Cerebrium

Cerebrium is a serverless infrastructure platform for real-time AI workloads. Teams use it to deploy large language models, voice agents, image and video models and other AI applications without managing their own GPU clusters. It focuses on low latency from the first request, with fast cold starts and automatic scaling as demand rises and falls.

Performance

Cerebrium uses memory and GPU snapshotting so containers can restore quickly instead of loading from scratch. The site publishes cold-start comparisons. For a vLLM Qwen deployment, it reports 3.8 seconds with snapshots, against tens of seconds or more for other setups, and it shows similar results for Stable Diffusion XL. The platform handles sudden bursts and scale-outs automatically. It lists capacity of more than 2,500 GPUs across several regions, including US East, two European regions and South Asia.

How it works

Developers deploy their own code through a command-line tool. One example on the site runs a training script on eight H100 GPUs with a single command, then streams progress logs back to the terminal. There are no reservations or long-term commitments: you choose the hardware for each app and Cerebrium provisions it on demand. Observability features help track how deployments behave in production.

Pricing

Cerebrium charges for actual compute time, measured per second. Listed GPU rates range from $0.000164 a second for a T4 to $0.00167 a second for a B200, with H100 at $0.000944 and A100 80GB at $0.000583. CPU, memory and storage are billed separately, and the first 100GB of storage is free. The Hobby plan is free plus compute and includes three seats, three apps and five concurrent GPUs. The Standard plan costs $100 a month plus compute.

Things to consider

Per-second billing suits spiky or unpredictable workloads, but steady high-volume traffic should be costed carefully against reserved alternatives. Cerebrium is built for developers comfortable with containers and the command line. Teams with compliance or data location requirements should check which regions and plans meet their needs.

Typical workloads on the site include LLM inference, real-time voice agents, image and video generation and model training. The same deployment workflow applies to each, so a team can run several kinds of model on one platform rather than stitching together separate hosting services for each project.