Anyscale

Picture this job: you've got a few million documents sitting as Parquet files in S3, and you need embeddings for all of them by tomorrow so your retrieval system can ship. On Anyscale, the example on the homepage handles that in about 25 lines of Python. You read the files with ray.data.read_parquet, wrap a SentenceTransformer model like bge-large-en-v1.5 in a class, and call map_batches with concurrency set to 16, one GPU per worker and a batch size of 512. Then you write the results back to your warehouse bucket. You don't write any queueing logic, retry loops or cluster scripts. Ray spreads the work across the GPUs and Anyscale runs the cluster.

That pattern is the core of the product. Anyscale is the managed platform from the team behind Ray, which it calls the world's most widely adopted AI compute engine. It's aimed squarely at teams building and scaling foundation models, and it groups its pitch into four workloads: multimodal data curation across video, images, text and audio, distributed training across GPU clusters with elastic scaling, batch embedding generation, and post-training with Ray-native frameworks like SkyRL and veRL. The code samples are real Ray code, not a proprietary DSL. If your team already writes Ray, the move to Anyscale is mostly about who babysits the cluster. A 64-GPU TorchTrainer run on a Llama 3.1 70B config is pitched as the same few lines you'd write locally.

The pricing model is usage-based. You start with $100 in credit, pay only for compute you use, and can move to committed contracts for volume discounts. The bigger decision is deployment. The Hosted option gets you running with zero setup, but it comes with limited regions and business-hours support capped at 5 case submissions. Bring Your Own Cloud runs inside your VPC or on-prem, works on VMs or Kubernetes, lets you use GPU reservations you've already paid for, and adds 24x7 enterprise SLAs. That's a sensible split, though it means the version most serious teams will want sits behind a sales conversation.