Governments are building AI-driven productivity on infrastructure they cannot price, govern, or quickly replace, and the economic gain flows to whoever controls the bottleneck.
AI can make a country richer and more dependent at the same time.
Factories get faster and public services improve while the same economy runs on chips, cloud capacity, and data centers it cannot price, govern, or quickly replace.
That dependence isn’t on anyone’s radar until access tightens, and then whoever controls the bottleneck holds the leverage over everything built on top of it.
A country’s protection is to keep more than one route to compute: owning some, renting some, pooling the rest, holding on to the ability to switch. That is sovereignty over AI, and it costs far less than a national hyperscaler.
Where does control of AI compute actually come from?
Owning GPUs does not give you control over AI compute. Nvidia may design the accelerator, but manufacturing, packaging, and memory supply depend on TSMC, Samsung, SK Hynix, and Micron.
Then somebody has to put those chips in a data center and supply the power to run them.
Nvidia’s CUDA platform has spent years becoming the default environment for many AI developers, which makes switching hardware more complicated than simply buying a different chip.
The biggest cloud companies are already trying to reduce that dependence. Google has been building TPUs for more than a decade, Amazon has Trainium, and Microsoft has Maia.
All three still use Nvidia hardware, but their own accelerators give them another option for at least part of their workloads. The clearest evidence that dependence matters is that even the largest buyers are trying to escape it.
Each of them reduced one dependency and ran into another further down the stack. Once a buyer can no longer switch suppliers easily, the provider sets the terms, and every product built on that infrastructure absorbs the consequences.
When does corporate dependence become economic dependence?
A company can buy GPUs from a private supplier for years, and then finds that access depends on a government decision. US export controls set where advanced AI accelerators can be sold and to whom, and those lists are revised as policy shifts.
For one company, losing access delays a product or pushes it toward more expensive capacity elsewhere.
When most of a country’s businesses and public services sit on the same infrastructure, those delays aggregate into something an economy feels: investment slips, and projects that made financial sense at one price stop making it at another.
At that point, a rule written in a capital where the affected economy holds no vote reaches past the companies buying the chips. It sets how fast that economy can adopt AI, and how much of the resulting productivity gain stays inside it rather than leaving as payment for access.
Does AI sovereignty mean every country needs its own hyperscaler?
If dependence is the problem, the obvious answer is national control: domestic data centers and domestically owned capacity.
Copying AWS or Google Cloud is expensive on a scale most smaller economies will not clear. The realistic answer is several sources of compute rather than one.
Europe is already moving that way with its network of AI Factories, where countries and institutions share access to large-scale computing infrastructure instead of each building the same capacity from scratch. LUMI in Finland works on the same principle: a group of European countries funded the supercomputer together and now draws on its resources, which adds shared capacity to what each of them runs nationally.
That opens a third position between building everything and renting everything from the same handful of providers. A country can own some compute, share some, and reach into a regional pool when it needs more.
Can countries share compute without giving up sovereignty?
Once countries start sharing compute, a new problem appears: if the hardware belongs to someone else, how do you use it without exposing your model or the data it processes?
For inference, confidential computing can solve part of that trust problem. It runs sensitive workloads inside hardware-protected trusted execution environments (TEEs), which limit what the infrastructure operator can see. Remote attestation adds another layer: it can then check that the expected protected environment is actually running before anything sensitive is sent to it.
This does not remove the dependency entirely. It shifts it. Instead of trusting whoever operates the machine, you trust whoever built the hardware and security stack underneath it, and today that stack relies heavily on Intel, AMD, and NVIDIA. In principle, nothing ties the approach to these three companies, and other chipmakers could build comparable protections. For now, though, that is where the mature options are.
What does change is ownership. The machine no longer has to belong to whoever owns the model. Third-party compute can still be used while keeping the model, prompts, and other sensitive data better protected from the operator.
Training is the same problem in a different shape. A group of hospitals may each hold useful patient data, but consolidating every record in one country or one database is often legally impossible.
Federated learning lets them train together while the raw data stays where it is. Each hospital contributes to a shared model without handing its records to the others.
Researchers are pushing the idea toward larger models. Photon demonstrated federated pre-training of models up to seven billion parameters on geographically distributed, weakly connected GPUs.
That is still far from replacing the clusters used to train frontier models. It does open a middle ground where institutions share infrastructure and still keep control of the models and data they cannot afford to hand over.
What does sovereignty over compute actually require?
Very few countries can build every layer of the AI stack at home, and none of them need to. The position worth holding is one where critical workloads keep running when a supplier pulls back or a policy changes. The mix will differ by economy: some domestic capacity, some rented, some shared regionally, with confidential computing or federated training where the model and the data have to stay protected.
What matters is that the work has somewhere else to go. Sovereignty is the credible ability to choose another route.
About the Author
Anastasia Matveeva is a Senior Product Manager and researcher at Product Science and a co-creator of the Gonka protocol. Her work focuses on machine learning infrastructure, large language model inference, and distributed computing systems. Anastasia holds a PhD in Mathematics from UPC Barcelona, where she worked as a researcher and lecturer.
