Same belief, different bet

Sam Altman and Dario Amodei agree about the most important technical question in the field. They both believe deep learning systems continue to acquire capability as a predictable function of compute and data, and that the trajectory holds far enough into the future to underwrite the largest capital deployments in the industry’s history. Almost everything else about how they have built the institutions around that belief differs.

Altman, 41, has run OpenAI since 2019. He co-founded it in 2015 as a nonprofit research lab, presided over its 2019 restructuring into a capped-profit subsidiary controlled by the original nonprofit board, and oversaw the 2025 reorganization that converted the operating company into a public benefit corporation. Amodei was OpenAI’s vice president of research and led much of its early scaling work, including the research behind GPT-3. He left OpenAI in December 2020 along with his sister Daniela and several senior researchers. The OpenAI organizational update accompanying his departure described his next step as focused “less on product development and more on research.” In 2021 the Amodeis founded Anthropic, which was incorporated from inception as a public benefit corporation rather than a standard for-profit.

The first choice is structure. Altman built OpenAI inside the original 2015 nonprofit, then negotiated the 2019 capped-profit hybrid that allowed Microsoft’s first $1 billion investment, then announced in May 2025 that the operating company would convert to a public benefit corporation with the nonprofit retaining control, then completed the recapitalization in October 2025 that produced OpenAI Group PBC underneath the renamed OpenAI Foundation. The structure was rebuilt twice, in 2019 and 2025, to accommodate the capital. Anthropic’s structure, by contrast, has been consistent since founding. The corporate form has not had to absorb the company’s growth because it was designed to.

The second choice is regulation. In the summer of 2024, California’s SB 1047, the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, became the most consequential legislative test of how the frontier labs would relate to government. OpenAI publicly opposed the bill, on the grounds that frontier AI regulation belonged at the federal level rather than the state level. Anthropic’s position, articulated in a public letter to Governor Newsom after the bill’s amendments, was that the legislation’s benefits “likely outweigh its costs.” Altman extended the federal-preemption argument the following March in the tech-industry newsletter Stratechery, telling the publication it would be “very, very good” if states were not allowed to impose a “crazy complex set of differing regulations.” Newsom ultimately vetoed the bill.

The third choice is capital tempo. In January 2025, OpenAI announced “Stargate,” the $500 billion four-year infrastructure commitment with SoftBank, Oracle and MGX. Within twenty-four hours Amodei told Bloomberg the situation looked “a bit chaotic.” His specific complaint was about funding clarity. The larger disagreement was whether the public choreography of capex announcements at that scale serves the field. Anthropic has done large funding rounds; it has not done a $500 billion announcement standing next to a sitting president.

The fourth choice is rhetoric. Altman writes in declarative cadences about the imminence of transformation. “We are now confident we know how to build AGI as we have traditionally understood it,” he posted in Reflections on January 6, 2025. He has used the phrase “a few thousand days” to describe the path to superintelligence. He has called the present an “intelligence age.” Amodei’s writing, including the October 2024 essay Machines of Loving Grace, runs longer, hedges more, and leans toward enumerated technical conditions for the optimistic outcomes it describes. The two men are arguing for similar timelines using opposite styles.

Altman’s May 5, 2025 announcement of OpenAI’s PBC conversion was the moment the divergence became hardest to read as anything other than a partial concession. “Anthropic and X.ai” were named in the framing, public benefit corporations were described as the standard model for other AGI labs, and OpenAI’s structural complexity was implicitly conceded as a problem to be solved. The lab that had started as a nonprofit in 2015 was moving closer to the structure adopted by the lab that had been founded six years later by people who had left OpenAI in part because of disagreements about its trajectory.

The counter-reading is that the Anthropic-cleanness narrative is partly retrospective. Anthropic’s funding stack as of 2025 and 2026 is heavily reliant on capital from the companies that compete most directly with OpenAI for frontier customers, principally Google and Amazon. Amazon had committed $8 billion by November 2024, and Google is also a major investor. Whatever institutional independence the original PBC structure was designed to confer is, in practice, substantially mediated by the strategic priorities of two of the world’s largest cloud providers.

The equivalent observation about OpenAI is that Microsoft’s roughly $13 billion in commitments since 2019, structured through preferred returns and compute credits inside the capped-profit, gave OpenAI exactly the kind of single-investor concentration that the 2025 PBC restructuring was meant in part to soften. The financial dependencies are different in shape, not in fact. Each lab has built itself around a relationship with a hyperscaler. Altman’s choice was Microsoft and one giant relationship. Amodei’s was Google and Amazon and the friction between them.

Where the divergence remains substantive is in disclosure. Helen Toner, the Georgetown CSET researcher and former OpenAI board member, gave a public account of the November 2023 firing, on The TED AI Show podcast in May 2024, that identified a pattern of incomplete disclosures from Altman to his board, including that he had not told the board he owned the OpenAI Startup Fund, and that the board learned of ChatGPT’s launch on Twitter. The WilmerHale review concluded that the board had acted within its broad discretion, but that Altman’s conduct did not mandate removal. Anthropic’s leadership has, to date, faced no comparable public episode. What is observable is that one lab’s CEO was fired by his own board for a candor problem, and the other lab’s CEO has not been.

Amodei’s most pointed public argument against Altman’s approach has been about pace rather than direction. In interviews through 2025, he returned repeatedly to the proposition that the field is dangerous in the specific sense that capabilities can outrun the institutional infrastructure needed to deploy them safely. Altman’s response, in his own writing, was to argue that the deployment infrastructure was being built at the pace the technology required and that delaying capability deployment was itself a category of risk. In 2026 the two views moved toward each other. On August 18, OpenAI said it had temporarily slowed scaling while it strengthened safeguards. In September, Amodei called for pacing the frontier so that safety work could keep up, and Altman said he agreed, adding that further training should wait on a safety case for controllability and alignment. What remains of the divergence is how much risk each will accept along the way.

The institutional case for Anthropic is that its structure, its disclosures, and its public posture have been consistent since founding, and that consistency is itself a form of evidence about how it will behave under pressure. The institutional case for OpenAI is that whatever the irregularities of its corporate evolution, it has shipped the most consequential AI products of the era and has built the largest user base of any frontier lab. Both arguments have support, and in September 2026 both companies backed some form of pacing and independent evaluation. Their remaining split is over acceptable risk: Amodei puts more weight on precaution and regulation, while Altman argues for broad access and accepts some harms in exchange for the technology’s benefits, short of catastrophic ones. The bet on Anthropic is that consistency compounds. The bet on OpenAI is that distribution does. The two men running the two companies appear to disagree principally about which of the two postures will be remembered, ten years from now, as the one that was needed.