Fly.io — Fly GPUs

What it does

GPU-backed Fly Machines for model inference and training. A top-level docs section, peer to Fly Machines and Managed Postgres.

Target ICP

• AI engineers deploying open-source model inference (Llama-family, Mistral, etc.).
• Training workloads that need short-duration GPU bursts without a long-term cloud contract.
• Ring 2 overlap with the Sprites pitch — AI builders who need both isolation and accelerated compute.

Inputs / Outputs

• Input: Docker image + GPU spec (type, count).
• Output: Running GPU-backed Machine, billable per second.

How it works

GPU SKUs attached to Machines (specific GPU models not captured from the docs index). Per-second billing consistent with Machines pricing.

Pricing / limits

Metered. Specific $/hour rates not captured.

UX entry points

• Docs: top-level "Fly GPUs" section.
• CLI: presumably fly launch with a --gpu flag or config in fly.toml.

Differentiators (claimed)

• Per-second billing on GPUs — most hyperscalers bill per hour or per minute minimum.
• Same platform as Machines — no separate GPU product line with its own CLI / console.

Sources

• fly.io/docs — Fly GPUs section (referenced in nav)