Anthropic: company profile, products, funding and valuation (2026)

Some companies you profile because you can invest in them. This is not one of those. Anthropic is a private company, and for almost everyone reading this it is not purchasable in any direct way. We are writing about it because it is one of the most reported-on private businesses of the decade, and because it is a near-perfect case study in a question our members ask constantly: how do you read a company you cannot see inside?

So this is a profile written the way we would write any pre-diligence note — sourced where we can source, and hedged where we cannot. We hold no position in Anthropic and offer no access to it. What follows is public reporting, read carefully.

What Anthropic is

In one line: Anthropic is an AI research and safety company that builds and sells the Claude family of models. It competes at the frontier of large language models, and it makes money mainly by charging developers and enterprises for access to those models rather than by selling a consumer subscription at scale.

It is a private, US-based public benefit corporation — a for-profit structure that also carries a stated public-good mission in its charter. That structure is not decoration; it is central to how the company describes itself and to the safety thesis it was founded on.

Every figure in this table is drawn from press reporting and company announcements. Private company disclosures are partial by nature. Read all numbers as approximate and potentially out of date.

The founding story and the safety thesis

Anthropic was founded in 2021 by a group who left OpenAI, led by Dario Amodei, who became chief executive, and his sister Daniela Amodei, who became president. The reported reason for the departure was a disagreement about pace and safety — how quickly powerful AI systems should be built and released, and how much caution should sit alongside capability.

That disagreement is not a footnote; it is the company's entire premise. Anthropic positions itself as building frontier models and researching how to make them safe, on the argument that you cannot make AI safe from the sidelines — you have to be at the frontier to understand it. Its best-known technical contribution to that argument is a training method it calls "constitutional AI," in which a model is trained against an explicit written set of principles rather than purely on human feedback.

The bet underneath Anthropic is that safety and capability are not opposites — that you have to build the most powerful systems in order to learn how to control them. Whether that bet holds is the whole question, and no one can yet mark it.

For an investor, the safety framing cuts both ways. It is a genuine differentiator with enterprises that fear reputational and regulatory risk. It is also a constraint the company has chosen to accept, and constraints have costs. Both things are true at once.

Products, and where the money comes from

The product is Claude, offered in tiers. On the company's own framing, Opus is the most capable tier, built for deep reasoning and long, complex tasks; Sonnet balances speed and capability for everyday enterprise work; and Haiku is the fast, high-volume, lower-cost tier. These are sold through three main channels:

• The API. Developers and companies pay per token to build Claude into their own products. Reporting consistently describes this — enterprise and developer API usage — as the core of the business.
• The Claude apps. Consumer and enterprise-facing chat products, including paid tiers, though consumer subscriptions are generally reported as a smaller share of revenue than the API.
• Claude Code. An agentic coding tool that has been reported as one of the fastest-growing products in the company's history, reaching around $1bn in annualised revenue within roughly six months of launch, as reported. Coverage in 2026 credited it with a large share of the enterprise coding market.

The revenue trajectory reported through 2026 is the headline that made Anthropic a household name in finance. Coverage described run-rate revenue climbing from around $9bn at the end of 2025 to roughly $30bn by around April 2026 and, per later reporting, crossing about $47bn by late May 2026. These are run-rate figures — an annualisation of a recent short period — and run-rate is a flattering lens in a fast-growing business, so read them as directional rather than as audited annual revenue.

The big backers: Amazon and Google

Anthropic's two anchor investors are, as widely reported, Amazon and Google. What makes these more than ordinary equity cheques is that they are bound up with cloud-compute deals — the investors also supply the infrastructure Anthropic needs to train and serve its models.

On Amazon: reporting through 2026 described a total commitment building to tens of billions of dollars — earlier investment topped up by a further reported tranche (around $5bn, with a much larger amount tied to commercial milestones) — alongside a deal in which Anthropic committed to very large multi-year spending on AWS and access to large quantities of Amazon's Trainium chips. On Google: reporting described a major expanded commitment, including a reported immediate investment with a larger amount potentially to follow, and an expansion of Anthropic's access to Google's TPU compute.

The strategic logic is straightforward, and worth stating plainly: the cloud providers earn back much of what they invest as Anthropic spends it on their compute, while also securing a stake in a leading model maker and a marquee customer for their AI infrastructure. That is a rational deal for them. It also means a meaningful part of Anthropic's reported "funding" is circular — money that flows back out as compute spend — which is exactly the kind of nuance a headline valuation hides.

Funding history and the reported valuation trajectory

The valuation story is one of the steepest on record, and it must be read with care. Private valuations are set by the terms of a single round — one price, one set of investors, one moment — not by a liquid market. A rising private mark tells you what the last investors were willing to pay; it does not tell you what the shares would fetch if many holders tried to sell at once.

With that caveat carried throughout, the reported trajectory runs roughly as follows:

• A Series G was reported at around $380bn post-money, roughly February 2026, as reported.
• A Series H was then reported at around $965bn post-money, roughly May 2026 — a very large step up in a matter of months, as reported.
• Coverage in mid-2026 also reported that Anthropic had confidentially filed for an IPO, which, if it proceeds, would eventually replace private marks with a public price.

Every one of those figures is a press-reported, unaudited, private number. We have deliberately not invented a precise date or a decimal you could not trace to public reporting. If you need exact terms, the only reliable source is Anthropic's own disclosures and, in time, any IPO prospectus.

Competitive position

Anthropic's clearest rival is OpenAI, the company several of its founders left. The two are frequently compared on model quality, on revenue and on enterprise adoption. Reporting through 2026 described Anthropic as having drawn level with or ahead of OpenAI on some measures of enterprise and developer spend, and as strong specifically in coding — while other datasets, using different methodologies, still put OpenAI ahead on overall share. On the evidence, the two are close and the measurement is contested; be sceptical of any single market-share figure quoted without its source and method.

Beyond OpenAI sit Google's own DeepMind and Gemini models — notable because Google is simultaneously an investor in Anthropic and a competitor to it — along with Meta, a field of well-funded challengers, and increasingly capable open-weight models that pull down the price of "good enough" intelligence. The competitive risk is not only that a rival ships a better model; it is that the category itself commoditises, and pricing power erodes for everyone at the frontier.

The bull case and the bear case

Held in balance, without a thumb on the scale:

Both columns are built from the same public facts. Which one you weight more heavily is a judgement, not a calculation — and anyone who tells you the answer is obvious is selling you something.

What the public record cannot tell you

This is the section we would refuse to skip, because it is the one that matters most and is most often left out. A private company discloses what it chooses to. Here is what, on the public record, we genuinely do not know:

• The unit economics of inference. What it actually costs Anthropic to serve a token, and therefore the gross margin on its revenue, is not disclosed. This is the single biggest unknown in the whole AI category.
• Cash burn and runway. Fast revenue growth and heavy compute spending can coexist with enormous losses. The net cash position is not public.
• Contract terms with backers. The precise structure of the Amazon and Google deals — what is equity, what is committed spend, what triggers what — is only partially reported.
• The real capitalisation table. Exact ownership percentages, preferences and liquidation terms are not fully public, and preferences can dramatically change who gets paid what in an exit.
• Whether the reported valuation is durable. A private mark is one round's price, not a market clearing price. It can move sharply in either direction.

A named gap builds more trust than a confident guess. We would rather tell you plainly that these things are unknown than paper over them with a number that reads well and cannot be sourced.