The Feud That Built the AI Race: Inside the Altman-Amodei Rift and OpenAI's Secret 2017 'Countries Plan'

The single most consequential rivalry in artificial intelligence did not begin with a product launch or a benchmark. It began, according to a major new book excerpt published Tuesday in The Atlantic, with two men who stopped trusting each other — and a previously unreported 2017 proposal inside OpenAI to auction rights to its future AGI to the governments of the United States, China, and Russia.

The excerpt is the first from The AGI Chronicles: The Inside Story of the Race to Create an Artificial Superintelligence, the forthcoming book by New York Times technology columnist Kevin Roose, which goes on sale October 6 from Farrar, Straus and Giroux. Roose spent more than a year reporting the book, drawing on over 150 insider interviews and internal documents. And its opening salvo lands at a moment when the two companies at its center are locked in the most expensive arms race in the history of computing: OpenAI, reportedly in talks to raise at least $30 billion at a $1.4 trillion valuation, and Anthropic, whose IPO prospectus targets a valuation above $2 trillion.

Why Amodei really walked

The public record has always been thin on why Dario Amodei left OpenAI in 2020 with six colleagues to found Anthropic. The standard account cites “differences over safety and commercialization.” Roose’s reporting makes the rupture far more specific — and far more personal.

By 2020, after helping lead the work on GPT-2 and GPT-3, Amodei had concluded that OpenAI’s leadership could not be trusted with the technology he was building. In Roose’s telling, Amodei came to regard Sam Altman as “a two-faced manipulator” and OpenAI co-founder and president Greg Brockman as “power-hungry and reckless.” He objected to OpenAI’s conversion into a commercial enterprise — the capped-profit pivot of 2019 that later became a full for-profit restructuring — believing the lab was drifting from its founding nonprofit mission of building AI that benefits all of humanity.

The departure itself was operationalized like a covert operation. While Amodei was still negotiating with Altman over terms to stay, the group of seven who would become Anthropic’s founders formed a private Slack channel. They named it BATNA — the negotiation-theory acronym for “best alternative to a negotiated agreement.” The channel’s very name says everything about the state of trust: even as talks continued, the founders were planning their walk-away.

The ‘countries plan’

The most striking revelation in the excerpt is an internal proposal that dates to 2017, when OpenAI was a struggling nonprofit in need of compute money. According to five people familiar with the matter, Greg Brockman and Ilya Sutskever explored what became known internally as the “countries plan” — an auction among national governments, including the United States, China, and Russia, for rights to OpenAI’s future AGI technology.

Roose reports that the idea began as a more benign concept — an international AGI coalition — but evolved into a fundraising strategy. Brockman and Sutskever rationalized it on national-security grounds: any future AGI would carry such immense strategic implications that governments might simply seize it anyway, so OpenAI might as well choose its partners and get paid.

OpenAI disputes the framing. A spokesperson told Roose the company had “briefly considered whether to turn OpenAI into an international cooperative endeavor” but “did not consider selling AGI to China or Russia.”

Amodei was appalled when he learned of the plan. Selling AGI to dictators, he believed, was the exact opposite of OpenAI’s mission. He warned internally that the scheme would intensify international arms races, risk enabling human-rights abuses, and undermine the lab’s moral legitimacy. That warning reads today as remarkably prescient, given that US-China AI competition has since become the organizing principle of technology policy on both sides of the Pacific.

The Kanye West beta access episode

The excerpt also documents smaller judgment calls that, accumulated, corroded Amodei’s trust. In one vivid episode, Altman allegedly offered Kanye West access to the GPT-3 private beta in 2020, citing two people familiar with the offer. Jack Clark, then OpenAI’s policy director, shut the idea down. It is a minor story in itself — but as a window into the unguarded instincts of the man running the world’s most important AI lab, it is the kind of detail that made Amodei conclude the guardrails were cultural, not structural.

Why this matters now

It is tempting to file this as Silicon Valley gossip — ancient history from a nonprofit that nearly ran out of money in 2017. That would be a mistake, for three reasons.

First, the feud is load-bearing. Roose argues that the personal enmity between Altman and Amodei directly fuels today’s race dynamics. Two companies now racing toward superintelligence — each spending hundreds of billions on compute, each warning in its investor documents about catastrophic risk — are run by two men who, by this account, have despised each other for years. When Anthropic’s prospectus warns of “catastrophic or existential risk to humanity” while committing to $518 billion in future compute, and OpenAI shelves its own GPT-6.1 Astra model over internal safety failures, the question of whether the people in charge can cooperate is not a personality story. It is a governance question for the entire field.

Second, the “countries plan” is a data point in a live debate. The idea of nation-states bidding for frontier AI rights sounds outlandish until you look at 2026: sovereign wealth funds are now among the largest investors in OpenAI and Anthropic, the White House has convened tech CEOs to draft “accords” on superintelligence, and the US and China have held dedicated AI safety talks. The 2017 proposal was, in a sense, simply early — a crude sketch of the sovereign-AI entanglement that has since arrived through investment vehicles instead of auctions.

Third, the book itself will be an event. Based on 150+ interviews, The AGI Chronicles is positioned as the definitive insider account of the race so far, and future excerpts — and the responses from OpenAI and Anthropic — will keep this history in the news through the fall.

OpenAI’s denial is categorical but narrow: it did not consider selling AGI to China or Russia. The question of what it did consider — an international coalition, a cooperative, an auction among allies — remains exactly the kind of ambiguity that has long let each side in this feud tell a different story about the same events. Roose’s bet, evident from this first excerpt, is that the truth of the AI era will be found not in the benchmark charts but in the Slack channels: BATNA, and what was said there in 2020 by seven people who no longer believed their boss should hold the future in his hands.

The AGI Chronicles goes on sale October 6, 2026.