AI sovereignty is a dumb idea.
The truth about AI or tech sovereignty more generally is that the truest version of this is basically impossible today. There is just too much interdependence.
Just take a look at the AI supply chain and you can see this.
The whole chain involves so many actors and no one part of the chain works in isolation. Chip designers depend on foundries, foundries depend on toolmakers, toolmakers depend on specialty materials and subcomponents, and AI model deployment depends on servers, memory, networking, and electricity.
AI sovereignty, often described as a state’s capacity to develop, control, and govern AI according to its own national interests, should be treated as nothing more than a pipe dream.
China has been trying to domesticate the AI stack for years now, and while it has made real progress, full self-sufficiency likely remains out of reach. The dependence still shows up in certain ways; Nvidia chips keep getting smuggled across the border and Chinese AI companies keep facing accusations of distilling US frontier models.
Most countries should forget about trying to rebuild the whole stack domestically. This just seems doomed to fail and waste of resources. The plan should not be to build the next OpenAI or Anthropic or some other frontier lab.
But this does seem to be what a lot of national AI policies seem to focus on. There appears to be a bias towards first-mover advantage and introducing and dominating some new innovation in the AI supply chain.
It is an attractive route because the upsides are immense. Generally when a country is the first to a technological innovation, this drives economic growth, in turn expanding productive capacity and wealth which then ultimately translates into greater influence and power on the world stage.
But this is quite hard to do, and such a policy choice also takes focus away from an aspect of this trajectory that often gets overlooked.
For a technological innovation to achieve higher growth and productivity, it requires effective proliferation. It needs to be adapted and embraced across industries so that its value can really be unlocked.
This is especially important for AI, a general-purpose technology that can be applied in many different ways. But the transition from theory to practice requires the development of infrastructure that makes implementation possible. It is a focus on diffusing innovative technology widely rather than being the first to invent it.
Diffusion potentially provides a pathway for countries to achieve economic growth from a certain technology despite not being the first to develop such technology. It may take many years to realise these benefits, but if they are eventually realised they could be significant.
Because general-purpose technologies spread gradually rather than rewarding initial breakthroughs, the decisive competition among leading powers occurs not in pioneering the tech but in diffusing it widely. This is shown in what happened with electricity, where several industrial powers innovated almost simultaneously yet the US pulled decisively ahead through faster adoption.
There is a difference between trying to replicate existing innovative tech and building the infrastructure to make it easier to implement in different domains. And for most countries, the more lucrative opportunity is the latter rather than the former.
Chasing the frontier is a bet on being first. Building for diffusion is a bet on being useful. Most countries can only win one of those, and it isn’t the first one.