which has a governance structure designed to prioritise safety over profit, may have already sufficiently disclosed the risks in its S-1. Once its disclosures are signed off by the SEC, it limits Anthropic’s liability for securities fraud if disaster strikes, he says. Product liability is a different matter, but the firm is likely to have also disclosed the risks of lawsuits in the S-1.
The market may force Anthropic’s hand, however. The firm will want to see how the furore affects the share prices of listed AI firms over the coming days, Mr Coffee says. If AI is under a “deep, dark cloud” and the sector as a whole suffers, it may prompt a serious rethink about the IPO. A modest fall in stock prices might not set the IPO back for long.
Supporters of Anthropic say Mr Amodei’s long history of warning about AI risks offers reassurance. In an essay published over the weekend, Anthropic’s boss said the triggers for the latest outpouring of concern were twofold. First, AI capabilities are compounding faster than the industry’s understanding of how to control them. That, he wrote, is primarily because of “AI’s growing ability to build the next generation of AI.” Second, several recent incidents of rogue AI agents carrying out illegal hacks, such as the attack by swarms of OpenAI agents on Hugging Face, an artificial-intelligence platform, could be harbingers of bigger dangers to come.
Mr Amodei outlined several ways the industry could slow the pace of AI model development to ensure their safety. He said Anthropic would give third-party evaluators the same access to its models as Anthropic employees. They would then be able to verify safety practices and report incidents. Mr Altman posted that OpenAI would do the same.
Mr Amodei also proposed regulation of frontier AI models in America. Until that happens, he called for voluntary co-operation between the labs, with a legal waiver from America’s government to prevent the discussions triggering antitrust concerns. In addition, he pushed for a broad international regulatory regime, encompassing China.
President Trump shrugged off the AI safety concerns on September 13th, saying that outpacing China in building AI was the most important thing. “Whoever wins AI wins,” he told reporters on a trip to Ireland. Mr Sacks, his former adviser, also made light of the need for regulation, arguing that if Anthropic and OpenAI believed it was important to slow down model development, they should go ahead and do it themselves.
None of Mr Amodei’s proposals would seriously hurt Anthropic’s commercial prospects. Slowing progress on the AI frontier is likely to have only a minimal impact on revenues, because its models already possess more than enough intelligence for the business users that provide most of its income. A pause might even be of benefit, if open-source model-makers, who are winning business from Anthropic and OpenAI by providing cheaper AI models to enterprise customers, are also forced to slow down.
But potential IPO investors, many of whom seem cocooned in an AI bubble, should not take the latest safety panic lightly. If it adds to fears, widespread in the West (less so in Asia), that AI will wipe out jobs, drive up electricity prices, and widen the gap between haves and have-nots, it may increase political pressure to temporarily halt the technology’s development and put a freeze on data-centre construction. Even those not worried about AI models annihilating humans should worry about anxious politicians trying to annihilate AI. ■