Cerebrium

Cerebrium is a serverless GPU platform for deploying and scaling AI workloads such as voice agents, image and video models and large language models. You point it at your entry point or Dockerfile and it runs your application as is, versioned and reproducible, with no custom SDK, decorators or Kubernetes to manage. Memory and GPU snapshotting restore containers quickly, and autoscaling handles bursts without capacity planning or reservations. It is aimed at developers and ML teams who need low latency inference in production, as well as batch jobs and training runs launched from the CLI. Capacity spans multiple clouds and regions, currently listed as US East, EU West, EU North and Asia Pacific South. Built in observability covers logs, metrics, scaling events and per request visibility, and plans add custom domains, private Slack support and ML engineering services at the top tier. Compute is billed per second: B200 at $0.00167, H200 at $0.001166, H100 at $0.000944, RTX PRO 6000 at $0.000694, A100 80GB at $0.000583, A100 40GB at $0.000555, L40s at $0.000542, A10 at $0.000306, L4 at $0.000222 and T4 at $0.000164, with CPU at $0.00000655 per vCPU second, memory at $0.00000222 per GB second and storage at $0.05 per GB per month after the first 100GB. Hobby is free plus compute, with 3 seats, 3 deployed apps and 5 concurrent GPUs. Standard is $100 per month plus compute, with unlimited apps, 30 concurrent GPUs and custom domains. Enterprise is custom, with volume discounts and unlimited concurrent GPUs. The pricing page is inconsistent on some limits, listing Standard seats as unlimited in one place and 10 in the comparison table.