Ownership of AI sounds simple. You can use it whenever you want, on your own terms, and you can open it up and change it. But when I sat down to write Own 2026, I wanted to get more precise. Where does ownership of AI actually sit? And how do you know whether you have it?
The book comes down to one idea. Access is the easy part. Ownership is where the power sits.
Four layers
Every AI you use is made of four layers:
- The weights: the model itself, the file you can download and run.
- The compute: the hardware it runs on.
- The data: what it was trained on, and everything you feed it.
- The interface: the app you type into.
Most of us only ever touch the interface. It feels like ownership because it’s on our screen and it knows our name. But the interface is the one layer that gives you nothing on its own. If someone else holds the weights, runs the compute and keeps your data, you are using AI on terms you didn’t write.
Real ownership starts at the bottom. A model you possess, running on hardware you control, with your data staying in your hands.
This is also where I see a lot of confusion. Plenty of models call themselves open. Some really are. Others come with user caps, limits on what you may use them for, or terms the provider can change whenever it likes. That isn’t ownership. It’s a lease with a nicer label. The label on the box is not the contents, and the licence decides what you actually get. My rule of thumb from the book: ownership you cannot describe in a sentence from the licence is ownership you do not have.
Why it matters
Owning the weights is only worth something if you can run them. For years that meant data-centre hardware, which is why AI pulled so hard towards a few large providers.
That gap is closing. Techniques like quantisation shrink models so they run on ordinary laptops, and they are open and free for anyone to use. And most useful work doesn’t need the biggest model anyway. It needs a good model that knows your work. That’s why I don’t believe centralisation is inevitable. It has momentum, but momentum isn’t destiny.
When you own all four layers, nobody can switch your AI off, change how it behaves without telling you, or decide what it may be used for. What it learns from you stays yours. And when a better model comes along, you move and take everything with you.
When you only own the interface, every one of those decisions belongs to someone else. That might be fine today. It’s a risk for tomorrow, especially as more of our work, our companies and our lives run on AI.
How to check what you own
Knowing this is one thing. Checking it for the model in front of you is another, and it’s harder than it should be. Documentation is scattered, licences are long, and marketing pages say “open” either way.
That’s why we started the AI Ownership Index. It’s a library, not a ranking. For each open model, and the providers that serve it, it documents what you need to know to own it: openness, licence, provenance, hardware, how to run it, how to tune it, and how it’s supported. Every section shows where the evidence comes from. Either we cite an existing source, we tested it ourselves, or we mark it as a gap and say why. It’s in alpha and grows every week.
A note on independence: the index is a OneHill project, and Anthill runs open models of the kind it documents. That’s exactly why every piece of evidence is labelled, so you can check it rather than trust us.
If you want the full argument, Own 2026 is free to read: onehill.org/own2026. It’s short. I wrote it for founders, builders and anyone who wants to use AI rather than be used by it.
And next time you pick a model, look it up on the AI Ownership Index first and check what you’d actually own. If something is missing or wrong, tell us. That’s how the library gets better.
Christoph Welsbach, founding director of OneHill. 17 October 2026.
Written by Christoph. His own AI was used to help edit and structure this piece.