From the archive
Fly.io launches GPU Machines for hosted AI workloads
A100 accelerators become available alongside Fly Machines, with on-demand pricing and regional deployment.
Fly.io opened access to GPU Machines, adding NVIDIA A100 accelerators to its application-hosting service. The launch included 40GB and 80GB configurations, with hardware coming online in Chicago, Northern Virginia, Amsterdam, San Jose, and Sydney.
Models beside an application
The service uses virtual machines attached to graphics processors, with a deployment experience similar to other Fly Machines. Fly described inference with existing models as a main use case, alongside custom workloads and some smaller adaptation tasks. It was not positioned as a service for training very large models from scratch.
Pay for the accelerator in use
The announced on-demand GPU rates were $2.50 per hour for the 40GB A100 and $3.50 per hour for the 80GB version. Applications also choose the CPU, memory, and storage they need. Reserved machines and dedicated hosts were offered as alternatives to on-demand use.
The Fly.io profile covers the company's application-hosting service and available deployment options.