Anyscale
by Anyscale (being acquired by Nscale)
Managed Ray platform for scaling AI data processing, training, inference and RL across thousands of GPUs on any cloud
Anyscale is the managed compute platform from the creators of Ray, the open-source framework for scaling Python and AI workloads. It is for ML platform teams and AI-native companies that need to run multimodal data processing, distributed training, batch inference, serving and reinforcement learning across large GPU fleets on multiple clouds without building that orchestration themselves.
Anyscale is the commercial platform built by the team behind Ray, the distributed computing framework that Anyscale reports has 500M+ all-time downloads, 41K+ GitHub stars and 1.2K+ contributors, and which was donated to the PyTorch Foundation in 2025 where it remains community-governed. The platform wraps Ray with developer workspaces, jobs and services, observability dashboards, lineage tracking, a Global Resource Scheduler and Multi-Resource Cloud scheduling that pools GPUs across regions and providers (AWS, GCP, Azure, Nebius, CoreWeave). Its proprietary Anyscale Runtime, generally available since Ray Summit 2025, is a Ray-compatible engine Anyscale claims delivers 6x lower cost for image inference, 10x faster feature preprocessing and 40% faster video serving. Customers run multimodal data curation, batch embeddings, distributed training and post-training/RL frameworks such as SkyRL and veRL, integrating with PyTorch, vLLM, SGLang and XGBoost. It comes as a hosted service or Bring-Your-Own-Cloud into the customer's VPC or on-prem Kubernetes, with SSO/SAML/SCIM, audit logs, SOC 2 Type II and ISO 27001. Named customers include Coinbase, Character.ai, TwelveLabs, Physical Intelligence, Runway, Grab, TripAdvisor and Recursion. On July 30, 2026 neocloud Nscale agreed to acquire Anyscale for about $1.65 billion; TechCrunch reported Anyscale will keep its brand and customers, its roughly 200 staff join Nscale, and the platform stays free to run on non-Nscale infrastructure, with closing expected in the second half of 2026.
Heads of ML platform or AI infrastructure at companies training, post-training or serving their own models across multi-cloud GPU capacity.
Run Ray-based data, training and inference pipelines at thousands-of-GPU scale with managed orchestration, observability and cross-cloud GPU pooling.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Usage-based, Contact for pricing
- Target Market
- CTOs, ML Platform Engineers, AI Infrastructure Teams, Data Scientists
- Deployment
- Multi-cloud, Hybrid, Cloud-first
- Team Size
- 201-500
Key Features
- ✓Anyscale Runtime
Ray-compatible proprietary engine that Anyscale claims cuts image inference cost 6x and speeds feature preprocessing 10x.
- ✓Multi-Resource Cloud and Global Resource Scheduler
Pools and queues GPU capacity across regions and providers so jobs land wherever reserved compute is free.
- ✓Bring Your Own Cloud
Deploys into your own VPC or Kubernetes on any cloud or on-prem, reusing existing GPU reservations and marketplace billing.
- ✓Observability and lineage
Ray Data, Train, Task and Cluster dashboards plus lineage graphs mapping datasets to models make distributed failures diagnosable.
- ✓Workspaces, jobs and services
Develop interactively in workspaces then promote the same Ray code to production jobs or services without rewriting.
- ✓Enterprise governance
SSO, SAML, SCIM, audit logs and multi-team access controls, backed by SOC 2 Type II and ISO 27001 certification.
Capabilities
Use Cases
- •Multimodal data curation
AI labs filter, transcode and embed massive video and image corpora for training, scaling across thousands of CPUs and GPUs.
- •RL post-training
Model builders run reinforcement-learning post-training with frameworks like SkyRL and veRL on elastic Ray clusters instead of bespoke orchestration.
- •Batch embedding generation
Search and RAG teams generate embeddings for billions of documents as scheduled batch jobs on pooled, spot-friendly GPU capacity.
- •Fraud and risk ML pipelines
A PeerSpot reviewer uses Anyscale for fintech fraud-prevention workflows, citing reliable scaling and GPU-memory monitoring dashboards.
- •Cost reduction on spot capacity
G2 reviewers describe Ray decorators spinning up clusters of 50+ spot instances, cutting compute costs by roughly 60 percent.
Ideal For
Best For
- ✓Multimodal data curation over video, image, audio and text at petabyte scale
- ✓Distributed model training and reinforcement-learning post-training (SkyRL, veRL)
- ✓Large batch inference and embedding generation jobs on pooled GPU capacity
- ✓Teams already standardised on open-source Ray that want managed clusters and governance
- ✓Organisations with GPU reservations spread across several clouds that want one scheduler
Not Ideal For
- ✗Teams that only need a hosted model API for a chatbot — a serverless inference provider is far simpler than running Ray pipelines
- ✗Buyers wary of vendor ownership changes — the pending Nscale acquisition adds uncertainty about long-term neutrality and roadmap
- ✗Small teams without distributed-systems skills, since partial failures, OOMs and debugging across distributed logs are inherent to Ray workloads
Integrations
Deployment
Market Analysis
Pros
- ✓Deep Ray expertise from the framework's creators and a huge open-source install base
- ✓Published per-hour pricing and $100 credit make evaluation easy
- ✓True multi-cloud and BYOC deployment with SOC 2 Type II and ISO 27001
- ✓Blue-chip AI-native customers including Coinbase, Runway and Character.ai
Cons
- ✗Pending Nscale acquisition raises neutrality questions; Futurum flags customer hesitation and talent-retention risk
- ✗Distributed Ray workloads bring partial failures and OOM debugging across scattered logs — Anyscale's own blog acknowledges this
- ✗Reviewers note beginner-unfriendly documentation, occasional platform lag and no built-in CI/CD integration
- ✗Latest GPUs (H100/B-class) are not list-priced, so costs need a sales conversation
Pricing
Pay-as-you-go (Hosted)
From $0.0135/hr (CPU); T4 $0.5682/hr, A100 $4.9591/hr
- ✓$100 starting credit
- ✓Anyscale-managed compute in limited regions
- ✓Business-hours support, 5 cases
BYOC / Committed contract
Contact for pricing
- ✓Any cloud, region or on-prem VPC/Kubernetes
- ✓Use existing GPU reservations
- ✓AWS/Azure/GCP marketplace billing
- ✓24x7 SLAs, unlimited cases
Hosted usage is billed per instance-hour with published rates (CPU $0.0135/hr, NVIDIA T4 $0.5682/hr, A10G $1.3635/hr, A100 $4.9591/hr); H100/B200/GB-class GPUs are contact-for-pricing. BYOC and committed-use contracts with volume discounts and 24x7 SLAs are negotiated with sales; new users get $100 credit.
Security & Compliance
Connect
Sources
This page was written from 9 sources, 6 on domains other than anyscale.com.
- 1.anyscale.com — anyscale.comvendor
- 2.anyscale.com — pricingvendor
- 3.anyscale.com — ray summit 2025 anyscale product updatesvendor
- 4.trust.anyscale.com — trust.anyscale.com
- 5.docs.anyscale.com — certifications
- 6.techcrunch.com — nscale buys anyscale as it seeks to own more of the ai compu
- 7.nscale.com — nscale acquires anyscale
- 8.futurumgroup.com — nscale acquires anyscale the neocloud land grab continues
- 9.peerspot.com — anyscale platform reviews
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Chroma
Open-source (Apache 2.0) vector and hybrid search database for AI, with a serverless Chroma Cloud on object storage
LiteLLM
Open-source AI gateway: 100+ LLM APIs behind one OpenAI-compatible endpoint, with cost tracking and guardrails
CIQ Fuzzball
Sovereign AI and HPC orchestration: train, fine-tune and serve models on infrastructure you control
Crusoe Cloud
Vertically integrated AI cloud — own the power, own the datacentre, rent the GPUs by the minute