OpenRelay

OpenRelay provides hosted model inference and dedicated GPU compute on a network that routes work across many providers. The vendor describes it as an inference delivery network that sends each request to the best provider based on latency, cost, availability and model performance, and that fails over automatically when a node goes down. There are three ways to use it. Serverless inference puts hosted models behind one endpoint that scales with traffic and bills for what is generated. Dedicated compute launches NVIDIA GPUs with SSH access and a persistent volume, metered per minute, and the vendor lists AMD and Google TPU options too. Batch jobs queue heavy workloads at a lower price than interactive traffic. An orl command line tool wraps the full API for launching and scripting VMs, and a single command serves the platform as MCP tools so that Claude or another MCP client can deploy machines and read usage, with an optional read only mode. The company describes its network as built on owned GPU infrastructure plus distributed capacity, and it is backed by Y Combinator. It is aimed at teams that want to run models or training jobs without a hyperscaler bill. Pricing is usage based. Inference is billed per token, and starting requires adding a card and making a deposit. A pricing calculator shows example GPU rates of $0.18 per hour for an RTX 3090, $0.35 per hour for an RTX 4090 and $2.60 per hour for an H100 SXM. An uptime SLA is available on dedicated capacity.