A serverless GPU platform for deploying and scaling custom AI workloads. You bring a Python app or Docker container, and Cerebrium runs it on GPUs from B200 and H200 down to A10, L4, and T4, billed per second with automatic scaling and low cold starts. It supports real-time endpoints, WebSockets, async tasks, and batch jobs, and exposes an OpenAI-compatible endpoint so existing SDK code can point at your deployment without changes. The company reports SOC 2, GDPR, and ISO 27001 compliance and lists Tavus, Deepgram, Vapi, and ResembleAI as customers. Pricing starts with a free Hobby tier (3 seats, 3 apps, up to 5 concurrent GPUs) plus per-second compute, Standard at $100/month plus compute, and custom Enterprise. GPU rates are published per second (e.g., A10 $0.000306/s, H100 $0.000944/s). Founded in 2021 (Cape Town, now New York), YC-backed, with an $8.5M seed round in 2025 led by Gradient. Product Hunt shows 5.0/5 from 4 reviews, and independent reviews note that third-party customer feedback is still limited.
Pros & Cons:
Pros: Bring-your-own Python/Docker avoids locking you into a fixed model catalog, unlike more opinionated inference platforms; per-second GPU billing plus automatic scaling means you pay for actual compute rather than idle capacity; OpenAI-compatible endpoints let existing SDK code point at your deployment with minimal changes; supports real-time streaming, WebSockets, async tasks, and batch jobs, covering both latency-sensitive and throughput-oriented workloads; published per-second GPU rates (A10 through B200) make cost modeling more transparent than “contact us” pricing; SOC 2, GDPR, and ISO 27001 compliance plus named customers (Tavus, Deepgram, Vapi) give it more credibility than typical early-stage infra startups.
Cons: Pure infrastructure—requires engineering work (Python app or Dockerfile, deployment config), so it is not usable by non-technical teams; total cost depends on GPU type, CPU/memory, storage, cold-start frequency, and runtime, and Cerebrium’s own docs include a dedicated page on estimating costs, which signals that it is not a single headline number; independent third-party validation is thin—Product Hunt has only 4 reviews, and independent reviews explicitly note limited public customer feedback; cold-start performance varies by model size, dependency image, GPU type, region, and availability, so the “~2s” claim should be verified with your own workload; the free Hobby tier caps concurrent GPUs at 5, so production use generally requires Standard at $100/month plus compute.
**Who is it for?**:AI engineers deploying custom models or Python apps who do not want to manage Kubernetes or GPU clusters; teams building real-time voice, video, or agent products that need low-latency streaming and WebSockets; startups that want per-second GPU billing instead of reserved instances; organizations that need SOC 2/GDPR/ISO compliance while running custom inference.