Why EU CSPs Are Launching AI GPU Clouds

The GPU-as-a-Service market is $8.2 billion today. MarketsandMarkets puts it at $26.6 billion by 2030 – a CAGR of 26.5%, one of the fastest-growing infrastructure categories on record.

That is not a prediction. That is a pipeline. And right now, the EU portion of it does not have a clear owner.

Here is the uncomfortable truth for most EU cloud service providers: the customers you already serve are running AI workloads somewhere. If it is not on your infrastructure, it is on AWS, Azure, or a neocloud you have never heard of. The GPU cloud category is being built with or without regional providers. The only question is whether EU CSPs show up while the market is still open, or arrive later to compete against incumbents who moved first.

The case for moving now is not complicated. But it does require understanding three things that most CSP conversations miss.

Inference changed everything, and most CSPs haven’t caught up

---

Everyone assumed GPU cloud was a hyperscaler game. And for training, they were right.

Training frontier AI models requires tens of thousands of GPUs running in tightly coupled clusters for weeks at a time. That is AWS territory. No regional CSP was going to compete for that.

But training is not where the market is going. Inference is.

Inference is what happens after a model is trained – every time a user gets a response, a document gets summarized, a recommendation gets generated. It runs continuously, at distributed scale, close to users. It does not need a 10,000-GPU hyperscale cluster. It needs reliable, low-latency GPU endpoints that run 24 hours a day.

The shift is already underway and the numbers are decisive. Gartner projects 55% of AI-optimized infrastructure spending supports inference workloads in 2026, rising to 65%+ by 2029. McKinsey puts inference demand growing at 35% CAGR through 2030 – faster than training, faster than traditional compute. By 2030, inference alone will account for more than 40% of total global data center demand.

This is not a niche workload. It is THE workload. And, it is structurally suited to exactly what EU CSPs already operate: sovereign, latency-sensitive infrastructure, close to end users, running in jurisdictions that matter.

The hyperscaler dominance in AI infrastructure is real, but it is training-era dominance. Inference opens the map. EU CSPs are sitting on exactly the right infrastructure profile at exactly the right moment – but most of them are watching it happen from the sidelines.

Compliance is not a constraint. It is a product.

---

Here is something hyperscalers cannot fix with a press release.

EU enterprises are not just shopping for cheap GPUs. They are shopping for GPUs they are legally allowed to use for their workloads. Those are different markets.

GDPR restricts how personal data moves outside the EU. The EU AI Act – now in full enforcement for high-risk systems – creates documentation, auditability, and transparency requirements that are materially easier to satisfy on infrastructure you control and can verify. Get it wrong and you are looking at penalties up to €35 million or 7% of global revenue. Stack GDPR and the AI Act together and a single compliance failure can reach 11% of global turnover.

The result: 20% of European companies have already begun repatriating business-critical data to local facilities. 61% of Western European CIOs are now actively prioritising local cloud providers: not out of preference, but out of board-level risk management.

And here is the part that the hyperscaler sovereign cloud announcements miss. Choosing a European region on AWS addresses data residency – where the server physically sits. It does not address data sovereignty – who controls the infrastructure and under whose law it operates. The US CLOUD Act means that data held by US-headquartered entities remains subject to US legal jurisdiction regardless of where the rack is. For defence contractors, financial institutions, healthcare systems, and public sector agencies, that distinction is not theoretical. It is the deciding factor.

EU CSPs don’t have this problem. They are European entities, operating European infrastructure, under European law. That is not a compliance checkbox. That is a genuine competitive moat, one that AWS cannot replicate with a €7.8 billion investment announcement, however large the number.

The sovereign cloud market is projected to grow from $154 billion in 2025 to $823 billion by 2032. North America currently holds 88% of neocloud GPUaaS revenue. That share drops to 72% by 2030 as sovereign cloud initiatives build out in Europe and beyond. That redistribution does not have a predetermined winner. It goes to whoever has built credible sovereign GPU infrastructure when enterprise procurement opens in force.

The economics: what is actually true and what isn’t

---

GPU cloud is a large market. Whether it is automatically a good business depends entirely on one variable: utilization.

This is worth being honest about, because a lot of the narrative around GPU cloud economics glosses over the hard part.

H100 on-demand pricing across providers currently ranges from under $2/hr on spot neocloud instances to $6.88/hr on AWS and $12.29/hr on Azure – a six-fold spread depending on provider, commitment model, and availability. The pricing is not a simple hyperscaler-vs-neocloud story. It is a capacity, commitment, and utilization story.

An H100 costs $25,000-$30,000 to buy. At $2/hr rental, you need roughly 12,500 hours of billable time to recover the hardware cost alone, before power, cooling, depreciation, or operations. At 60% utilisation, that is just over 20,000 hours. At 40%, closer to 30k+. The operators making strong margins on GPU cloud are not doing so because the price per hour is attractive. They are doing so because they have solved the utilization problem, consistently filling capacity across a diverse enough customer base that the economics work.

That is the real insight, and it cuts both ways. For a CSP entering GPU cloud without owned hardware – accessing capacity from a sovereign supply network and building the commercial layer on top – the utilisation risk sits elsewhere. The business model becomes: how well can you match supply to demand, and how efficiently can you fill the hours you have committed to? That is a commercial and operational challenge. It is not a CapEx challenge.

The customers who make the economics work are not the ones hunting spot pricing. They are enterprises with compliance requirements, public sector agencies with data residency mandates, and platform companies with consistent inference workloads running 24 hours a day. These buyers are less price-sensitive, sign larger contracts, and churn at structurally lower rates, because switching GPU infrastructure for a compliance-certified AI workload is a significant operational decision, not a one-click migration.

EU CSPs that position correctly – sovereign, compliant, locally operated – are competing for exactly those buyers. That is a better customer profile than the commodity GPU spot market, and it supports pricing that the utilization model can actually sustain.

The EU players who moved early, and what they understood

---

The EU CSPs already in the GPU cloud did not wait until they had the perfect product. They moved on to a specific segment with a minimum viable offer and built from there.

• Scaleway secured first-to-market European availability of NVIDIA’s Blackwell Ultra B300 GPUs and positioned it explicitly around sovereign compliance: EU jurisdiction, GDPR-compliant, no CLOUD Act exposure. The GPU is the product. The compliance wrapper is the differentiator.

• OVHcloud published reference architectures for deploying large-scale AI models in sovereign EU infrastructure, giving enterprise buyers a concrete, auditable deployment path that no US hyperscaler can offer with equivalent legal clarity.

• STACKIT built the GPU offering as a natural extension of its compliance-first positioning into German enterprise and public sector – segments where BSI C5 certification matters more than whether you have the latest silicon.

Different entry points. Different customer segments. Same underlying insight: the GPU market in Europe is not won on hardware specs. It is won on trust, proximity, and legal clarity, the three things EU CSPs have always sold.

None of them waited for a complete product. None of them needed to own a hyperscale GPU fleet. They found a specific segment where their existing positioning gave them the right to win, and moved into it.

The window is open. It will not stay that way.

---

Europe’s data center GPU market stood at $10.6 billion in 2024. It is projected to reach $82.2 billion by 2034 (ResearchAndMarkets). How that revenue is distributed across provider types is still being decided. It will be determined by which providers have credible GPU products when enterprise procurement cycles open. Those cycles are opening now, driven by the EU AI Act’s enforcement timeline and the mainstreaming of inference-heavy AI applications across every vertical.

Two forces are compressing the timeline faster than most CSP roadmaps assume.

• First: the EU AI Act’s full enforcement for high-risk AI systems has begun. Enterprises that have been deferring AI infrastructure decisions on compliance grounds now have a concrete deadline forcing the conversation. Sovereign GPU infrastructure goes from “nice to have” to “required for sign-off” in regulated sectors, almost overnight.

• Second: the hyperscalers are coming. AWS has committed €7.8 billion to a European Sovereign Cloud. Microsoft has deployed Sovereign Private Cloud across France and Germany. Google has secured NATO contracts for AI-enabled sovereign services. These moves validate the market for enterprise buyers and raise the competitive bar simultaneously. EU CSPs that are not in the GPU market before this hyperscaler investment lands (at scale) will find themselves competing against sovereign offerings from the three largest cloud operators in the world.

The window is not indefinitely open, but it is open now.

EU CSPs built their businesses on being the trusted, local, compliant alternative to hyperscalers. That positioning has never been more commercially valuable than it is in AI infrastructure right now. The market is asking for exactly what regional operators offer. The infrastructure complexity that previously made GPU cloud a specialist undertaking has been substantially reduced. The supply chain is accessible. The customer demand is demonstrable.

The CSPs who move in the next twelve months will be operating from a structurally different position than those who move in twenty-four. First-mover in a compliance-defined market segment is not a temporary advantage. It compounds.