Cerebrium

Short Intro: Cerebrium is a serverless GPU infrastructure platform founded in 2021 by Michael Louis and Jono Irwin, who previously co-founded OneCart, a grocery delivery platform acquired by Walmart, where they built the production AI systems that showed them how broken ML deployment tooling was. Coming through Y Combinator's W22 batch and backed by an $8.5M seed led by Gradient Ventures in July 2025, the platform targets the latency floor that real-time voice agents and multimodal AI pipelines require. Deepgram, Tavus, and Vapi run workloads on Cerebrium, and the company is headquartered in New York City with roots in Cape Town, South Africa.

Key Capabilities:

• Serverless GPU instances across 12+ GPU types from T4 through H100

• Average cold starts of 2-4 seconds with 35ms added latency per request

• Pay-per-inference billing with no minimum commitments or idle GPU charges

• Autoscaling to 10,000+ requests per minute with minimal engineering overhead

• Real-time voice application support targeting sub-500ms response times

• WebSocket and streaming endpoints for bidirectional real-time communication

• Multimodal inference pipelines combining LLM, vision, and audio workloads

• Multi-region deployments with data residency controls

• Custom Dockerfiles and runtimes for flexible environment configuration

• Adaptive batching and concurrency for GPU utilization optimization

• Distributed storage and secure secrets management built in

• LLM fine-tuning and large-scale batch job support

• One-line deployment for both custom models and open-source LLMs

See Cerebrium pricing details →