Reserved vs on-demand GPUs: when does prepaying pay off?

On-demand GPUs bill every hour they run; a reserved node is paid once for the whole term at a lower, locked rate. Reserving wins once the node is busy more than a certain share of the time. Put in the on-demand rate you're quoted to see where that line is.

Break-even calculator

Costs are for one 8-GPU node. Avian rates are prepaid and locked for the term; prices as of October 2026.

Use the on-demand rate you've been quoted for the same GPU. For an 8-GPU on-demand node to beat a reservation you also need one to be available when you need it.

Avian reserved, per node per month
$0
Avian reserved$0
On-demand at your usage$0

Reserve this node

How prepaid reserved pricing works

You choose a GPU and a term, pay once for the whole term, and the node is yours: all 8 GPUs, single-tenant, at a per-GPU-hour rate that can't change until the term ends. There's no meter, no monthly invoice and no surprise bill.

When on-demand is the better deal

  • Short experiments, evaluations and one-off jobs measured in hours or days.
  • Spiky workloads that sit idle most of the time, below the break-even line above.
  • Work that doesn't need a full 8-GPU node with NVLink.

When reserving wins

  • Steady inference traffic or long training runs that keep the node busy.
  • Teams that need a guaranteed 8-GPU node every day, not whatever capacity is free that hour.
  • Budgets that need one predictable number: the whole term is paid and fixed upfront, and 12-month and longer terms can be invoiced (NET-30 or PO).

For workloads that don't need a whole node, Avian's per-token inference API runs DeepSeek V4.1, MiniMax M3, GLM-5.3, Kimi K3, MiMo-V2.6 with no reservation, from $0.1275/M tokens.

NVIDIA B300 pricingEvery term, per GPU-hour and per node. NVIDIA B200 pricingEvery term, per GPU-hour and per node. B300 vs B200Which node to reserve for your workload.