GPU Spot vs. Reserved Instance Savings Guard

Benchmark your AI training budget by comparing spot capacity against reserved and on-demand rates. Input your expected GPU hours and spot discount to see the effective blended rate after reruns plus the savings or shortfall against other purchasing models.

Total GPU runtime you expect to consume this month.
Percent off of on-demand list pricing when using spot capacity.
Optional. Defaults to $3.00 per GPU hour when blank.
Optional. Reserved instances assumed to be 25% cheaper when left blank.
Optional. Enter extra runtime you expect to repeat because spot jobs are interrupted.

Pricing and discounts vary by cloud provider and region. Validate numbers with your vendor agreements before budgeting.

Examples

  • 300 hours, 45% spot discount, $3.00 on-demand, 25% reserved discount, 12 rerun hours ⇒ Effective spot rate including reruns: $1.65/hr • Monthly savings vs. on-demand: $385.20 • Monthly savings vs. reserved: $160.20
  • 500 hours, 35% spot discount, defaults for on-demand, reserved, and reruns ⇒ Effective spot rate including reruns: $1.95/hr • Monthly savings vs. on-demand: $525.00 • Monthly savings vs. reserved: $150.00

FAQ

How should I estimate rerun hours?

Track historical interruption frequency for similar workloads. Multiply the typical restart time by expected interruptions per month to populate the field.

Can I compare multiple GPU types?

Run the calculator separately for each instance family. Update the on-demand rate to reflect current pricing for A100, H100, or L4 configurations.

Does the tool include sustained-use discounts?

No. Apply those to the on-demand rate before entering it to avoid double-counting savings.

What if the result shows negative savings vs. reserved?

Negative values indicate spot interruptions erode the advantage relative to a committed reservation. Consider mixing strategies or raising the rerun buffer.

Additional Information

  • Rerun hours capture job restarts triggered by spot evictions and increase both billed hours and blended rate.
  • Reserved savings assume a one-year commitment with steady utilization and no interruption overhead.
  • Networking, storage, data egress, and orchestration costs are excluded so you can isolate compute economics.