ANYSCALE

Ion Stoica is one of the most influential computer scientists in distributed systems — he co-invented Apache Spark while at Berkeley and co-founded Databricks. He then built Ray: an open-source framework that makes it easy to distribute Python code across many computers, which turned out to be exactly what the machine learning community needed to train large models.

Ray became the default distributed computing framework for AI companies at scale — OpenAI trains on Ray, Uber uses Ray, Spotify uses Ray. Anyscale was founded to commercialize Ray: offer a fully managed Ray platform so companies could run distributed AI without managing infrastructure.

Anyscale charges for cloud compute used to run distributed AI and Python workloads on its managed Ray platform. It offers a "Ray on Anyscale" service that abstracts away the complexity of managing Ray clusters — companies pay for the compute consumed plus a platform fee.

It also offers training and inference infrastructure specifically for LLMs.

Anyscale Platform (managed Ray clusters), RayTurbo (optimized LLM training), Anyscale Endpoints (LLM inference API), Ray open-source framework

Every major AI company's adoption of Ray was a distribution case study. OpenAI's public acknowledgment that it trains on Ray drove enormous inbound interest.

Anyscale launched RayTurbo (a faster version of Ray for LLM training) to differentiate from self-managed Ray. Partnerships with major cloud providers for marketplace listings expanded distribution.

Ray is open-source and free. Anyone can run Ray themselves on AWS, GCP, or Azure without paying Anyscale anything.

The challenge is convincing companies that the managed service is worth the cost — Anyscale has to demonstrate that its managed platform saves enough engineering time and avoids enough operational headaches to justify the premium. As AI infrastructure spending exploded, this pitch became easier.

Andreessen Horowitz, NEA, Intel Capital, Foundation Capital, Addition (Lee Fixel)