October 20, 2026
Nations, businesses and individuals all want the same thing from AI, and all hit the same shortage: the behaviour that shows how people actually live sits on their phones and barely reaches a model. New DataSapien White paper
Sovereign AI is the phrase of the year. Until now, it has mostly been said about nations: a country that runs its own models on its own infrastructure. The logic is control. However, with rising AI token costs and concerns over sharing proprietary data, increasing numbers of organisations are looking to post-train and tune their own models, as part of a wider model mix.
But the model is the easy part. Models are converging, and compute can be bought or built. The hard part, the part that decides whether an AI is actually sovereign, is the data it is grounded in. And here every sovereign-AI programme meets the same wall. Its corpus is the public internet, and the public internet does not hold the one thing that predicts how people will behave. That behaviour sits on hundreds of millions of personal devices, under consent, in a form no central crawl can reach.
The same wall stands in front of a business.
AI models have Incomplete knowledge
A frontier model is a compression of the public web. It is extraordinary at language and reasoning, and it knows the median of what people like your customers tend to do. It does not know what your customer did yesterday, what they abandoned at eleven at night, or what they are in-market for this week. That signal, the human context that predicts the next decision, never reaches a training pipeline. A model rented by the token gives you the market average every competitor can also rent. What it cannot give you is your own customers.
Here is the tell. The open web has been scraped dry, and the scramble for the next source has turned undignified. Meta, sitting on one of the largest behavioural datasets ever assembled, has started recording its own employees’ keystrokes to train its models on how people work. Brokers are buying up the private Slack archives and email of shuttered companies and selling them to AI labs, often without the people in those messages ever knowing. When firms are reduced to mining their own staff and picking over the data of dead startups, it tells you what behaviour is worth, and how little of it is there to buy cleanly.
Scarcity
The reason it is scarce is that the old ways of reaching it are closing. To use behavioural signal, you have always had to collect it: pull it off the device, into a warehouse, under your control and your liability. Regulation is tightening around bulk collection. The mobile platforms are shutting it down at the operating-system level. The metering tricks the panels leaned on for a decade are degrading in the field. The signal is getting harder to reach at the exact moment it becomes most valuable.
Which is where the third level comes in, the one the sovereign-AI conversation keeps leaving out: the individual.
The behaviour everyone wants to train on belongs to a person, and it is generated in a place that person controls, on the device in their hand. Most companies treat that as an obstacle to engineer around, and inherit the liability and the mistrust that come with pulling the data out. The alternative is to build on it. Bring the model to the data. Let it be shaped by a customer’s behaviour on their own device, and let only the result cross back. The person stays sovereign over their human context, because the behaviour never leaves. The business becomes sovereign over its intelligence, because it can post-train its own model on the consented, anonymised signal, a corpus no competitor holds, without ever holding the behaviour behind it. And a national programme can stand on the same ground: a population that consented, device by device, rather than a dataset taken from it.
Three problems, one cause.
These are not three separate problems. They are one pattern at three scales. The individual keeps their data. The business keeps the intelligence it trains. The nation keeps the capability. In every case the behaviour was used where it sat, and never copied into a warehouse to do it.
There is a hard edge that proves the point. After WWDC this year, Apple changed its developer terms to ban using output from an Apple model to train or improve another model. Read it closely and it catches brands, not just rival labs. The platform’s model will personalise your moment; it will not let you build one of your own from it. To own the intelligence, you have to run your own model, on the device, against behaviour that stays there.
This is what sovereign AI actually requires, under the national headlines. Not a bigger model, which anyone can rent. It takes the behaviour the open web never had, used where it is generated, with the person’s permission, and left on their device. A third-party vendor can sell every competitor a plausible model of the customer. Only the business, the nation, or the person who holds the relationship can ground an intelligence in what actually happened.
The behaviour is already there, generated every day by the people you already serve, in the apps they already use. The only question is whether you build the architecture to learn from it before the window closes further.
Human-centred. Not data centres. DataSapien.
The full 11 page white paper is here: Training Sovereign AI on Behavioural Data