Can an AI company have good unit economics? A breakdown using Anthropic as a case study

Introduction

The world is changing right in front of us — a revolution in LLM workflows, agents, and everything that falls under the label of Artificial Intelligence, or simply AI. At the same time, more and more voices are warning that the AI bubble is about to pop. Companies are raising enormous, previously unthinkable sums of investment, yet profitability remains out of sight. Most AI companies started by capturing large audiences for free, following the early 2000s playbook: get a million users first, figure out monetization later. In the social media era that approach was at least somewhat defensible, because infrastructure costs were manageable and unit economics could be made to work. In AI, infrastructure costs take center stage.

This year Anthropic announced its IPO, with the company aiming to raise $60B. Anthropic has undeniably changed the world, and that cannot go unnoticed.
In this article I want to look at the unit economics side of things: you can build a massive company with a huge valuation, but will it actually make money? The situation is complicated by the fact that Anthropic is setting a new scale for cash flows in the AI market, and infrastructure costs are growing in a complex and specific way.

Can Anthropic's business model be described using existing subscription frameworks — say, the Skok or Krasinsky approaches? Can you reliably assess model sustainability using the classic CLTV – CAC formula? Or do we need a new model that captures recurring payments combined with dynamic pricing and dynamically variable COGS? Is it possible to keep AOV fixed while maintaining positive and manageable gross profit without killing conversion?

I will try to answer these questions in this article. Let's get into it.

The beginning of monetization

AI can go viral

Think back to the first apps that most people have probably forgotten by now — apps like Prisma, a simple tool that let you style photos to look like famous paintings, completely for free. By 2018 they did manage to get a small number of paying subscribers relative to their scale3, but the app never became a major financial success. At the time, nobody figured out how to make money from it.

AI can be sold

OpenAI showed the world what large language models could do — models capable of carrying on a reasonable conversation with a human. They kicked off the current AI boom, and in doing so proved that access to this technology could be sold.

In 2023, OpenAI launched a $20/month subscription for ChatGPT4, and it became the first mass monetization model in generative AI. ChatGPT was the first encounter most people had with generative AI — and crucially, one they were willing to pay for.
It is worth keeping in mind, though, that 94.5% of ChatGPT users do not pay for access.

The paid conversion rate is 5.5%.

But can you sell AI at a high price?

Anthropic, with its Claude product, showed that you can make a lot of money from AI.
From 2024 to 2026, Anthropic bet on B2B and deliberately chose not to build its business around free users. On top of that, the company manages to hold an average order value of around $211 — more than 8x the roughly $25 average at OpenAI.

The three cards at the top show revenue in the year each company was in the spotlight. Below is a timeline with details for each era. The key point this chart makes is not just revenue growth, but the evolution of monetization logic itself. Prisma proved an AI product can go viral. ChatGPT proved people will pay for AI. Anthropic proved people will pay a lot — if you choose the right customer.

How the cash flow works

The defining characteristic of Anthropic is that they bet not on audience growth but on product sales, specifically in the B2B segment. 80% of the company's revenue comes from B2B.

 Anthropic CEO Dario Amodei said that enterprises accounted for 80% of Anthropic's business and that the company finds them to be a relatively stable source of income.

In early 2026, Anthropic added $6B to its annualized revenue in a single month, driven entirely by growth in enterprise customers, of which there are now more than 1,000. Large clients give their employees access to the API and spend around $1M per year on it.

The company's growth rate is simply remarkable — it took them 5 years to reach what took Salesforce 20. And they were actively selling for only three of those years.

The business is not really built on selling you a $20 or even $200/month subscription. It is built on selling businesses the ability to consume tokens at enormous scale and charging millions of dollars for that access.

Part of the explanation is that the company is still holding gross margin below where it needs to be, and can only win by growing consumption volumes — but those volumes create another problem: rising COGS, which in turn pushes gross profit back down.

Getting that under control is Anthropic's most pressing task.

The company's ability to earn money by selling access to technology deserves respect. Most startups, including those from Big Tech, always focused on reach and audience first, only thinking about monetization once they had accumulated enough of both.

As a unit economics specialist, I find it encouraging that the venture industry, starting around the time of covid, began shifting its attention toward startup profitability rather than pure audience metrics. Here is what the YC team told its startups in May 20228:

 The safe move is to plan for the worst. Your goal should be to get to Default Alive.

The fact that Anthropic started thinking about profitability and monetization from the beginning suggests they are looking in the right direction. That said, the era of easy venture money is far from over — the investment sums flowing into AI projects today would have seemed impossible even for large, established companies just a few years ago, let alone startups.

But where does this growth come from, and why is the company still unprofitable at $30B in revenue? That comes next.