Neither Washington nor Beijing: the third way is a necessity

Control has moved from hardware to models

Export restrictions on training hardware were meant to slow a competitor down. They did not produce the expected containment: models from laboratories subject to those restrictions kept improving, by optimising what they could no longer buy.

Control then moved one notch further, towards the models themselves and towards the distillation techniques that transfer a capability from one model to another. The question is no longer who owns the machines, but who has the right to learn from whom.

This shift matters to a European buyer, because it changes the nature of the risk. A hardware risk is managed with stock and lead times. A risk bearing on the right to use a model materialises overnight, through a change of licence or of access conditions, with no operational warning.

An internal contradiction that is hard to hold

This approach runs into a contradiction its promoters struggle to resolve. An overwhelming share of the American start-up ecosystem itself depends on open models of Chinese origin, for reasons of cost and freedom of deployment.

Close to two hundred of those companies have publicly asked that these models not be taken away from them. The request is coherent from their point of view: they built products on open components, and losing access would cost them their economic advantage.

Above all it illustrates a reality a European decision maker should absorb: in a globalised ecosystem, the country of origin of an open model is less determining than the conditions under which you can download it, audit it and run it yourself.

The moral argument and the geopolitical moment

The argument that certain training corpora were assembled illegitimately deserves examination, and it is too rarely examined on the merits.

It is however hard to separate from the geopolitics of the moment. The same practices of mass content ingestion are the subject of ongoing lawsuits in the jurisdictions of those who make the argument. A criterion that applies only to competitors is not a criterion, it is an instrument.

For a European organisation, the practical consequence is clear. The debate on the legitimacy of corpora is a genuine one, and it will be settled in courts and through regulation. In the meantime, protection consists in favouring models whose training chain is documented, whatever their origin, because documentation is what will let you answer if the question is put to you.

What the third way is not

One reading comes up often and weakens the position, so it should be set aside. The third way is not technological autarky. Refusing to depend on a single supplier does not mean refusing to use what others produce well, wherever it comes from.

Nor is it national preference in disguise. A poorly documented European model, impossible to run on reasonable infrastructure and with no community to maintain it, offers no additional guarantee over an open model produced elsewhere. Origin is not a criterion, conditions of use are.

Finally, it is not a wait-and-see position. Nothing in this analysis justifies postponing a project in the hope that the market will clarify. It will not, and organisations that wait accumulate a usage gap costlier than the imperfect trade-offs they are avoiding.

What it is: a reversibility requirement applied to every component, verified by a real test rather than by a clause, and revisited at regular intervals.

Two dependencies, neither of them tenable

An organisation building on a closed American model inherits exposure to an extraterritorial jurisdiction, a pricing policy it does not control and a roadmap over which it has no hold.

An organisation building on an open Chinese model inherits a risk of access withdrawal, uncertainty about training documentation and a political exposure whose extent depends on events beyond it.

The two risks are not of the same nature, and one may be preferable to the other depending on the case. But neither can be controlled from inside the organisation, and that is the only criterion that matters for a system you will be responsible for over ten years.