We’re launching a new Nebius platform built from the ground up

When we first launched Nebius, we started by using what we had learned from building a general public cloud, which helped us get up and running quickly. Now we see better ways to do things, so we have built a brand new, efficient and user-friendly AI cloud platform.

Our new platform features a faster storage backend, support for new GPUs and our latest ML services, better observability and a more intuitive UI. With a strong focus on AI needs, it provides enthusiasts and ML practitioners with robust, functional environments for their ambitious initiatives. By the way, we sometimes call it Newbius, the new Nebius, you know.

Faster storage for better performanceFaster storage for better performance

Storage is crucial — especially for ML training. To better support AI workloads, we’ve made technical tweaks and low-level upgrades to our file storage, boosting performance to up to 100 GBps and 1M IOPS for aggregated read operations. Here’s what’s changed:

• We increased filesystem throughput by removing architectural bottlenecks in its internal structure.

• Our file storage now has higher read throughput and lower latency thanks to extending the minimum data chunk size. These improvements reduce IO to CPU, freeing up computational resources.

• We redesigned the way files and their metadata are transferred to the Filestore, making it faster with parallel downloading.

• Optimized settings now allow for faster work with Pytorch Dataloader.

These changes ensure seamless data streaming during model training and prevent disruptions when saving checkpoints, uploading model code or sharing model weights across cluster nodes.

New GPUs, new possibilitiesNew GPUs, new possibilities

The new platform allows access to NVIDIA H200 Tensor Core GPUs, which are coming to our server racks these days. Stay tuned to learn more about these offerings.

We’ve also implemented several network changes defining how our cloud communicates with our physical sites, paving the way for the smooth integration of new data centers into our infrastructure.

Easier access to ML servicesEasier access to ML services

We’ve recently launched two new managed services to improve our customer’s ML operations: Managed Spark™ and Managed MLflow. Both are now available on our new console.

Managed Service for Apache Spark is a fully managed data processing engine designed to simplify and accelerate data engineering and ML workloads. Apache Spark is renowned for its speed and ease of use in big data processing.

Managed Service for MLflow is a fully managed, industry-leading tool for managing the ML lifecycle. It collects and stores key metrics and parameters of machine learning iterations, tracks experiment runs and helps distill the best-performing models for further deployment.

Improved observabilityImproved observability

We’ve made significant improvements to observability on the new platform. You now have real-time access to key hardware metrics on service dashboards. There’s no need to connect external tools like Grafana — everything you need is at your fingertips.