Anyscale

Anyscale has no free plan, but every new account gets a $100 credit. Startups can also apply for up to $20K in credits through the AI Startup Program. After credits run out you pay per minute of compute, billed in Anyscale Credits, with no monthly fixed fee.

Anyscale's pricing page lists an instance containing an NVIDIA A100 at 4.9591 Anyscale Credits per hour. On the dollar toggle that is $4.5916 per hour on Hosted, or $0.9184 per hour as Anyscale's fee on BYOC, where your cloud provider bills the machine separately.

Anyscale's program page lists up to $20K in Anyscale credits, dedicated field engineers and the Anyscale Runtime. Eligibility is generally open to companies at various stages, though funding, technical maturity and current cloud setup may be considered. Anyscale reviews each application, and its sales team follows up within 24 to 48 hours. The page does not state how long the credits last.

Hosted runs on Anyscale's own infrastructure in limited regions with credit card billing and business-hours support with 5 cases. BYOC runs in your AWS, GCP, Azure, OCI, CoreWeave or Kubernetes environment, keeps data in your VPC, and comes with 24x7 enterprise SLAs and unlimited cases.

Yes, on BYOC. The machines run in your own cloud account, so your provider credits pay for them. Anyscale says its startup program credits stack with existing cloud provider credits.

Ray is open source and free. Anyscale adds Developer Central, a Cluster Controller for managed scaling and monitoring, and the Anyscale Runtime, which keeps APIs compatible with open-source Ray. You pay for that managed layer per minute of use.