Anyscale

Anyscale is a commercial platform built around Ray, the open-source distributed computing engine that came out of UC Berkeley's RISELab. Teams use it to run Python workloads such as multimodal data curation, batch embedding generation, distributed model training, LLM post-training and model serving across large CPU and GPU pools. Developers work in managed workspaces with VS Code or JupyterLab, then move code into production as Jobs or long-running Services using the Anyscale CLI and SDK. There are two ways to deploy it. The Anyscale-hosted option is serverless, is the default for new organizations, runs in a limited set of regions, and is billed monthly by credit card. Bring Your Own Cloud runs Ray clusters inside the customer's AWS or Google Cloud account, or on Kubernetes through EKS, GKE, AKS or on-premises clusters. Anyscale also lists Nebius and CoreWeave as supported compute. For Azure there is a Terraform module. Pricing is usage-based, metered in Anyscale credits per instance-hour. Published rates run from 0.0135 credits per hour for CPU to 4.9591 for an NVIDIA A100, while H100 and Blackwell-class GPUs require a sales quote. New accounts receive $100 in credits, and committed contracts with volume discounts are available. BYOC can be invoiced through the AWS, Azure or GCP marketplaces. Anyscale suits ML and data teams that already use Ray, or plan to, and need to scale across many GPUs. Its main limitation is that it is focused on Ray. It is not a general web hosting platform, reviewers describe a learning curve and hard-to-predict costs, and an independent comparison estimates that managed clusters cost noticeably more than bare-metal GPU rental. In July 2026 Nscale agreed to acquire Anyscale, and the companies said the Anyscale brand will continue.