Nvidia Corp. today announced plans to invest $2 billion in Nebius Group NV, a Dutch cloud provider that specializes in artificial intelligence workloads.

The capital will help the company expand its data center infrastructure. Nebius plans to deploy servers equipped with Nvidia’s latest Rubin graphics processing units and Vera central processing units. It will also adopt the BlueField chip series, which offloads auxiliary tasks such as managing storage hardware from a server’s CPU.

The most advanced Nvidia graphics card series that Nebius currently offers to its customers is Blackwell. Rubin, the new GPU that Nebius is adopting, runs inference workloads 10 times more cost-efficiently.

Vera, the CPU that Nebius will deploy alongside Rubin, is also more efficient than its predecessor. Appliances equipped with both chips reportedly provide 10 times better performance per watt than Nvidia’s previous-generation hardware. The systems, which weigh about two tons, include 72 Rubin GPUs, 32 Vera CPUs and more than 1,000 other chips.

The company hosts its infrastructure in internally operated data centers and co-location facilities. The company’s largest site is a 300-megawatt cloud campus currently under construction in New Jersey. Its data center network also spans the U.K., France, Finland and Iceland. 

Nebius offers hosted GPUs alongside managed services designed to ease AI model development. It provides a managed version of MLFlow, an open-source tool that analyzes telemetry from AI training runs to identify inefficiencies. Another Nebius service offers a cloud-hosted implementation of Jupyter Notebook, a code editor that is widely used in AI projects.

In conjunction with Nvidia’s investment, Nebius today made the chipmaker’s Nemotron 3 Super model available on its cloud. It’s a mixture-of-experts algorithm optimized to power AI agents. Nemotron 3 Super can ingest prompts with up to 1 million tokens and activates one-10th of its 120 billion parameters to generate answers.

The model is available via a Nebius service called the Token Factory. The offering extends open-source AI models with an autoscaling feature and safety guardrails. In addition, there’s support for a performance optimization method called speculative decoding. The technology speeds up inference by enabling AI models to offload some calculations to a smaller, more hardware-efficient neural network. 

Nebius plans to deploy more than 5 gigawatts of additional computing capacity by the end of the decade to power customer workloads. Besides supplying hardware for the initiative, Nvidia will also provide AI cluster design materials, “technical reviews” and other forms of support.

The investment comes a few weeks after the GPU maker inked a similar deal with CoreWeave Inc., another AI cloud operator. Nvidia agreed to invest $2 billion in the startup after earlier placing a $6.3 billion order for cloud infrastructure. CoreWeave plans to deploy more than 5 gigawatts of computing capacity by 2030.

Photo: Nebius