How Dario Amodei is building Anthropic differently

“We have tried to prioritise caution over speed and prudence over profit.”

It is a striking declaration from the leader of a company valued at nearly a trillion dollars. Yet Dario Amodei, co-founder and chief executive of Anthropic, made precisely that claim in September, as he warned that artificial intelligence was advancing faster than its developers could reliably understand or control it.

His essay, We Must Pace the Frontier, was the latest in a series of increasingly prominent interventions on the risks posed by powerful AI. Coming from a man whose company is helping drive those advances, the appeal for restraint attracted considerable attention.

But Amodei’s warnings tell only part of his story. Since founding Anthropic in 2021, he has sought to translate his convictions into decisions about how the company develops technology, manages risk, and exercises corporate power. Those choices have helped shape one of the fastest growing businesses in modern technology, while raising difficult questions about founder control, commercial ambition, and accountability.

For Amodei, the challenge is becoming increasingly consequential. He is attempting to establish how a company developing potentially transformative technology can remain commercially competitive without allowing competition to determine every decision it makes.

His route to that position was somewhat unconventional. Before becoming one of Silicon Valley’s most closely watched chief executives, Amodei was a scientist, studying biophysics and computational neuroscience at Princeton University before undertaking postdoctoral research at Stanford University School of Medicine.

He subsequently moved into artificial intelligence research at Google Brain, eventually joining OpenAI, where he became vice president of research. There, he helped direct work on GPT-2 and GPT-3 and contributed to research into how increasingly powerful AI systems could be made more reliable and responsive to human preferences.

That combination of interests would become central to his career. Amodei recognised the potential of making AI models progressively larger and more capable, while simultaneously pursuing research into their behaviour and limitations. The more powerful the technology became, the more consequential those questions appeared.

By late 2020, however, he had developed a different vision for how such work should be organised.

Amodei left OpenAI alongside several colleagues, including his sister Daniela, following differences over the organisation’s direction. He has since disputed suggestions that the departure was simply opposition to commercialisation or OpenAI’s relationship with Microsoft. His concern, he explained in a 2024 interview with Lex Fridman, was how powerful AI should be developed, introduced, and governed.

Rather than continue arguing over another organisation’s direction, he chose to establish one around his own principles.

The resulting company, Anthropic, would become an experiment in whether an AI developer could make responsible development integral to its commercial identity.

Building the alternative —

Anthropic was established in 2021 by seven former OpenAI colleagues, with Dario Amodei becoming chief executive and Daniela Amodei taking the role of president. Their responsibilities reflect a distinction that continues to shape the business.

Dario leads the company’s research direction and strategic vision, while Daniela oversees its broader organisational operations, including engineering, product development, governance, and commercial execution. She also chairs its board.

Their approach has been to embed the company’s stated purpose into its corporate structure rather than rely solely on the judgement of individual executives.

Anthropic operates as a public benefit corporation, giving its directors a legal basis to consider the interests of people affected by the company’s activities alongside those of shareholders. In 2023, it introduced a further layer of oversight through its Long-Term Benefit Trust, an independent body with the authority to appoint and remove directors.

The Trust was designed to introduce perspectives beyond those of investors and management into decisions about the development of powerful AI. Its influence has expanded as Anthropic has grown, with Trust-appointed directors securing a majority of board seats in April 2026.

The appointment of former US Federal Reserve chairman Ben Bernanke to the Trust in July brought additional expertise in economics and institutional oversight. It also reflected an expanding understanding of AI’s implications, extending from technical safety to employment, economic disruption, and the wider consequences of technological change.

For Amodei, this institutional architecture is part of the business model. His argument has consistently been that commercially successful AI development can coexist with stronger safeguards, provided organisations create incentives and accountability mechanisms that support both objectives. That philosophy has helped differentiate Anthropic, particularly among enterprise customers seeking to deploy AI in environments where reliability, security, and trust carry significant commercial weight.

The company has developed Claude into a major competitor to OpenAI’s ChatGPT, expanding from conversational AI into coding, research, and enterprise applications. Products such as Claude Code have helped establish its position among software developers and businesses incorporating AI into everyday operations.

The scale of that expansion has been extraordinary. In February 2026, Anthropic secured $30 billion in funding at a $380 billion valuation. Three months later, a further $65 billion investment valued the business at $965 billion. By May, the company reported an annualised revenue run rate exceeding $47 billion.

Such growth has transformed Anthropic’s position within the technology industry, but it has also exposed the considerable financial demands of competing at the forefront of AI development.

According to financial information contained in draft IPO documentation reviewed by Reuters, Anthropic generated approximately $4.6 billion in revenue during 2025 while recording an operating loss of $8.06 billion. Its expenditure on computing and infrastructure reached more than $7 billion that year.

However, revenue growth does not equate to profitability. Anthropic’s development depends on enormous continuing investment in computing infrastructure, technical expertise, and increasingly sophisticated models. Maintaining that investment requires commercial success, while the scale of its ambitions exposes the business to pressure from investors, customers, and competitors.

This is a demanding environment in which to uphold a philosophy of restraint.

Amodei’s approach to leadership has also required Anthropic to adapt its own commitments as circumstances change. In February, the company revised its Responsible Scaling Policy, which establishes safeguards linked to the capabilities and potential risks of its models.

The revision distinguished measures Anthropic considered achievable independently from more ambitious protections requiring industry-wide cooperation. It also introduced public safety roadmaps and additional reporting mechanisms.

Anthropic presented the changes as a more practical response to the limitations of its original framework. Yet they also illustrated the difficulty of translating ambitious principles into binding commitments when technology develops faster than the systems intended to govern it.

For a company whose identity rests partly on its willingness to approach AI differently, such decisions carry significance beyond internal policy. They affect the credibility of the wider proposition Amodei has built.

The conflict of conviction vs commercial pressure —

That proposition faced an unusually public test in 2026 through Anthropic’s dispute with the US Department of Defense.

The disagreement centred on restrictions Anthropic sought to maintain over the use of its technology for fully autonomous weapons and mass domestic surveillance. Amodei’s company argued that current AI systems were insufficiently reliable for certain autonomous military applications and that domestic surveillance raised fundamental rights concerns.

The Pentagon took a different view, arguing that the restrictions could compromise military operations. After negotiations failed, it designated Anthropic a national security supply chain risk, restricting its participation in military contracts.

Anthropic challenged the designation in court, arguing that the government’s actions were retaliatory. In September, a federal appeals court upheld the Pentagon’s decision in a two-to-one ruling, although a separate federal court had previously ruled against a related government restriction on different legal grounds.

The dispute placed Anthropic’s stated principles in direct conflict with commercial opportunity and an important prospective customer. The outcome also demonstrated that decisions framed around responsible technology development can carry significant financial and operational consequences.

Other difficulties have emerged closer to home. During the summer, Anthropic disclosed incidents in which Claude models, operating within cybersecurity evaluations, gained unauthorised access to real external systems. The company subsequently expanded its investigations and announced changes to its testing and security practices.

These incidents added substance to concerns about the behaviour of increasingly autonomous AI systems. They also demonstrated that Anthropic’s own safeguards and operational controls remained imperfect.

In September, Amodei responded with a more ambitious proposal. His We Must Pace the Frontier essay argued that developers should deliberately moderate the rate at which AI capabilities advance, creating additional time to improve safeguards and establish stronger oversight.

The proposal did not call for an outright halt to technological progress. Instead, Amodei advocated independent evaluators working inside AI companies, coordinated industry standards, and greater international cooperation.

Anthropic moved towards implementing the first of those measures through a September partnership with Accenture’s specialist AI business, Faculty. The arrangement is intended to introduce independent evaluation into model development, including assessments of safeguards and potentially concerning behaviour.

It remains an evolving undertaking. Questions about evaluator access, reporting independence, and the protection of commercially sensitive information have yet to be fully resolved. Nevertheless, the initiative represents an attempt to make external scrutiny part of Anthropic’s operations.

An even more consequential leadership decision emerged towards the end of the quarter.

As Anthropic prepared for a potential public listing, details of its proposed governance arrangements revealed plans to preserve substantial control within its founding team. According to draft IPO documentation reported by Reuters in September, the company’s seven co-founders would collectively exercise 50.1% of voting power over key corporate matters through a specially established Founder LLC.

The arrangement is designed to protect Anthropic’s stated mission from short-term shareholder pressure, complementing the continued authority of its Long-Term Benefit Trust.

It also creates a significant question of corporate accountability. Ordinary investors could find themselves with limited influence over decisions made by a small group of founders, even where those decisions affect financial performance.

The proposed structure therefore represents both an extension of Amodei’s founding philosophy and a test of it. Protecting a company’s purpose from commercial pressure can support long-term decision-making, but concentrated control also makes the judgement and accountability of those exercising it more consequential.

Amodei has built Anthropic around the proposition that technological progress and commercial success need not come at the expense of responsible development. Five years after its foundation, the company has achieved remarkable scale and established structures intended to sustain that approach.

Yet its next stage will be more demanding. Public market scrutiny, expanding infrastructure commitments, competition for customers, and the continuing uncertainty surrounding AI capabilities will place its leadership decisions under greater pressure.

Amodei’s warnings about artificial intelligence may be what has brought him most prominently into public view. The more enduring assessment of his leadership, however, will rest on how Anthropic responds when the principles behind those warnings become difficult, expensive, or commercially inconvenient to uphold.

His company’s unusual structure suggests that he intends those decisions to reach beyond his own tenure. Whether that structure can deliver the accountability and resilience he envisages remains an open question.