Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

The great cloud migration in reverse

Why AI is now pulling back some of what we moved

[Photo: Getty Images]

Profile photo of Pedro Henriques

Founder & CEO · BRIDGE IN

Over the last decade, companies spent considerable time and effort moving everything to the cloud. Files, databases, and company and customer data. It was the obvious call. Cheaper than a rack in a closet, easier to run, and someone else’s problem when it broke at 3 am.

The artificial intelligence tsunami is making some of us reconsider that decision. The worry is data sovereignty, though that phrase hides more than it explains. When we run a prompt through someone else’s model, our inputs leave our control, and our customers’ data often goes with it. We are also trusting that the model behind the application programming interface stays the same from one week to the next, and that nobody is learning how our business works from the questions we ask. For a compliance company, a law firm, or anyone whose product is trust, that worry adds up.

So, the old instinct comes back. Bring the AI in-house. Run the models on our own machines, in our own building, where nothing crosses the perimeter.

WHAT IT TAKES TO TAKE AI IN-HOUSE

Until recently, running your own capable model was a nice idea with no hardware behind it. That changed this year. In late July, Moonshot launched Kimi K3, an open-weight model big enough to compete with the best closed systems. You cannot run that one in your office. It needs a rack of specialized chips. Days later, DeepSeek refreshed its far smaller V4 Flash and came surprisingly close to it, on a model small enough to run on a single high-memory Mac Studio, with the weights free to download under an MIT license. Running a genuinely useful model on hardware you own stopped being theoretical.

Another option worth considering is neither on-premises nor an API call: You can rent GPU capacity by the hour and run your own open-weight model on it. You have full control over the model without any of the racks to babysit. But if you are renting someone else’s computer, did you end up getting the sovereignty you were after, or did you just hand the problem to a different landlord?

To answer that, we need to stop using data sovereignty as a simple, single concept. It is four separate things, and most setups win some and lose others.

1. Model control. Do you own the weights, and can the model change without warning? Your own open-weight model inference is fully controlled. An API model can be updated, repriced, or retired out from under you.

2. Usage exposure. Does a third party get to see how your business uses AI, which prompts you use, which workflows and integrations, and which instructions you give on how to process your data?

3. Runtime confidentiality. While the model is running, can the operator read your data or your weights straight out of the machine?

4. Jurisdiction. Who can legally force access to the data? This is about where the operator is incorporated, not only where the servers physically sit.

TIME TO TEST

Now run the AI options through those four parameters. Start with the hosted API, where you lose all of them. The no-retention and no-training promises are contract clauses, and contracts get breached and subpoenaed. These are also the same labs that trained on the open web first and argued the copyright questions in court afterward. Decide how much of your business should rest on their good behavior.

Running your own open-weight model on a rented cloud gives you back model control and usage exposure, but the operator can still read the machine while it runs. Adding confidential computing closes that gap. Only on-premise models win on all four.  But you pay for it in maintenance.

Jurisdiction follows the operator, not the server location. Whoever runs the box sets the law that can reach your data. You choose the jurisdiction by choosing where that company is incorporated. A U.S. provider answers to the U.S. CLOUD Act even with the servers in Frankfurt, which is why Amazon’s European Sovereign Cloud, live since January, still sits under U.S. law. The same goes for AI labs themselves: Nearly all frontier ones are American or Chinese, so a hosted API puts your prompts under one of those two legal systems, wherever you are.

Confidential computing—GPUs that encrypt their own memory so even the operator cannot read it—is the newest piece companies should consider. It is further along on the big U.S. clouds than on the European sovereign providers, where the GPU instances are still catching up.

THE HUMAN ELEMENT

Whichever option you pick, a person somewhere can still reach your data. You need to decide who and how many.

Run it on-premises, and you are trusting your own staff and door locks. Run it through an API, and you are trusting the lab and everyone it answers to. Run it on a confidential rented box, and you are trusting the security built into the chips, the attestation that proves the machine’s claims, and the provider’s honesty about which government can knock on its door. Every option requires you to trust someone.

Security was never a single wall. It was, and is, a stack of layers. The layer that gives way is almost always a person: a reused password, a careless contractor, a social engineering attack. So, the honest goal is not a guarantee but mitigation through a safer design.

A U.S. company picking a U.S. cloud keeps most of this invisible, because the law and the vendor sit on the same side of the ocean. For everyone else, the question is sharper. Do you want to keep renting your intelligence from a handful of American and Chinese labs, on their terms and their legal turf? Or do you want to own the model your business runs on, and choose, deliberately, the short list of people you can trust?

The cloud simplified everything. Confidentiality, data sovereignty, and protecting business alpha around AI are now making leaders reconsider their cloud strategy.

Pedro Henriques is the founder and CEO of BRIDGE IN.