Blog

Nebius and Positronic are running an open competition in robot manipulation on the DROID setup, a Franka arm with a Robotiq gripper. Teams train on Nebius AI Cloud and compete in simulation until 13 December; the top 10 policies then run on a real arm in December. Prizes total $20K+ in Nebius GPU credits and Positronic evaluation credits, and no team needs a robot of its own.

A first-person recap of Nebius Build Week and GTC Berlin from developer advocate Marouane Khoukh. The Sunday hackathon at CODE University drew more than 250 builders and produced 74 projects, including a children’s news show scripted with NVIDIA Nemotron on Nebius Token Factory and a robot arm whose team found real demonstrations made the biggest difference.

We benchmarked NVIDIA Cosmos 3 Super text-to-video serving through vLLM-Omni across four topologies on NVIDIA HGX B200 and NVIDIA HGX H200 nodes on Nebius AI Cloud. On HGX B200, eight single-GPU replicas produced 42.5% more valid video per node-hour than one 8-GPU replica, at 5.5 times the mean latency. The post includes the full method and an open-source workflow to rerun it.

The compromise of eight eMudhra intermediate certificate authorities may affect systems and workloads that still trust the affected certificate chains. In this post, we explain how Nebius is addressing the issue across its services, what mitigations have already been applied or scheduled, and what actions customers should take to protect existing environments.

Nebius now runs three new NVIDIA retail agent blueprints, and only one needs a GPU cluster the team manages. Inventory Lab runs supply-chain experiments as Slurm jobs on NVIDIA H100 GPUs. The store operations agent runs on a CPU virtual machine with models served by Nebius Token Factory; the fine-tuning flywheel runs as serverless jobs, with no cluster to provision.

Customers using Nebius Physical AI Workbench report up to 40% less engineering time with AI agents operating robotics workflows directly. Built in collaboration with NVIDIA and open source on GitHub, it runs data generation, simulation, training, and evaluation in one place. This release adds NVIDIA Cosmos Data Factory and a sim-to-real workflow co-designed with Sereact for Zalando returns.

NVIDIA and Nebius are expanding the AI Builder Program into Europe with $10 million in compute grants across Nebius AI Cloud and Token Factory, founder funding opportunities through Accel and our broader European VC network, and hands-on community programs. Applications are open.

Three new open-weight models, Reflection Beam, Meta’s Muse Spark Open, and NVIDIA Nemotron 3.5 Super VL, are now available on Nebius Token Factory. Built for different agent workloads, they span coding and reasoning, multi-file coding and tool use, and visual understanding for documents and web content. All three are served through the same OpenAI-compatible API, making it easier to compare quality, latency, and cost on real tasks.

Nebius Academy and NVIDIA are launching an education collaboration at GTC Berlin, featuring: NVIDIA Deep Learning Institute (DLI) open-source courses hosted on Nebius Academy starting with Fundamentals of Deep Learning, co-developed courses in agentic and physical AI that run on Nebius, and Nebius GPU compute behind NVIDIA DLI’s sold-out full-day workshop at GTC.

NVIDIA Run:ai platform is now open source, and is available in Nebius AI Cloud Applications for deployment on Nebius Managed Service for Kubernetes. This post covers what’s in the release (fractional GPU allocation, gang scheduling, hierarchical quotas) and the joint Nebius–NVIDIA benchmarks: 77% of full-GPU throughput at half a GPU allocation.

Nebius has achieved NVIDIA Exemplar Cloud on NVIDIA HGX B300 for training workloads, adding to existing Exemplar recognition on NVIDIA H200 and GB300 NVL72. The validation reflects ongoing engineering collaboration with NVIDIA to test and optimize production infrastructure for demanding AI training workloads.

The team behind Nebius Managed PostgreSQL explains how and why it was built specifically for AI workloads.

We describe how it’s built on CloudNativePG and Kubernetes, why replacing Barman with WAL-G cut a 1.5 TB backup from more than a day to 2 hours, and how PgBouncer, point-in-time recovery, pgvector, and pgvectorscale fit bursty AI workloads.