Blog

NVIDIA and Nebius ran joint benchmarks using NVIDIA Run:ai, the AI workload orchestration and optimization software platform on the Nebius AI Cloud. The goal was simple: test whether fractional GPU allocation could improve efficiency and scalability for real-world inference workloads — without compromising performance.

When inference becomes distributed, routing strategy can determine whether a system scales or stalls. In this article, we examine a real agent-style workload where cache-aware routing in vLLM reduced average step time by nearly 50 percent and cut P95 latency from over a minute to under 20 seconds — with the same model, hardware and traffic.

As AI adoption accelerates, energy use increasingly sets the boundaries of how far the systems can scale. Power availability, efficiency and infrastructure design are becoming practical constraints. This shift is prompting to think of concrete ways to manage the energy footprint of AI systems by optimizing energy use and creating measurable efficiency gains. In our latest whitepaper, we explain how Nebius improves efficiency across the stack, from software engineering to hardware design and data center operations.

We recently introduced support for exporting billing data from Nebius AI Cloud in the FOCUS format. This update is a small but important step toward making cloud integration simpler and financial operations smoother for teams building and scaling AI workloads. At Nebius, we believe billing data should be easy to work with, easy to integrate and easy to trust — especially when AI infrastructure costs are a core part of your business model.

Large mixture-of-experts language models promise significant gains in model quality, but deploying them in real products often exposes hard latency limits. This article explains why MoE systems that look efficient in benchmarks struggle under production constraints, and how architectural decisions around routing, batching and serving determine whether latency budgets hold under worst-case inputs and real user behavior.

While reinforcement learning drives agents to state-of-the-art performance, rejection fine-tuning serves as a powerful baseline. Stemming from our extensive experiments with different models and scaffoldings, we are sharing a dataset of 67k high-quality OpenHands trajectories from Qwen3 Coder 480B for research purposes. We also include two RFT checkpoints — Qwen3 Instruct 30B and 235B, achieving ~50% and ~60% on SWE-bench Verified respectively.

Bioscience seeks to unravel the structure of protein complexes that drive nearly every cellular process. Open-source tools and data banks are essential for accelerating advances in this field. With support from Nebius AI Cloud and a Nebius Academy fellowship, researchers at the University of California, San Francisco are adapting widely used protein modeling software to modern GPU cloud platforms, expanding access and impact for the scientific community.

This guide teaches the essentials of serving large language models with vLLM. It builds from foundational neural network concepts, like transformers and attention, to introduce practical inference workflows, explore useful vLLM features and offer hands-on, operational guidance for efficient deployments.

The Nebius AI Cloud 3.1 release continues the evolution of our platform into a transparent, orchestrated, developer-first environment — where performance is accessible, capacity is visible and control feels effortless. Across new platforms based on NVIDIA Blackwell Ultra, smarter capacity management and an improved experience for both developers and platform operators, every enhancement moves us closer to that goal.

Today, we’re expanding Nebius AI Cloud’s elasticity even further with the introduction of Capacity Blocks and Capacity Dashboard — two new capabilities that bring additional clarity to GPU capacity management and give users full confidence in the compute they rely on.

Europe’s first operational NVIDIA GB300 NVL72 systems are now live in Nebius AI Cloud. The Blackwell Ultra-based platform is fully deployed in our Finland data center, delivering supercomputer-class performance for large-scale training and high-throughput inference. This rollout expands the compute foundation available to teams building next-generation models across Europe.

In this blog, we will take a look at how teams can deploy the four NVIDIA NIMs available now in our platform, how they can be consumed via their API and how they can fit into larger scientific workflows. It also covers the Virtual Screening Blueprint, which turns folding, molecule generation and docking into a ready workflow that becomes available in JupyterHub shortly after deployment.