Nebius and Anyscale have partnered to offer deeper platform integration and enable teams to deploy and scale Python and AI workloads with Ray easier, faster and cheaper. The partnership combines Nebius AI Cloud and Anyscale’s managed Ray platform, delivering a complete stack for developers and platform teams to move multimodal AI — across data processing (structured and unstructured), training and serving — from code to production.

Multimodal AI is the new normalMultimodal AI is the new normal

Modern AI no longer trains on tidy, structured datasets. Today’s systems learn from and act on vast amounts of video, images, audio and sensor data — often all at once and in real time. This shift toward multimodal introduces new infrastructure challenges that traditional environments weren’t built for. Use cases such as smart city video analytics, robotic simulation and navigation or generative agents that blend speech and vision, require data streams constantly across CPUs and GPUs.

For teams building AI pipelines — typically scaled by manually orchestrating multiple Python runtimes — this can mean spending time and effort patching together manual scripts and infrastructure pieces to optimize for efficient scheduling and processing across the heterogeneous (CPU+GPU) compute required for multimodal data. Datasets can routinely exceed memory limits, network throughput becomes a bottleneck and orchestration between compute nodes can break down under production load. The result is slow iteration cycles, idle GPUs and unpredictable performance that prevents teams — from AI researchers to robotics engineers — from scaling models reliably in production.

Ray, the open standard for scaling AI, lets developers take any workload — from simple functions like resizing an image to advanced AI steps like generating frame descriptions with an LLM — from a single machine to thousands of nodes effortlessly. It provides one cohesive fabric to scale every part of the workflow — data prep, training or inference — across any type of AI, from traditional ML to emerging agentic systems. But Ray is still a compute framework: it needs a production platform to handle infrastructure setup, proactive node replacement, developer tooling and observability knob-tuning to get peak performance from every GPU.

Nebius and Anyscale: end-to-end scalability for distributed AINebius and Anyscale: end-to-end scalability for distributed AI

To address the challenge of seamless scaling and maximize efficiency, Anyscale, the managed platform from the creators of Ray is now integrated with Nebius AI Cloud, enabling teams to build and deploy production-ready Ray infrastructure with a few clicks.

With this integration, teams can deploy Anyscale directly from the Nebius AI Cloud console or marketplace — no cluster configuration, no networking setup and no quota delays. In addition to Ray cluster lifecycle management (deployment, scaling, upgrades, etc.), Anyscale provides developer tooling and observability purpose-built for Ray as well as a performance-optimized runtime. Anyscale remains a control plane for all of your workloads across environments, while Nebius works as a data plane. The Anyscale Runtime is fully compatible with the open source Ray API, so teams already using Ray can reduce operational burden without code modifications, seamlessly integrating with Nebius AI Cloud’s managed Kubernetes infrastructure and storage layers.

The solution can also deliver significant efficiency gains, by consolidating underutilized nodes and schedule reliable execution on Nebius’ lower cost, preemptible Virtual Machines. Flexibility to combine reserved and on-demand instances ensures capacity to continuously run scheduled tasks alongside new experiments without interruption or downtime. The result is higher performance, higher utilization, simple management and production-grade reliability — whether processing petabytes of data or training multimodal models across tens or thousands of GPUs.

Get started nowGet started now