Anthropic: The Safety Company That Had to Win

Anthropic is the company that left OpenAI because it thought OpenAI was moving too fast and too commercially — and then raised tens of billions of dollars, signed on Amazon and Google, shipped one of the most capable AI models on earth, and agreed to pay $1.5 billion to settle a lawsuit over pirated books. The gap between "we are the safety-first lab" and "we are a $380-billion frontier-AI business" is the most important thing to understand about it, because the company lives inside that gap on purpose.

This is not a simple story of hypocrisy. It is the harder story of what happens when you genuinely believe a technology is dangerous, conclude that the safest thing is to build it yourself before someone reckless does, and then discover that building it requires becoming exactly the kind of well-capitalised commercial juggernaut you set out to be an alternative to.

The split

Anthropic was founded in 2021 by seven people who left OpenAI, led by the siblings Dario Amodei — a physicist by training, born in 1983, who had been OpenAI's vice president of research — and Daniela Amodei, who became Anthropic's president. They brought with them a cluster of senior researchers: Jared Kaplan, Jack Clark, Chris Olah, Ben Mann, Sam McCandlish, and Tom Brown, several of them central to the work behind GPT-3.

The stated reason for leaving was "directional differences" — a polite phrase for a real disagreement about how safely and how commercially OpenAI was developing increasingly powerful models, particularly after its deepening entanglement with Microsoft. The founding thesis of Anthropic was, in essence: frontier AI is coming whether we like it or not, so it had better be built by people who take the risks seriously. They structured the company as a public-benefit corporation and created a "Long-Term Benefit Trust" intended to put mission ahead of shareholders in governing the board.

The contradiction at the core

Here is the tension that defines Anthropic, and that the company has never been able to fully resolve: you cannot build a safer frontier model than the reckless labs unless you are at the frontier, and being at the frontier costs tens of billions of dollars. So the safety company had to become a fundraising machine.

It succeeded spectacularly. Amazon committed on the order of $8 billion across 2023–2024; Google put in well over a billion more. By a February 2026 round, Anthropic was reportedly valued at around $380 billion. To raise that kind of money you must promise commercial returns, ship competitive products fast, and chase enterprise revenue — all the behaviours that "moving too fast and too commercially" was supposed to describe. Anthropic's defenders argue this is the only coherent way to influence the technology's trajectory; its critics argue it is the precise process by which a safety mission gets quietly converted into a growth mission, with the safety language retained for branding. Both readings are supported by the evidence, which is what makes it interesting.

Claude, and "Constitutional AI"

The product is Claude, a family of large language models that, by 2025, was among the most capable in the world and especially strong at coding, long-document reasoning, and following nuanced instructions. Anthropic's genuine technical contribution to the safety conversation is real: techniques like Constitutional AI (training a model to critique and revise its own outputs against a written set of principles, reducing reliance on armies of human labellers) and a serious, well-funded interpretability research program aimed at actually understanding what is happening inside these models.

Credit where due: this is not safety theatre. Anthropic publishes substantive research, takes interpretability seriously when most of the industry treats it as a cost centre, and has been more willing than rivals to talk concretely about catastrophic-risk scenarios. The uncomfortable counterpoint is that the same company is also racing to make the models more powerful and more widely deployed — and you cannot fully do both at maximum speed.

The FTX money

No honest account skips this. In April 2022, FTX — Sam Bankman-Fried's crypto exchange — invested about $500 million in Anthropic, money associated with the effective-altruism network that Bankman-Fried courted before his fraud collapsed the company. When FTX imploded, that stake became an asset of the bankruptcy estate, which sold it in 2024 for roughly $884 million to help repay FTX's defrauded creditors.

Anthropic did nothing illegal in taking the investment, and it is a real irony — even a kind of poetic justice — that a stake bought with what turned out to be fraud proceeds was later sold at a profit to compensate the victims of that fraud. But it sits awkwardly with the company's moral framing. A lab that presents itself as the conscience of the industry took half a billion dollars from one of the most notorious fraudsters of the era. That is part of the record.

The pirated books

The cleanest puncture of the halo came in the courts. Authors sued Anthropic over how it built its training data, and the facts that emerged were unflattering: alongside legitimately acquired text, the company had used large troves of pirated books. A judge's 2025 ruling drew a careful line — training a model on lawfully obtained books could be transformative fair use, but acquiring and storing pirated copies to do it was a different matter. In September 2025, Anthropic agreed to pay about $1.5 billion to settle, on the order of $3,000 per book.

It is hard to overstate how on-the-nose this is. The company founded to be the responsible adult in the room paid one of the largest copyright settlements in history because it had, in plain terms, used stolen material to build its product. Anthropic would frame it as an industry-wide practice it is now resolving responsibly; the less generous and equally accurate framing is that "we are the careful ones" and "we trained on pirated books" are both true, and the second one cost $1.5 billion.

The honest verdict

Anthropic is the most intellectually serious of the frontier labs and the clearest illustration of why "safety-first AI company" may be a contradiction in terms at this scale. The Amodeis are not cynics; the interpretability work, the published research, and the genuine alarm in Dario Amodei's public writing all read as sincere. They really do appear to believe that the safest path runs straight through the frontier, and that someone thoughtful had better get there.

But sincerity does not dissolve the contradiction. To execute that belief, Anthropic took fraudster money, trained on pirated books, raised to a $380-billion valuation by promising commercial returns, and now races the very rivals it formed to be safer than. The most truthful sentence is the uncomfortable one: Anthropic may be both the most responsible frontier lab and a living demonstration that "responsible" and "frontier" cannot be maximised at the same time. Watch what it does when those two goals conflict — because they will, repeatedly, and those moments are the only reliable test of which mission is actually in charge.

---

Editor's note: HustleMemo writes founder-led case studies grounded in public reporting. We hold the company's stated mission and its documented conduct in the same frame rather than choosing one. We do not interview subjects on a paid basis and disclose conflicts in-line. Corrections: editorial@hustlememo.com.

Sources

• "Anthropic" and "Dario Amodei," Wikipedia (founded 2021 by seven ex-OpenAI staff incl. Dario & Daniela Amodei; public-benefit corporation; Long-Term Benefit Trust; Claude).
• Reporting on the Amazon (~$8B across 2023–2024) and Google (>$1B) investments and the ~$380B valuation (Series G, 2026).
• The April 2022 FTX $500M investment and the FTX estate's 2024 sale of the stake for ~$884M.
• The authors' copyright litigation, the 2025 ruling distinguishing fair-use training from pirated acquisition, and the September 2025 $1.5B settlement ($3,000/book).
• Anthropic's published research on Constitutional AI and interpretability.