Past roughly 500 GPU-hours a month, reserved capacity is cheaper and more predictable than the public rate card. That is the point at which it is worth a conversation.
Committed GPU-hours at a discount to the public rate, with guaranteed availability in your chosen region. One, three or twelve month terms.
Contiguous allocations of 8 to 512 GPUs on a non-blocking InfiniBand pod, placed together rather than scattered across the fleet.
Bring your own base image, driver branch or kernel. We validate it once and make it available to your account as a first-class option.
Cross-region private links and VPC peering back into your own cloud account, so data never crosses the public internet.
Net 30 terms, purchase orders, and consolidated monthly invoicing instead of prepaid credit. Multi-currency supported.
A named engineer, a shared channel, and a 1-hour response target on production issues rather than the standard business day.
Reserved pricing is a function of three things: how many GPU-hours you commit to, over what term, and how much flexibility you give us on placement. A twelve-month commit on a flexible region is the cheapest arrangement; a one-month commit pinned to a specific pod is the most expensive.
As a rough guide, expect 15-25% off the public rate at 500 hours a month, and 30-40% at 5,000. We will put actual numbers in writing after one call — there is no multi-stage discovery process.
If your workload is better served by a hyperscaler, or by buying hardware outright, we will say so. Sustained 24/7 single-GPU usage over two years, for instance, is usually cheaper to own than to rent from anyone.