Europe’s sovereignty ambitions require smarter infrastructure

The boom in AI tools is frequently framed as a race to secure GPUs before someone else does. The contestant’s position is typically measured in hyperscaler CapEx, chip partnerships, and the number of Nvidia units deployed.

Yet when two thirds of organizations struggle to centralize their data, and less than a quarter (23%) report robust GPU capacity, the race changes course. It no longer resembles a sprint for sparse silicon, but looks more like a rush-hour gridlock of poorly connected clusters.

Laurie Maclachlan

Head of International at vCluster.

cloud providers commanding roughly 85% of the European market – and sovereign cloud spending forecast to more than triple by $23 billion by 2027 – the pressure to secure control is well and truly on.

GPU scarcity is real. But much of today’s constraints are the result of fragmented Kubernetes environments, duplicated clusters, and IT infrastructure that was simply never designed for secure, efficient scaling at the scale required by AI. For many organizations, Kubernetes has become the default system for running modern applications.

It acts as an orchestration platform, managing containers, allocating compute resources, and keeping everything ticking along reliably across servers. These are grouped into clusters: pools of computing power that teams use to run workloads.

The problem is that Kubernetes was designed in a pre-AI era, when workloads were significantly lighter. It is far less optimized for sharing scarce, high-value GPU resources, and it’s for this reason that the cracks are starting to show.

The hidden (and growing) cost of cluster sprawl

In the rush to deploy AI safely, many organizations have defaulted to isolation by duplication. That means spinning up entirely separate clusters and environments for each team or workload, rather than securely sharing the same underlying infrastructure.

In practice, should one team needs autonomy, they get their own cluster. Should another team’s workload be sensitive, it is separated physically rather than logically. If governance is complex, the easiest answer is to ringfence infrastructure instead of rethinking how it’s shared. The result is cluster sprawl on a huge scale.

While each cluster may feel like a sensible risk-reduction decision, the reality is that it becomes an economic liability. GPU-backed environments are expensive to stand up, but even more expensive to leave underutilized. Left fragmented between teams, visibility of clusters drops while idle capacity increases.

This is so much more than an operational headache. Structural inefficiencies such as these erode return on investment. GPUs should be seen as capital assets, not experimental hardware tucked away in research and development budgets. Leaving them stranded behind siloes is the equivalent to building a power plant and running it at half capacity simply because the grid cannot distribute electricity efficiently.

Kubernetes may have transformed how infrastructure is built and operated, but it was never designed with multi-tenant GPU sharing in mind. In simple terms, this allows multiple teams or workloads to share the same underlying infrastructure safely, all with clear boundaries, isolation and governance without duplicating the hardware itself.

Without true multi-tenancy, shared environments creak under AI workloads. It leaves a choice between compromising on isolation or retreating to dedicated clusters. Neither scales, and the model is untenable. It’s for this reason that the narrative of GPU shortage is obscuring a much deeper flaw.

Sovereignty heightens the tenancy debate

Moving workloads either on-prem or into sovereign environments does not magically solve inefficiency. In fact, it can amplify it. When organizations bring AI infrastructure closer to home to meet regulatory, security or geopolitical concerns, they assume full responsibility for its utilization. This ownership, without orchestration, cannot be considered sovereignty, however. It is simply overhead.

As AI workloads continue to accelerate, the race won’t be towards hardware acquisition but secure deployment and efficient monetization. That requires a control layer purpose-built for the economics of AI, such as software that sits above the raw infrastructure and governs how compute is allocated, shared and isolated.

performance while unlocking higher utilization.

With sovereignty so high on the agenda, multi-tenancy will reposition resilience as a built-in process. Rather than respond to capacity constraints by spending evermore eye-watering sums, organizations can optimize what they already own. New workloads shouldn’t justify new clusters when it can be a shared asset that’s governed intelligently.

In any race, raw speed is only part of the equation. Control, coordination and the ability to navigate complexity often determine whether momentum is sustained. AI infrastructure is no different. Until it is designed for secure, shared scale, no amount of raw speed will deliver its true potential.