Anthropic Won. Here's the Part That's Still Up for Grabs

Welcome to edition 1 of our newly revamped The Preql Report, a bi-weekly note on the enterprise-AI stack from where we sit as Preql's co-founders. My aim is not to churn out takes. It is to send you two meaningful things every other week, one thing that changed in the market, and one thing it should change in how you buy, build or trust AI in your business.

There was really only one place to start.

Founders are told to have an opinion. Usually this is said by the same people who will later ask you to soften it (or even penalize you for it). I have decided to stop softening mine, which is convenient, because The Preql Report only works if I actually say what I think. So here we go.

Anthropic is expected to file its public S-1 next week, with a listing potentially in late September or early October. It will be the first time we get to read, in a company's own words, what winning the enterprise actually looks like on a P&L. And the word on the street as they rack up the trillions and are rumoured to out value SpaceX is Anthropic has, for now at least, won it.

Menlo Ventures estimates Anthropic took 40% of enterprise LLM spend in 2025, ahead of OpenAI at 27% and Google at 21%, up from 12% just two years ago. In coding, Claude sits around 54% share versus 21% for OpenAI. Anthropic has said more than 300,000 business and enterprise customers now generate about 80% of its revenue.

This is nothing short of genuinely impressive! I mean hats off to Dario and Danielle. But I also see the framing as genuinely narrow. Anthropic has won the model layer of the enterprise. Companies are perfectly comfortable letting Claude near their code, their tickets and their long, sad Notion pages. What they are not comfortable doing and let us all please stop pretending otherwise, is letting a model answer in its entirety what is our revenue this quarter, which customers are actually profitable, or is this deal on track or not. Not because the model can't phrase an answer. But because nobody inside the company can agree on the numbers behind it.

If you have ever sat in a meeting where four people opened four dashboards and produced four different active-customer counts, congratulations, you get what I mean. If you have not, you're lucky you've never worked in enterprise and may successfully have avoided tech altogether!

Here is my not-very-softened opinion. Enterprise AI does not have a data problem.

Enterprises have plenty of data. Warehouses full of it. Spreadsheets guarded like family heirlooms. What they do not have is agreed meaning on top of it. And certainly no way to manage, maintain and share that meaning within the organization let alone to an agentic workforce.

The ERP defines revenue one way. The CRM defines a customer another. Payroll, expenses and that one heroic ops person's who is the backbone of every team's spreadsheet has another private view of the truth. Point a very fluent model at all of it and you do not get an intelligent assistant. You get a very confident one. Those are not the same thing.

A semantic layer is the boring, load-bearing fix. Decide once what revenue is, what an active customer is, what churn is, and how each is calculated from which system. Then every dashboard, every board number and every AI assistant answers from the same definitions. It is the least glamorous slide in the deck and the one standing between most enterprises and AI they can actually trust with a number.

That is what we build at Preql.

We sit between a company's systems and its AI, dashboards and workflows, and make sure the business context underneath them is defined, governed and auditable, with the source data and calculation logic visible behind every answer. Not to make CFOs take AI on faith. So they don't have to.

The S-1 will be picked apart line by line next week. I'm less interested in the numbers themselves than in what they'll quietly reveal: how concentrated that enterprise revenue really is, how much of it is still coding, and whether customers are actually spending more in year two. That is the difference between AI as an experiment and AI as a workflow.

But none of it will tell us Anthropic has lost. It will tell us where the frontier is. And the frontier, in my very unsoftened opinion, is no longer can the model answer. It is can the business trust the answer.

That is the question the next phase of enterprise AI is really about. It is the question we started Preql to answer. And it is the question me or Leah will keep coming back to in this report every other week, using whatever the market hands us as an excuse to think about it out loud.

See you in two weeks. Bring your own definition of truth!

Gabi