Anyscale

GPU Cloud / Infra

Anyscale, founded in 2019 by UC Berkeley researchers Ion Stoica, Robert Nishihara, Philipp Moritz, and Michael I.

Reviewed by 7wData
Share Email

On this page

Profile

Anyscale provides a managed platform for scaling AI workloads—data pipelines, model training, and inference—built on Ray, an open-source distributed computing framework.

Anyscale, founded in 2019 by UC Berkeley researchers Ion Stoica, Robert Nishihara, Philipp Moritz, and Michael I. Jordan, commercializes Ray, an open-source distributed computing framework designed to simplify AI workload scaling. The company raised $20.6M in Series A funding in December 2019 from Andreessen Horowitz and others, followed by a $40M Series B in October 2020 and a $199M Series C in October 2021, reaching unicorn status.

A Series C-II extension of $99M arrived in August 2022, bringing total capital to $259.6M. As of 2023, Anyscale reported $111.9M in annual recurring revenue at a $1B valuation. In July 2024, co-founder Robert Nishihara ceded the CEO role to Keerti Melkote, an industry veteran who founded Aruba Networks and sold it to Hewlett Packard Enterprise for $3 billion in 2015.

Nishihara transitioned to focus on product while Melkote leads commercial expansion. The company operates a platform for production-scale AI workloads, enabling data curation, distributed model training, batch embedding generation, and post-training inference on frameworks like SkyRL and veRL across AWS, GCP, Azure, and other cloud providers. Ray itself has crossed 237 million downloads, and 120,000 unique GPU clusters run the software weekly.

In October 2025, Ray was formally welcomed into the PyTorch Foundation, positioning itself as a core component of an open-source AI compute stack alongside PyTorch and vLLM. Recent product launches include Anyscale Endpoints for LLM inference and fine-tuning (June 2024), Private Endpoints for self-hosted models, and major platform updates announced at Ray Summit 2025 including Lineage Tracking, Anyscale Runtime delivering 6x cost reductions for image inference, and first-party Azure integration. The company employs approximately 400–685 people depending on source (mid-2026 estimates vary).

Notable customers named publicly include OpenAI, Apple, Netflix, Uber, Shopify, ByteDance, Coinbase, Character.ai, TwelveLabs, Runway, Recursion, Attentive, Notion, and Grab. Despite strong adoption, Anyscale faces competition from cloud hyperscalers, specialized AI platforms, and open-source alternatives, and the company has faced industry scrutiny over CVE-2023-48022, an unauthenticated remote code execution vulnerability in Ray that affects publicly exposed clusters; Anyscale maintains the issue stems from misconfiguration rather than design flaw, noting Anyscale-hosted clusters are unaffected.

Track Anyscale and 240+ vendors.

48k+ subscribers read the daily AI & data note. One email, both newsletters. Unsubscribe anytime.

Who buys this

  • Large language model and foundation model training teams at AI labs and enterprises
  • Data-intensive ML teams at tech giants (OpenAI, Apple, Netflix, Uber, ByteDance) running inference and fine-tuning at scale
  • Organizations building multimodal AI applications requiring distributed GPU orchestration across clouds
  • Enterprises and Azure customers running Ray workloads in regulated or on-premise-preferred environments

Publicly disclosed clients

  • OpenAI
  • Apple
  • Netflix
  • Uber
  • Shopify
  • ByteDance
  • Coinbase
  • Character.ai
  • Runway
  • Recursion
  • Attentive
  • TwelveLabs
  • Notion
  • Grab

Strengths and what to watch

Strengths

  • Ray commands de facto standard status in distributed AI compute with 237M+ downloads and 120K GPU clusters active weekly; PyTorch Foundation backing (October 2025) signals institutional entrenchment
  • Operational credibility: serves high-stakes customers (OpenAI, Apple, Netflix, Uber) training models at billion-dollar scale with 120K GPU clusters weekly; Semafor reports OpenAI chose Ray over half a dozen competing systems
  • Experienced leadership transition: Keerti Melkote's appointment (July 2024) added seasoned operator with proven track record ($3B Aruba exit) while founder Nishihara retained product focus, signaling maturing commercial discipline

Watch for

  • CVE-2023-48022 remains unpatched and actively exploited in the wild (ShadowRay campaign); Anyscale's framing as 'misconfiguration' rather than design flaw has not resolved widespread compromise of public Ray clusters, creating reputational and security perception risk
  • Pricing volatility and incomplete solution perception: industry commentary notes Anyscale's usage-based compute model produces unpredictable bills and that modern AI apps require REST APIs, web frontends, and CI/CD, leaving users to stitch together multiple vendors
  • Flat valuation and no new funding since August 2022 ($1B, $259.6M total capital); no announced Series D; could signal slowing deal momentum or deliberate profitable-growth strategy, but venture-backed infrastructure peers typically seek later rounds

Recent moves

Key Information

Industry
GPU Cloud / Infra
Founded
2019
Employees
1118
Headquarters
San Francisco
Country
United States

Frequently Asked Questions

What is Anyscale?

Anyscale is a managed platform for scaling AI workloads—data pipelines, model training, and inference—built on Ray. Founded in 2019 by UC Berkeley researchers, it serves companies like OpenAI, Apple, and Netflix. The platform handles distributed computing across AWS, GCP, and Azure at billion-dollar scale.

What is Ray?

Ray is an open-source distributed computing framework created by UC Berkeley researchers. Anyscale commercializes and manages Ray as a platform, simplifying how teams scale data pipelines, train models, and run inference. Ray has 237M+ downloads and powers 120K GPU clusters weekly across major AI companies.

What is Anyscale used for?

Anyscale serves LLM training teams, data-intensive ML teams at tech giants, and organizations building multimodal AI. OpenAI, Apple, Netflix, Uber, and others use it for distributed model training, fine-tuning, batch embeddings, and inference. It orchestrates GPU resources across multiple clouds at scale for high-stakes AI applications.

What are Anyscale's main strengths and weaknesses?

Anyscale competes with cloud hyperscalers, specialized platforms, and open-source frameworks. Its strength: Ray's de facto standard status with 237M downloads and PyTorch Foundation backing. It serves high-stakes customers like OpenAI and Apple. Challenges: unpredictable usage-based pricing and incomplete solution—users need REST APIs, frontends, and CI/CD tools separately.

Is Anyscale secure?

CVE-2023-48022 is an unauthenticated remote code execution vulnerability in Ray affecting publicly exposed clusters. Anyscale frames it as a misconfiguration, not design flaw—open-source Ray lacks default authentication. Anyscale-hosted clusters are unaffected, but the vulnerability remains unpatched and actively exploited in the ShadowRay cryptomining botnet campaign.

Who founded Anyscale?

Anyscale was founded in 2019 by UC Berkeley researchers Ion Stoica, Robert Nishihara, Philipp Moritz, and Michael I. Jordan. In July 2024, co-founder Robert Nishihara stepped down as CEO; Keerti Melkote, who founded Aruba Networks and sold it to HPE for $3B, took over. Nishihara refocused on product development.

How Anyscale compares

Direct head-to-head against 3 competitors. Picked by 7wData.

This company

Anyscale

Positioning
Anyscale provides a managed platform for scaling AI workloads—data pipelines, model training, and inference—built on Ray, an open-source distributed computing framework.
Customer segments
Large language model and foundation model training teams at AI labs and enterprises
Strengths
Ray commands de facto standard status in distributed AI compute with 237M+ downloads and 120K GPU clusters active weekly; PyTorch Foundation backing (October 2025) signals institutional entrenchment
Watch for
CVE-2023-48022 remains unpatched and actively exploited in the wild (ShadowRay campaign); Anyscale's framing as 'misconfiguration' rather than design flaw has not resolved widespread compromise of public Ray clusters, creating reputational and security perception risk
Recent moves
PyTorch Foundation Welcomes Ray to Deliver a Unified Open Source AI Compute Stack

Databricks

Positioning
Unified lakehouse platform for enterprise data and ML teams needing data pipelines, model training, and governance in one contract.
Customer segments
Fortune 500 data engineering and ML teams seeking integrated compute, storage, and model governance under one vendor contract.
Strengths
Mosaic AI Model Training delivers custom inference kernels achieving up to 1.5x faster throughput than vLLM-v1, integrated with lakehouse data.
Watch for
Platform fees stack on top of cloud compute. Community forums document consistent cost overruns on Spark-heavy ML workloads.
Recent moves
$4B Series L at $134B valuation closed December 2025, citing enterprise AI expansion and a possible IPO timeline.

Modal

Positioning
Serverless GPU platform for AI inference and batch workloads, targeting developer teams who want GPU access without cluster management.
Customer segments
AI-native startups and ML engineering teams building inference pipelines and batch jobs. Named customers include Mistral, Harvey AI, and Suno.
Strengths
Scales from 0 to 1,000 GPUs in minutes by pooling capacity across hundreds of data centers, with pay-per-second billing.
Watch for
No native Ray compatibility. Teams needing large-scale distributed training report Modal's single-function execution model creates an architectural ceiling.
Recent moves
$355M Series C at $4.65B valuation, May 2026, quadrupling the Series B valuation. Round led by General Catalyst and Redpoint.

NVIDIA Run:ai

Positioning
Kubernetes-native GPU scheduling platform for enterprise AI infrastructure, now NVIDIA-owned, serving on-premises and hybrid cluster operators.
Customer segments
Enterprise IT and MLOps teams managing multi-tenant GPU clusters in on-premises, hybrid, or regulated cloud environments.
Strengths
Fractional GPU allocation and gang scheduling for Kubernetes clusters, reducing idle GPU time in multi-tenant enterprise deployments.
Watch for
Post-acquisition roadmap prioritizes NVIDIA hardware. AMD ROCm and Intel oneAPI users face documented slower support compared to CUDA ecosystem.
Recent moves
NVIDIA completed $700M acquisition of Run:ai, December 2024. KAI Scheduler open-sourced under Apache 2.0 license, April 2025.

Sources

  1. techcrunch.com — Founders, Series A $20.6M from a16z, December 2019 launch details
  2. www.semafor.com — Keerti Melkote CEO appointment (July 2024), Robert Nishihara transition to product, OpenAI/Apple/Netflix/Uber customer names, Ray adoption metrics
  3. www.anyscale.com — Product capabilities: multimodal data curation, distributed training, batch embeddings, post-training; customer logos
  4. pytorch.org — Ray joining PyTorch Foundation (October 22, 2025), Ray's 237M downloads, integration with PyTorch and vLLM
  5. www.anyscale.com — Ray Summit 2025 announcements: Lineage Tracking, Anyscale Runtime (6x cost reduction), Anyscale on Azure, cluster controller, global resource scheduler
  6. getlatka.com — 2023 ARR of $111.9M, $1B valuation as of 2021 Series C
  7. www.anyscale.com — Named customer logos: Coinbase, Character.ai, TwelveLabs, Runway, Recursion, Attentive, Notion, Physical Intelligence, Riot Games, TripAdvisor, Grab, Torc Robotics, Reflection, Canva, Bedrock
  8. www.anyscale.com — Anyscale's response to CVE-2023-48022, framing as misconfiguration issue, no authentication-enabled default in open-source Ray