Sovereignty in a can

Yesterday, a government reached in and switched off a frontier AI model.

This kind of restriction to technology isn’t unprecedented, but it’s unusual in its speed and scope. One day, an enhanced capability was available to paying customers. The next day, it’s gone. The vendor hopes it’ll be back, and soon. But it’s out of their hands.

I was surprised by how much this rattled me. Vendor dependency and single-source procurement are risks that I consider. But this made it visceral. I’d used Fable 5 for a couple of end-to-end security reviews but I wasn’t depending on it. Unlike other models, I hadn’t wired it into any part of my tech stack or deployed it in a client’s environment. That would have been reckless anyway.

What I realised was that it wasn’t about the model. It was proof-of-concept. The concept: a capability you depend on can be removed by a party who is neither you nor your vendor. The proof: Fable went back behind closed doors.

The reality of any information technology is that it comes with risks. The three brought to mind by this particular event were data privacy, vendor dependency, and geopolitical volatility.

My main academic background isn’t in tech - it’s in the study of comparative politics, history, economics, terrorism and asymmetrical warfare. It wasn’t my CTO instincts that were triggered by the US government’s direct interference in AI commerce; it was my international studies background. It put me in mind of North Korea’s parallel internet, the NSA’s Clipper chip… even Napster (RIP).

The reflex action, naturally, is to switch vendors. And it’s worth considering, but that just moves the dependency. Another vendor is another hand on the switch. A portfolio of vendors is a sound move. We’ve seen car manufacturers learn that lesson the hard way when local manufacturing stopped because their suppliers collapsed; when they’re prudent, they don’t buy anything that they can’t get from multiple vendors

But in the case of AI, or more properly LLMs, the real hedge is the open-weights that you physically hold. That’s model files on your own disk, in your own metal box, that no directive can reach. Which led me to the conclusion that I needed to build just that.

It’s probably fair to say that between the longbow, the blacksmithing, the rucking, I’m not a million miles from a doomsday prepper. I don’t mind canned food. But in this case, it’s not so dramatic.

What needs building is a desktop PC with serious capabilities, not a bunker. I’ve just ordered the parts and it’ll come together over the next week or so. It’ll be open-weight models, on my hardware, and my network.

In one comic detail, the RAM and the GPU each cost twice what they would have a year ago. One vendor-related bottleneck remains the memory shortage, not ideology.

Once it’s built, it’ll be a self-contained server for an LLM, air-gappable if a client needs it, and with security that doesn’t rest on anyone’s terms of service. And not at the whim of a world power that is becoming increasingly unpredictable.

In practice, this is me undertaking a real-world experiment to ensure I understand these risks and the countermeasures.

If a client is concerned about how much of their work runs through a capability someone else controls, we need to be able to design a solution for that.

What happens when today’s subsidised LLMs are repriced, restricted, or switched off? Where does our data actually live? What survives the vendors?

For most businesses these risks are real and unpriced, but it’s time to consider them. We need to make sure that the solutions have been planned so we’re not caught flatfooted. I’m walking that path now, and I’ll share what I’ve learned along the way.

So back to the office: in this case sovereignty isn’t a manifesto. At this moment, it resolves to a screwdriver, a $2 anti-static strap, and a tower under my desk.

Build incoming. I look forward to sharing what I learn along the way.