Nvidia's European cloud partners scale up, and start fencing in agents

At GTC Berlin, Nvidia used its European showcase to argue that regional AI clouds are no longer a sovereignty talking point but an operating requirement: model builders want compute next to their data, open weights they can tune, and infrastructure that constrains what an agent can touch.

The blog post accompanying the announcements frames the demand as bottom-up. AI-native companies are pulling capacity into Europe, the Nordics, Africa and Central Asia, and cloud partners are answering with Nvidia's DSX AI Factory stack, including Vera Rubin systems, BlueField DPUs and Spectrum-X Ethernet, rather than with generic GPU rentals.

Power, not silicon, is the binding constraint in several of these markets, so Nvidia led with efficiency numbers. Nscale tested DSX MaxLPS on GB300 NVL72 systems and reported 37% more GPUs and close to 50% higher aggregate inference throughput inside the same provisioned power budget, with Vera Rubin capacity planned across five European markets.

Crusoe is building Spark, a liquid-cooled modular AI factory aimed at Vera Rubin and at inference close to users; its EMEA customers include Base44, Gradium, LTX and Wonderful. Aleria, Cassava Technologies, Firebird, G42, Radiant, SWI Compute and Vultr are expanding in parallel across Europe, the Nordics, Africa and Central Asia.

The more instructive details sit one layer up. Nebius adopted Nvidia Groq 3 LPX with Spectrum-X networking on its Token Factory to cut token generation latency, and Baseten trained a code-review routing model on Nemotron 3.5 Lightning that CodeRabbit uses: routing agreement rose from 75.8% to 80.7% while estimated inference cost for the evaluated workload fell about 50%.

Model builders are matching the infrastructure. Mistral trained Mistral Large 4 on 3,800 Grace Blackwell GPUs in its own European data centers and plans Vera Rubin delivery in 2027; Fireworks and Era4 are pairing a training and inference platform with Blackwell Ultra liquid-cooled capacity at Era4's Chesterfield site for financial services, insurance and life sciences.

The safety layer is the part that reads as new. Nvidia announced an OpenShell reference architecture in which Vera Rubin NVL72 handles inference, Vera CPUs run agent harnesses and tool calls, and BlueField-4 DPUs enforce policy in silicon via Nvidia Sentry; CoreWeave, DigitalOcean, Lambda, Nebius and Nscale are among roughly twenty partners backing it.

For developers, the shift that matters is not the GPU count but the ability to run agents where the data lives, inside boundaries someone can audit. Europe's sovereign-AI contest is moving from whether a country has accelerators to whether it can constrain what those accelerators' agents do, and the next test is whether this capacity comes online and holds up under enterprise workloads.