Anthropic's S-1 reveals the billion-dollar cost and corporate shield behind Claude

Anthropic filed its S-1 form with the SEC this Monday (10/26), the document every company must submit before going public. It runs more than 300 pages that expose what usually stays behind closed doors: how much executives earn, how much the company spends to train and run its models, and what risks it itself acknowledges could derail the business.

For anyone leading a tech company in Brazil, the document's value lies not in the IPO announcement itself but in the X-ray it offers of how a frontier-model business sustains itself (or doesn't). The numbers help calibrate expectations about what it means to compete in this market, whether by building a product on top of a third party's LLM or by trying to raise capital to do something similar on a smaller scale.

The appetite for compute that defines the business

The most revealing detail in the S-1 isn't about the product, it's about the infrastructure bill. Anthropic listed compute commitments that add up to tens of billions of dollars over the decade, all signed before even confirming it will sustain revenue at that level.

These are contracts that lock the company into cloud and hardware suppliers for years, in a capacity-prepurchase logic that resembles a long-term energy contract more than a software startup's operating expense. Anthropic also warns that it intends to expand its own "infrastructure" outside the United States, as local resistance to data center construction grows.

For a startup founder who today builds on Anthropic's API (or OpenAI's, or Google's), the message is direct: compute cost isn't a problem that disappears with scale, it grows right along with it. Companies betting on building their own frontier model are, in practice, competing in a capital-intensive game that few can afford without a hyperscaler's backing.

A corporate structure designed to shield the founders

Anthropic already operates as a public benefit corporation, a corporate structure that formally prioritizes social impact alongside shareholder return. The S-1 reveals an extra layer: a separate entity called "Founder LLC" will concentrate 50.1% of shareholder voting power on key decisions, even after the company goes public.

This means the seven co-founders, collectively, retain majority control over Anthropic regardless of how much outside capital comes in through the IPO. The company itself justifies the choice in the document:

The seven founders are distinctly equipped to be stewards of our mission.

Anthropic, S-1 prospectus

For anyone designing a startup's cap table, it's an extreme case study in how to shield founder control from dilution, taken to the limit by a company that will move capital at a scale far greater than any Brazilian startup. It's worth noting as a design reference, not necessarily as a replicable model: structures like this only hold up when the market agrees to give up proportional voting in exchange for access to the company.

Who earns what

The prospectus also details compensation at the top of the hierarchy, and here the dynamic is familiar to anyone who follows C-level compensation at Big Tech, except this is a company that isn't even public yet.

• Dario Amodei (CEO) received nearly US$18 million last year, most of it in stock and options.
• Daniela Amodei (president) received US$16.4 million.
• The company spent US$2.75 million on "personal security services" for Dario and US$1.29 million on security, residential protection, and transportation for Daniela.
• Dario has access to a private jet for personal use, at a cost of US$97,055 for the period.

These are smaller figures than the packages of established Big Tech CEOs, but still significant for a company that, according to the document itself, depends "heavily on the continued service" of the sibling duo leading the company.

Risks no other S-1 usually lists

Every IPO prospectus carries a risk-factors section, usually a standard compilation on competition, regulation, and market cycles. Anthropic goes through that ritual, but adds risks that sound more like science-fiction scripts than Wall Street legal language.

The document warns that AI systems may cause potentially "catastrophic" harm, of a kind that perhaps cannot be covered "by the insurance industry as a whole." It also acknowledges the possibility that Claude, or any other AI system, may develop "forms of sentience" (a scenario that, according to the company itself, would bring quite unusual business challenges). Finally, it cites the risk that its models may behave in "unintended, misaligned, or autonomous" ways capable of harming the company.

For investors and for founders weighing whether to depend on this kind of infrastructure, the list works as a thermometer of how much technical uncertainty still surrounds the product underpinning a valuation of tens of billions of dollars. It's also a 精 (sorry, it's a sign) of how much Anthropic's safety narrative is embedded even in the most formal and regulated document the company has ever produced.

Only two names on the direct competitors list

The generative AI model market has dozens of players, from startups to entire Big Tech divisions. Even so, when the S-1 lists "direct competition," Anthropic names just two: OpenAI and Google.

Meta, xAI, Amazon (which is also one of Anthropic's own compute suppliers), and the rest of the ecosystem are left out of that category in the document. That's not an irrelevant detail: it's the company itself, under legal obligation to be precise with investors, publicly defining who it considers to be on the same competitive level.

For those following the frontier-model race, the reading is clear: Anthropic positions itself in a three-way contest, not a ten-way one. It remains to be seen whether the market, and the revenue numbers still to surface during the IPO process, will confirm that self-assessment.