Intelligence sovereignty: the bet we made building SmartAlex

I watched All-In episode 279 this weekend. I watch it most weekends, because the show has a habit of landing on the thing everyone else will be talking about a couple of weeks later. This time the four of them spent the first forty minutes on one idea, and I sat there realising they were describing the exact bet we made when we built SmartAlex.

The idea has a name now. Chamath calls it intelligence sovereignty. Here is the line that stuck with me, somewhere in the middle of the argument: you cannot rent intelligence from the same place that rents it to your competitor. Read that twice. It quietly reframes almost every AI decision a business is about to make.

Privacy is the old worry: can someone see my data. Sovereignty is the new one: can someone use my data to compete with me, or use their AI to decide what I get to think. On the show the examples were blunt. A frontier lab watches where the value is being created on top of its model, then moves in and builds that product itself. It happened to Figma when Claude Design launched. It happened to the coding tools when Claude Code launched. The pattern is old. It is the Microsoft and Google playbook: win the layer everyone depends on, then use that position to take the most valuable things people build on top of it.

So what does AI safety mean if you actually run a business. Not the science fiction version. It means you keep control of four things: your compute, your models, your data, and your alpha, which is the proprietary know-how that makes you you. The moment you hand all four to a single provider, you have handed over the one thing that was genuinely yours.

David Sacks framed the market as three layers. Chips at the bottom, models in the middle, applications on top. Right now the middle is consolidating into a duopoly: two companies do almost all of the revenue at the model layer, and both have every reason to keep it that way. If you build an application, or you run an enterprise, that is the last thing you want. You want the model layer to stay competitive, because competition is what keeps your options open and your data yours.

The economics are already moving in that direction. On the show, Chamath shared a test his team ran on an ordinary enterprise task. Running an open model through their own control plane came out about sixteen times cheaper than calling a frontier model directly. It ran slower, but the price difference was enormous. And the cost of tokens is projected to fall by roughly ninety percent a year for the next three years. Intelligence is on its way to being nearly free. The question stops being can I afford it, and becomes who owns the part that matters.

When we built SmartAlex, we made a choice that looked stubborn at the time. We do not hard-wire the business to a single AI provider. SmartAlex runs on our own engine, and every part of it, the model that thinks, the layer that listens and speaks, the way calls are carried, is swappable underneath the application. Your conversations, your prompts, the way your business answers its phone, all of that is your asset, and it stays with you. When the model layer gets cheaper, and it will, you inherit the saving without rebuilding anything. That is intelligence sovereignty pointed at the one channel almost every business still runs on: the phone.

How we keep your intelligence yours

The model

Swappable. No single provider owns your stack.

Your data

Your calls and prompts stay your asset, not training fuel.

The cost curve

You inherit falling token prices automatically.

The hard moments

AI carries the volume, a person takes what matters.

The back half of that segment was a genuine argument about jobs, and it is worth your time. The short version: the data does not show AI erasing work. A study of more than twenty-one thousand firms found the companies that adopted AI hardest actually grew their headcount, entry-level roles included. What is happening is displacement and reshaping, not a cliff. And the most interesting prediction was this: as more gets automated, a human in the loop becomes the premium, not the cost. Klarna is the tell. They replaced their support team with AI, made a lot of noise about it, then quietly reversed course, because customers want to know a real person is there when it counts.

That is exactly how we think about voice. The goal was never to fire your receptionist. It is to let AI carry the volume, the after-hours calls, the routine questions, the overflow at lunchtime, and hand the calls that genuinely need a person straight to a human, with the context already gathered. The machine takes the boring nine calls. Your best person takes the tenth, the one that was worth taking.

If you run a business, here is what forty minutes of very sharp people amounts to in practice:

• Do not marry one model. Build or buy on something that lets you swap the intelligence layer without re-platforming.
• Treat your data as your edge. Before you hand a dataset to any AI provider, ask what they become once they have it.
• Design for cheap. Assume intelligence keeps getting cheaper and stop rationing it. The teams pulling ahead are the ones spending tokens on things that were not worth it last year.
• Put a human on the moments that matter. Automate the volume, not the relationship.
• Move now. The phrase from the show was slow down to speed up. The businesses setting this up today are the ones that will look untouchable in eighteen months.

None of this is a prediction I am making from the outside. It is the bet we already made. SmartAlex is what intelligence sovereignty looks like when you point it at the phone: your own engine, your own data, a person where it counts, and none of your edge leaking to a company that might wake up tomorrow and decide to compete with you.