Runcrate is a GPU cloud with a thesis: the customer is increasingly an AI agent. H100/A100 instances and per-token inference sit behind an OpenAI-compatible API, with MCP support so an agent can provision the compute it needs mid-task. For humans it reads as a tidy mid-size GPU provider with visible hourly pricing; the agent angle is the bet on where the market goes.
What it costs
Per-hour GPU rates and per-token inference listed openly (we saw rates from $0.27–0.74/hr class entries) — compare against the brokers before committing volume, as always.
What we can’t tell you
The boring clouds-are-hard stuff: availability when you actually want an H100 (quoted prices mean nothing at zero stock), egress costs, and what “agent provisions compute” means for your bill when the agent is enthusiastic. Set spend caps before you hand an agent the keys — that sentence will be obvious in two years and is advice today.
The details
| What it is | GPU cloud + per-token inference, OpenAI-compatible, MCP for agents |
| Price | Per-hour GPUs, per-token inference, listed openly |
| Best for | Developers (and their agents) needing compute without AWS ceremony |
| Think twice if | You skip spend caps — agent-provisioned compute needs them on day one |
