Compute overview

Infrastructure designed to maximize price-performance for your ML workloads

Foundry offers two core mechanisms for provisioning compute:

1. Short-term reserved instances: Reservations give you guaranteed access to interconnected nodes. Reserved instances are typically useful for large, preplanned workloads like training that need to run uninterrupted. You can reserve instances in advance for as little as 3 hours and for up to 2 weeks. Reserved instance pricing is adjusted gradually based on overall supply and demand.

2. Spot instances: Spot instances provide a more dynamic way to access compute. To use spot instances, you submit a limit-price order that specifies the maximum price per GPU-hour you are willing to pay. Whenever the spot price is lower than your limit price, your spot instances spin up. See Spot auction mechanics to learn more.

Provisioning compute

Making reservations

Reserve interconnected nodes for as little as 3 hours for preplanned workloads.

Creating spot orders

Burst capacity by placing limit-price orders for spot instances.

Spot auction mechanics

Understand how the Foundry spot auction works and best practices for limit-price bids.