Could Anthropic Be the Most Important Company in History?

The IPO has put Anthropic’s economics in focus. Its founders deserve a closer look.

There’s plenty of analysis of Anthropic’s business. What I find even more interesting is the kind of company they’re building: who the founders are, what brought them together, and how that shows up in their technology and the way they treat people. If Anthropic could become the most important company of our time, possibly in history, its character deserves as much attention as its financial statements. Looking at the people alongside the products and the S-1 has made me more favorable about the company.

// Scientific ambition belongs in the incentives analysis. Dario Amodei and Chris Olah coauthored practical AI-safety research in 2016. Several future founders contributed to scaling laws and GPT-3 in 2020. Their work predates Anthropic’s commercial success. That’s why I give their scientific motivations real weight.

// Anthropic’s influence could outlast any lead in model performance. Claude’s applications help people carry out work, while MCP gives other developers a shared way to connect AI to tools and data. I think that combination could give Anthropic lasting importance even as competing models improve.

// The mission still has to work as a business. Nine-month 2026 revenue reached $33.242 billion, with a $1.260 billion operating loss. The much larger $104.039 billion net loss includes $94.324 billion in financing revaluation expense. These unaudited results deserve careful reading: my enthusiasm for the research doesn’t settle the economics.

I’m favorable about the company. Its lasting importance will depend on what people accomplish with the technology, including progress in software, science, and other work that benefits people well beyond Anthropic’s customers.

Everyone is talking about Anthropic’s business. The IPO, the revenue, the losses, the enormous cost of computing power. You can find plenty of analysis of those numbers, and we’ll look at them here too.

What I find even more interesting is the kind of company they’re building. Who are the people behind it? What brought them together? What have they spent their careers trying to understand? And how does that show up in the technology they make and the way they treat people?

I think Anthropic could be the most important company of our time. Possibly in history. If that’s even a possibility, the character of the organization deserves as much attention as its financial statements.

I realize how that sounds. I use their tools, I’m a fan, and in my opinion they’ve built the most capable AI platform available. I also have a ridiculously expensive token bill. There’s something addictive about being able to get good work done that quickly. You start thinking about everything else you could do with it.

But I didn’t start out with this much confidence in the company.

I started with Charlie Munger’s habit of looking at incentives. When an AI company talks about changing the world while raising enormous amounts of money, it’s reasonable to wonder how much of the ambition is simply to make an enormous amount of money. That was part of my skepticism about Anthropic.

Then I spent more time looking at the people who founded it, the work they’d already done, and the problem they wanted to solve. I think I was giving too little weight to another incentive: the opportunity to solve the biggest problem in your field. Perhaps the biggest problem anyone has attempted to solve.

If you believe it’s possible to build machines that can help us discover medicines, write software, understand science, and do increasingly difficult intellectual work, I can see why you’d organize your life around it. I can also see why you’d want to understand what those machines are doing before handing them more responsibility.

That’s what interests me about Anthropic. Its IPO filing gives us a much better look at the business. Its products let us experience some of the technology ourselves. And its founders’ careers give us a way to judge whether the mission appeared before or after the money.

The company produced a conversation with seven of its cofounders in December 2024. It’s worth spending time with.

What interests me is how they describe the decision to build a company. Around 32 minutes in, Jared Kaplan explains why pursuing the work required an institution and access to considerable capital. Daniela Amodei talks about the shared mission and trust among the founders. They describe forming Anthropic as a way to pursue work they believed needed to happen.

It’s a company-produced interview, so of course they’re telling their own story. I still think it’s a revealing one. These are people who had been working on the underlying questions for years.

Here are the seven people in that conversation. Their current roles come from Anthropic’s leadership page; their earlier work helps explain what each brings to the company.

The mix matters to me. A group of physicists and engineers needs people who understand organizations, public institutions, and how technology affects people. Daniela and Jack are part of why this is a more interesting founding team than the usual shorthand suggests.

There’s also a personal connection. My last client was Opaque.co, led by my good friend Aaron Fulkerson. He joined Anthropic with key members of his team to help protect the data we entrust to its systems. They bring deep expertise in confidential computing: protecting sensitive data even while it’s being used. Apple’s Private Cloud Compute is a useful comparison. It’s designed to keep personal data private during cloud AI processing, even from Apple itself. That’s the kind of protection I want when I put my own information into an AI system.

I’ve never heard someone rave about the kindness of an organization the way Aaron has about Anthropic. That means something to me, especially coming from a good friend.

Anthropic didn’t invent modern AI. The field is the work of generations of researchers, universities, companies, and open-source communities. Several of its founders, however, were already contributing to important advances before they started the company in 2021.

In 2016, Dario and Chris were coauthors of “Concrete Problems in AI Safety”. The paper addressed practical questions: how to avoid unintended side effects, prevent systems from exploiting poorly specified rewards, and keep them working when circumstances change. This was years before Claude was a product.

In 2017, Dario and Tom contributed to research showing how systems could learn from human comparisons of behavior. Work like that helped establish the approach known as reinforcement learning from human feedback, or RLHF. Humans don’t have to write a precise mathematical rule for every desired behavior; their preferences provide part of the training signal.

In 2020, Jared, Sam, Tom, and Dario were among the authors of “Scaling Laws for Neural Language Models”. They studied relationships among model size, data, computing resources, and predictive performance. The research helped make decisions about building larger models more systematic. It didn’t prove that intelligence would improve indefinitely, but it gave researchers useful evidence about where to invest effort and compute.

That same year, Tom, Jared, Sam, Jack, and Dario were among the larger team behind GPT-3 at OpenAI. GPT-3 demonstrated how one large language model could perform many tasks from instructions and examples, without a separate round of training for every task. That generality helps explain why today’s assistants can move among writing, programming, analysis, and other kinds of work.

The chronology changed how I thought about the incentives. The research record shows longstanding interests in both capability and safety. I find that more persuasive than a mission statement written after a product becomes successful.

At Anthropic, those interests produced Constitutional AI. The 2022 paper, whose first-listed author is Yuntao Bai, describes using written principles to guide a model’s critiques and revisions, then using AI feedback in further training. People still choose the principles and design the process. It’s a method for shaping behavior at scale, with human judgments built into it.

Interpretability addresses a related problem: understanding the machinery inside a model. Chris’s earlier work on neural-network “circuits” illustrates the ambition to identify understandable components and how they interact. If AI becomes involved in scientific discovery or important decisions, being able to investigate why it behaves a certain way could become extremely valuable.

Claude is easy to describe as an AI assistant. That description leaves out much of what the platform can now do.

There are the underlying models, the applications people use, the tools that let those applications take action, and the developer services that let other businesses build their own products. Keeping those distinctions clear makes the product portfolio easier to understand.

You can see the progression. Claude launched in 2023 as an assistant. Computer use arrived in public beta in 2024, letting models interact with software through screenshots, mouse actions, and keyboard input. Claude Code followed in 2025, giving the model a working environment for software development. The newer products extend that approach into other kinds of work.

This helps explain why the tools can be so useful and why using them heavily can become expensive. A task may involve repeated model calls, reading files, using tools, revising output, and checking results. You’re paying for that activity, and the bill can grow as you find more work to hand over. Lower costs for individual operations don’t necessarily mean a lower monthly bill.

My token bill doesn’t prove Anthropic has the best economics. It does reflect my willingness to keep using the platform. The question I care about is how much useful work I get from the time and money I put into it.

MCP also deserves more attention than it usually gets in a company profile. A useful common protocol can influence products built by other companies. Anthropic could have a lasting effect on AI through the infrastructure it helps establish, even as the lead in model performance changes hands.

An S-1 is the registration statement a company files as it prepares to offer shares publicly. Anthropic’s November 2 amendment gives us more current information than its original October filing.

For the nine months ended September 30, 2026, its unaudited results show $33.242 billion in revenue, a $1.260 billion operating loss, and a $104.039 billion net loss. Those losses measure different things: the net loss includes $94.324 billion of expense from revaluing financing instruments. It should not be presented as cash burned running the business.

The filing also reports $895.4 billion in aggregate purchase and lease commitments, including short-term commitments and leases not yet commenced, as of and subsequent to September 30. Of that, $447.5 billion is expected before the end of 2030; substantially all the remainder falls between 2031 and 2045. The preliminary prospectus still leaves the offering price range and share count blank.

That is an extraordinary amount of money to commit. It makes the scientific ambition tangible: building and serving these systems requires resources on a scale that very few organizations can assemble.

The credit ratings deserve a look, too. The Financial Times reported that KBRA assigned Anthropic a BB+ issuer rating in June: the highest speculative-grade, or “junk,” rating, one notch below investment grade. That is a rating of the company, not every bond associated with it. The FT separately reported that junior debt in the Project North Star chip financing received BB+ from KBRA, with a positive outlook, and BB from Morningstar DBRS. I can be enthusiastic about what Anthropic is building while recognizing the financial risk involved.

My favorable view doesn’t make those commitments less consequential. Anthropic has to keep building products people will pay to use, manage the costs of delivering them, and make investment decisions before it knows exactly what future demand will look like. Scientific ambition and commercial discipline have to coexist.

I’d also separate my enthusiasm for the company from any judgment about its eventual shares. A company can do historically important work while an investor pays too much to own it. We don’t need to settle that question to examine why Anthropic matters.

The strongest case for Anthropic’s importance is that its technology could help people do more of the work that produces other advances.

Software is one example. Better tools for writing, maintaining, and understanding code can benefit every organization that depends on software. Scientific research is another. A system that helps researchers analyze results, test ideas, or reduce administrative work could contribute to progress across many fields.

Dario makes the ambitious version of that argument in “Machines of Loving Grace,” his essay about AI’s potential benefits in biology, neuroscience, economic development, and other areas. The essay describes what he hopes capable AI could make possible. Those outcomes still have to be achieved.

There are already smaller, concrete examples. Enveda reported using a Claude-based system to help review pharmaceutical manufacturing records. Its study covered four batch records from one clinical campaign: 1,061 scanned, largely handwritten pages. Reviewing the records with the system’s comments available took 9.4-fold less time than unaided manual review. People retained responsibility for the final decisions.

That’s a limited study reported by the company that built the system. It doesn’t establish that AI has solved drug development. It does show a specific way AI may reduce work that stands between a scientific idea and its practical use. A lot of important progress depends on getting thousands of unglamorous tasks done correctly.

For me, that’s where the historical possibility becomes credible. If a company helps improve the tools used to make discoveries, build organizations, and deliver services, its effects can spread well beyond its own products. The eventual importance would depend on what people accomplish with those tools.

The incentives deserve another look here, too. Anthropic says its founders pledged 80% of their equity holdings to philanthropy when they founded the company. A pledge isn’t a completed donation, but it is relevant to the explanation that their sole objective was personal wealth.

There’s an organizational commitment as well: Anthropic’s Long-Term Benefit Trust is financially independent of management and investors and has authority to appoint a majority of the board. Whether that structure works will depend on how people use it when decisions become difficult.

Its safety policies have changed since the founders recorded that conversation. In February 2026, Anthropic separated commitments it would make on its own from recommendations for the broader industry. It also explicitly described its safety-roadmap goals as goals rather than hard commitments. That distinction matters when judging how much protection a promise in its current policy actually provides.

There are substantial questions ahead. How reliably can these systems perform work without close supervision? Who gets access to the benefits? How much power should sit with the organizations that build the models? And can Anthropic preserve the qualities people admire as the financial and competitive pressure grows?

I’m favorable because of what I’ve used, what I’ve learned about the people building it, and what I’ve heard from people I know. I think their ambition is serious. If they succeed in helping us solve problems that have resisted decades of effort, Anthropic’s importance could extend far beyond anything we see in its IPO filing. That’s why I think it could become the most important company in history.

Your AI Sherpa,