Dario Amodei, the chief executive of Anthropic, was becoming a nuisance for his fellow technology industry executives.<\/p>\n
Anthropic and OpenAI, the leading artificial intelligence companies, were lobbying the White House to prevent some A.I. models from being developed and widely shared because, they argued, they were too dangerous. But companies like Nvidia and Microsoft saw those A.I. models as a big business opportunity.<\/p>\n
At the end of July, Satya Nadella, chief executive of Microsoft, and Jensen Huang, then Nvidia\u2019s chief executive, intervened. They emailed each other drafts of a letter saying those A.I. models should remain open, according to four people familiar with the discussions. When they asked other technology executives whether they\u2019d like to sign on, two dozen companies quickly agreed before the letter became public. Over the next week, 200 more companies \u2014 including a chastened OpenAI \u2014 signed onto the letter<\/a>.<\/p>\n Dr. Amodei refused.<\/p>\n Dr. Amodei has for years exasperated his peers with warnings about extreme risks of A.I., even as his company has built some of the most powerful A.I. systems in the world. Frustration has turned to hostility as Anthropic leaped into the lead in A.I. development, experienced stunning revenue growth and moved toward what is likely to be one of the largest initial public offerings ever later this month.<\/p>\n Along the way, Anthropic \u2014 which has cultivated an image as a company that tries to do the right thing \u2014 has also gained a reputation for playing hardball with customers and competitors. It has cut some off from its service and lobbied against other companies\u2019s free alternatives to its A.I.<\/p>\n Over the last year, other technology companies and investors have organized to push back, according to more than 20 people in the technology industry, many of whom spoke on the condition of anonymity to preserve their relationships with Anthropic.<\/p>\n They have tried to show a united front to counter Dr. Amodei\u2019s warnings about A.I. They have leaned on their ties to President Trump to keep regulators away. They have publicly complained about Anthropic\u2019s business practices. They have even spoken with another Anthropic executive about Dr. Amodei scaring people with his rhetoric.<\/p>\n But the message on A.I. safety from Dr. Amodei, who appears unbothered by criticism from other executives, is breaking through. A.I. agents going rogue and attacking websites have become increasingly common. And in the midterm elections, the threat of job losses from A.I. and data center construction were significant issues for voters<\/a> in both parties.<\/p>\n (The New York Times has sued OpenAI and Microsoft<\/a>, claiming copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)<\/p>\n The Schism That Launched Anthropic<\/p>\n Anthropic started with a split. In 2020, a group of OpenAI employees led by Dr. Amodei left the company amid disagreements over how to safely develop A.I. Those employees founded Anthropic in San Francisco the next year.<\/p>\n Its founders\u2019 ties to effective altruism<\/a> \u2014 a philosophy where donors try to measure and maximize the benefits of their philanthropy \u2014 created a foundation for the company, particularly its approach to A.I. safety.<\/p>\n But Anthropic played second fiddle to OpenAI, which made everyday use of A.I. tools mainstream with the release of ChatGPT in 2022.<\/p>\n That changed in February 2025, when Anthropic surprised competitors with a preview of Claude Code, an A.I. coding tool. Software engineers started using the tool to generate computer code<\/a> instead of writing it themselves. Anthropic\u2019s revenue exploded. <\/p>\n Then, its critics say, Anthropic got tough with customers.<\/p>\n A start-up called Windsurf had talks to sell to OpenAI, and those negotiations became public<\/a> in April 2025. Anthropic later cut off Windsurf\u2019s direct access to its models, which the start-up used to power its own A.I. coding tool.<\/p>\n \u201cI think it would be odd for us to be selling Claude to OpenAI,\u201d said Jared Kaplan, one of Anthropic\u2019s founders and its chief science officer, in an interview<\/a> in June.<\/p>\n The move crippled Windsurf and potential customers went elsewhere, according to three people familiar with the discussions. Even as Windsurf\u2019s talks with OpenAI broke down, Anthropic kept Windsurf cut off.<\/p>\n Windsurf struck a licensing deal with Google<\/a> in July 2025. But the message was delivered to the rest of the start-up world: Their businesses were being built on Anthropic\u2019s models, but they had to stay in the good graces of Anthropic to maintain that access.<\/p>\n \u201cSomehow, in the United States, we\u2019ve convinced ourselves that innovation should be in the hands of a few organizations,\u201d said Manos Koukoumidis, chief executive of the A.I. start-up Oumi, in an interview.<\/p>\n Last February, about a year after Anthropic previewed it, Claude Code was on track to make $2.5 billion in annual revenue<\/a>. In a financial filing, Anthropic disclosed overall revenue of $17 billion<\/a> in its most recent fiscal quarter.<\/p>\n \u2018Only They Should Do It\u2019<\/p>\n In early 2025, Dr. Amodei ruffled feathers at Nvidia by backing U.S. trade restrictions on chip sales to China. That June, Mr. Huang said he disagreed with Dr. Amodei on almost everything.<\/p>\n \u201cHe believes that A.I. is so scary that only they should do it,\u201d Mr. Huang said of Dr. Amodei and Anthropic at a conference in Paris. (Mr. Huang stepped down<\/a> from Nvidia on Thursday under pressure from his board of directors.)<\/p>\n Anthropic responded: \u201cDario has never claimed that \u2018only Anthropic\u2019 can build safe and powerful AI.\u201d <\/p>\n But industry leaders worried about public sentiment turning sharply against A.I. and asked Anthropic to tone down the doomsday talk.<\/p>\n Over the summer, Marc Andreessen and Ben Horowitz, of the venture capital firm Andreessen Horowitz, met with Tom Brown, a co-founder of Anthropic. Anthropic was scaring the public with its predictions of extreme job loss, Mr. Andreessen and Mr. Horowitz said, and should cool it down, according to three people familiar with the discussion.<\/p>\n Mr. Andreessen then took out his phone, where he had pulled up one of Dr. Amodei\u2019s television appearances. The text below Dr. Amodei\u2019s face summarized his prediction of a white-collar blood bath and mass unemployment. Mr. Brown appeared taken aback by the confrontation.<\/p>\n Then, last June, Anthropic released Fable<\/a>, a version of its most advanced model. The company said it would make the model less effective if it thought Fable was being used to research or develop A.I. But Anthropic wouldn\u2019t tell a user it was degrading the performance, infuriating some researchers.<\/p>\n \u201cIt is a very real sign of concentration of power that businesses can make such obviously user-harmful behaviors and still lead in the market,\u201d said Nathan Lambert, a co-founder of Trillium Labs, an open-source A.I. nonprofit, in a social media post<\/a>.<\/p>\n Within a day, Anthropic walked back the policy. Now, if it thinks Fable is being used to research or develop A.I., it switches to a worse model, and then tells the user.<\/p>\n \u201cWe made the wrong trade-off,\u201d Anthropic said at the time.<\/p>\n Anthropic also said<\/a> it would store the data of those using its most advanced models for 30 days \u201cto support our safety work.\u201d In a sign of how paranoid Silicon Valley had become about Anthropic, the change stoked fears that the company could train its models on customer data, and then compete with those customers, according to four people familiar with the changes.<\/p>\n In September, Anthropic started allowing customers to store their data on their own cloud computing services. Danielle Cohen, a spokeswoman for the company, said it treats customer data as confidential and, by default, prohibits training on it.<\/p>\n \u2018It\u2019s a Hoax\u2019<\/p>\n At the end of July, Silicon Valley\u2019s anger toward Anthropic peaked as Anthropic and OpenAI lobbied the White House to act against open-source A.I. models. Their campaigning came after OpenAI said its A.I. had gone rogue and hacked into other companies, including a start-up called Hugging Face<\/a>.<\/p>\n So Mr. Huang and Mr. Nadella shared the letter supporting open-weights A.I.<\/a>, which make the rules that govern a model publicly available. It opened the floodgates against Anthropic, according to five people familiar with the discussions. So many companies asked to sign the letter that Microsoft set up an email address so it could accept signatures for the letter more easily. They included Amazon and Google, which are both investors in Anthropic.<\/p>\n OpenAI eventually signed the letter, but Dr. Amodei said he disagreed with parts of it and Anthropic didn\u2019t sign it. He said in a blog post<\/a> that all A.I. models, no matter how they are developed and shared, \u201cshould go through mandatory safety testing.\u201d He said he believed bad actors would exploit open-weights A.I. models.<\/p>\n Silicon Valley\u2019s executives continued discussing how to present a united front on safety concerns, mostly keeping Anthropic involved in those discussions. In calls and texts, executives from Nvidia, Microsoft, Google, Meta, Anthropic and OpenAI kicked around ideas, according to five people familiar with the discussions. Some had informal discussions about making a joint statement or commitment on slowing A.I. development. Others brainstormed regulatory proposals.<\/p>\n But in September, three days after an Anthropic employee quit<\/a> over A.I. safety concerns and the issue burst into the mainstream, Dr. Amodei ran ahead of the other companies and made his own call for the industry to slow down<\/a>. <\/p>\n About three weeks later, Dr. Amodei struck a different tone at the White House during a lunch with Mr. Trump and more than 30 other technology executives, venture capitalists and politicians.<\/p>\n