The European Commission has published its response to reduce Europe’s dependence on non-EU digital infrastructure and give control of the related cloud and artificial intelligence service. The European Technological Sovereignty Package sets out a clear statement of intent with the likely introduction of the Cloud and AI Development Act which will outline laws on AI models and data sovereignty.
The timing couldn’t have been more pertinent, given almost a fortnight later Anthropic’s AI models were banned in the US over concerns of foreign access. It’s perhaps not surprising an event, and it’s one I will watch closely. However, for now it should come as a stark reminder how concerned governments, businesses and NGOs are about solving the data sovereignty conundrum, especially when it comes to AI models.
Governments are balancing national security risks with their growing dependence on AI-powered public services. Scaling businesses and multi-nationals alike are grappling with AI innovation and productivity gains against a backdrop of managing complex laws that come from expanding into and operating on a global stage.
There’s no doubt that the intersection of AI models, cloud infrastructure and security is murky for companies who are not expert in this field of technological advancement.
Who owns the data, who controls it, whose responsible and accountable for it? And more crucially, under which laws – where the AI was modelled, or where the AI is used, or where the data is hosted? Do we need network centric policy or data centric ones?
Big questions with no single, simple answer.
Where is the risk in AI models and sovereignty?
As I see it, the risk isn’t only about where the data sits. It’s about who has legal reach over it. For instance, you can use a US cloud provider to run a workload in a region of the UK. Plenty of companies do.
But as well as adhering to UK data laws, you are still exposed to US law through the Cloud Act. This means that “stored in Britain” and “controlled from Britain” are two very different things.
It’s a discussion I have almost daily with clients and prospects, and the bottom line is that that the gap between storage and control is where the real exposure lives.
What’s the answer? It’s not a drawbridge.
A clearly stated intent, as the EC has done, and collaboration from governments is an important part of the story. For that matter, I do believe governments should keep firm control of the things that genuinely matter, so things like citizen identity, sensitive data, the infrastructure those sit on. But AI models, and the data in them, is a bigger challenge.
I expect we will see more countries build frontier models to manage AI risk and deal with sovereignty through local ownership. But there will be a greater number who don’t or can’t simply because the compute power, the capital and the fact the talent isn’t there for most.
For most countries the achievable goal, and more useful one, is to own what sits underneath the AI models, so the data centres, the power, the connectivity, and the hosting that runs the inference. Get that right and your national workloads runs on infrastructure you control, even when the AI model itself was trained somewhere else.
In the meantime, and in terms of corporate responsibility and AI models and sovereignty, leaders need to identify what’s commercial risk, what’s strategic risk and what’s legal risk. In the main, leaders need to be cognisant of the difference between ownership and control, as that’s not always understood at board level.
Cloud partners like us do understand this challenge however, and will help build a strategy that meets the compliance needs of the jurisdictions of their host country (if it’s different to where the data is used) and the territories they operate in today, and in the future.
Residency matters but it’s not the finishing line
If you’re a leadership team tasked with thinking about this challenge, then I’d say this, though data residency matters it’s only a starting point, it’s not the finish line.
Keeping data in the country reassures people, partners, customers, governments, but if the operator is foreign-owned and answers to foreign law, the address on the building doesn’t change who can ultimately compel access.
You won’t solve this problem until you have both regulated ownership and legal control alongside residency, so data held in the UK is also held by a company that answers only to UK law.
The trap many companies fall into is assuming that by writing residency rules into a contract (that any large provider can tick off by opening a local region) has you covered.
It doesn’t. All that does is hand the market straight back to the same few hyperscalers and bake in the dependence you were trying to escape.
<<Refer to our guide here on what to consider if you need more help on this aspect of planning>>
Is this a lesson we need to learn now? You bet.
We’ve already lived through this shift to a few not the many with cloud, and it isn’t a happy lesson. Looking back, no governments collectively decided the country’s digital infrastructure should run on three American providers. It just happened, one decision at a time, and unpicking it now is slow and expensive.
Sovereign AI is on course to repeat that exact pattern. It’s a form of dependence by default rather than by choice. The lesson from telecoms and semiconductors is that this gets settled at the infrastructure layer, the data centres, the power, the networks, which is the part the UK can genuinely own. The model layer almost certainly isn’t.
Speak to CWCS
See what you could save and get a clear plan to move away from AWS. Speak to an expert to identify whether colocation or a hybrid approach could reduce costs without sacrificing flexibility.
