For some time now, democratisation has been the AI industry’s favourite word. It is one, arguably, it has earned the right to use too. Less than a decade ago, building anything with machine learning required a substantial research team and a budget to match. Only a handful of organisations could afford to unlock its potential. Today, however, the barrier to using advanced AI has fallen dramatically. What was once a technology accessible to the lucky few now sits in the hands of millions.
But access is only the surface layer. Underneath it sits another layer which matters more than access to itself: training. And training has not been democratised at all. This is the dichotomy at the centre of our industry. Accessing a model is easy. Making one is almost impossible.
What we have seen thus far
What democratisation of the access layer has actually opened up is consumption. Today, someone with little or no technical background can interact directly with an AI model or even build their own app in an afternoon. With the rise of open-weight releases, they can run one on their own hardware too. These are genuine and significant advances, and it would be wrong to dismiss them.
What has not opened up is the ability to train a model. Producing a frontier model still demands significant capital, specialist talent, vast datasets, and retraining cycles that must be repeated every few months just to stay current.
That is not a barrier a boundary-pushing startup can clear. It is not one a resource-strapped healthcare organisation can clear. And it’s not even one a well-capitalise, mid-sized manufacturer can clear. It is a barrier only a few well-funded organisations in the world can clear. And it is squarely in their interests to keep it that way.
The result is a curve running in two directions at once. The cost of using a model has fallen steadily toward zero. The cost of making one has climbed toward the very limits of what private capital can sustain.
Why open source is not as open as you think
The usual answered offered at this point is open source. Open weights, the argument goes, solve the ownership problem. Give a company an open-weight model, and it can download the model, run it, fine-tune it, and deploy it. Isn’t that democratised training? It is certainly closer than a closed API. But owning the weights is not nearly the same as owning the model itself.
An open-weight model remains the frozen output of a training process run by the organisation that produced it. What users receive is the result of that process, not a capability.
A business may be able to somewhat adjust the model, but doing so effectively still demands expertise, compute, and clean data most organisations do not have in-house. And as the world moves on, the model ages. Retraining becomes necessary again, on a timeline the users do not really control or influence.
The next phase of competition
If democratisation ever reaches the training layer, if users genuinely gain the ability to create, train, and own a model, the competitive landscape will look very different.
Firstly, ownership would survive the vendor. A model trained by its user, on data that user owns, with weights held free from licensing restrictions, would not evaporate the moment the company that supplied the tools to build it disappears. That is what happens routinely today. The relationship stops being a subscription the moment ownership becomes real, and that single change removes the dependency the current market is built on.
Secondly, models would keep on learning. If training becomes something that can be done continuously rather than a cyclical and centralised event, then the model no longer freezes the instant it is deployed. Rather, it can keep learning from new data while already in production, without the need for expensive retraining cycles and a team of specialists to manage the process. The workarounds the industry has built to compensate fir static models become unnecessary because the underlying limitation is gone.
Then, accountability can be traced more effectively. When control of what a model learns, and how, sits at the level of the user rather than a third-party vendor, the chain of responsibility becomes legible in a way that it currently isn’t. That is the kind of traceability regulators are increasingly demanding, and what black-box frontier models, however capable, simply cannot currently provide.
Finally, there are the distributed benefits. This is what would make “democratisation” the right word at last. If the ability to train a model is widely held, then the value, control, and responsibility are all widely held too.
Access to models is already cheap and getting cheaper. Access to training is not and that is where the next phase of competition will be decided.