Nebius Brings Up a Vera Rubin Rack in Finland, With Customer Readiness Still Ahead

Early access is not a unique market position

Nebius’s claim is to an early deployment, not to being first. CoreWeave, a Nasdaq-listed AI cloud provider, announced on June 1 that it had completed what it called the industry’s first bring-up and system-level validation of the full Vera Rubin NVL72 rack architecture. In its release, CoreWeave said its production-scale approach includes rack-level cooling controls, unified control hardware, multi-rail networking and tenant-isolation features. Those are company statements, but they also show the operating layers that separate a validated rack from a repeatable multi-tenant service.

The broader comparison is tighter still. NVIDIA said Rubin-based products would be available from partners starting in the second half of 2026, naming Amazon Web Services, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure alongside CoreWeave, Crusoe, Lambda, Nebius, Nscale and Together AI. That list makes the relevant competitive question less about who has photographed a first rack and more about who can deliver capacity, software and economics at scale.

NVIDIA has projected up to 10 times higher inference throughput per watt and one-tenth the cost per token versus Blackwell for NVL72; it also says large mixture-of-experts models can be trained with one-quarter as many GPUs. Those are NVIDIA performance and cost claims, not results from Nebius’s Finnish deployment. Neither Nebius nor an independent benchmark in the retained material supplies an equivalent workload measurement for the rack, so the claims cannot establish Nebius’s customer economics or a direct comparison with competing clouds.

NVIDIA has money, as well as hardware, in the buildout

NVIDIA’s interest in Nebius is financial as well as technical. A March 11 Nebius filing says the company entered a securities purchase agreement with NVIDIA to sell a pre-funded warrant for approximately $2 billion in gross proceeds. The warrant covers 21,065,936 Class A ordinary shares at an exercise price of $0.0001 per share, subject to its terms and adjustments.

Nebius said it intended to use the net proceeds for investments in its full-stack AI cloud and in greenfield data-center development. That makes the financing relevant to its ability to fund infrastructure, but it is not a disclosed budget for the Finnish rack or proof that planned capacity has been installed. The filing also does not turn a deployment timetable or a performance claim into an achieved result.

The company’s Finland plans illustrate the same distinction. A report says Nebius is constructing a 310-megawatt AI factory in Lappeenranta and has a stated aim of more than 5 gigawatts of NVIDIA capacity by 2030. Those figures describe a reported construction project and a company target, respectively; they should not be conflated with the capacity of one validated NVL72 rack.

What would show commercial readiness

The next evidence should be operational rather than ceremonial. Nebius needs to complete the software and production testing it identified, then show when customers can use Rubin through AI Cloud or Token Factory, under what service and price terms, and with what workload results.

For a technology vendor, the comparison NVIDIA offers is between product generations; for customers, the comparison will be between accessible clouds and the cost and reliability of actual workloads. A clear availability date, independently comparable performance data, and evidence that the Finnish integration can be repeated across more capacity would answer whether Nebius has converted an early rack deployment into a competitive service.