We rent NVIDIA accelerators by the second to people who would rather be running experiments than negotiating contracts.
Getting a GPU used to mean one of two bad options: buy hardware that is obsolete in eighteen months, or sign an annual commit with a hyperscaler whose quota system says no anyway. Neither works for the way research and product development actually happen — in bursts, with long quiet gaps.
So the model here is deliberately boring. Prepaid credit, per-second billing, no contract, no sales call, no minimum. Deploy what you need, destroy it when the run finishes.
Hourly rounding is a quiet tax on bursty workloads. If your job takes 71 minutes, being charged for two hours is a 69% markup on the tail. We bill per second and publish every rate on the pricing page — there is no rate card behind a login, and no discount you only get by asking.
The same applies to egress. Charging for data leaving a platform is a retention mechanism dressed as a cost. We give the first 10 TB free each month and charge a cent per gigabyte after, which is roughly what it costs us.
We are not a hyperscaler. There is no managed Kubernetes, no proprietary training service, no model marketplace. You get a machine with a GPU in it, root access, and a network connection. If you need a platform rather than infrastructure, we are probably the wrong choice, and we would rather say that up front.
We also do not hold HIPAA or PCI attestations. The security page sets out exactly where we stand on compliance instead of implying more than is true.