Cost decisions

Reserved vs on-demand GPUs: what are you committing to?

Commit to a reservation only after identifying exactly what it buys: a discount, a specific allocation, or both. On-demand is more flexible for uncertain usage. Reserved capacity can suit a predictable window, but unused time and cancellation rules can outweigh a lower advertised rate.

Read the product definition

AWS distinguishes Capacity Reservations from savings and reserved-instance products, with different capacity benefits. Vast.ai describes reservations as prepaid discounts on an instance. The word reserved therefore does not carry one universal meaning.

Ask for the GPU variant, count, region, start time, end time and replacement terms. Separately ask how billing changes if you use less than expected. These questions should have written answers before the commitment.

References: EC2 capacity reservations · Vast.ai instance types

Price unused hours explicitly

Suppose an illustrative reservation costs $1.50 per hour for 100 hours, while an on-demand offer costs $2 per used hour. The reservation costs $150. It breaks even at 75 used hours; at 50 used hours, on-demand costs $100.

This calculation assumes equivalent performance and ignores other charges. Add storage, setup and any minimum billing periods before making a purchase. A nominal discount can still increase the total bill.

Price unused hours explicitly
Useful hoursReserved exampleOn-demand example
50$150$100
75$150$150
100$150$200

Separate deadline risk from price risk

A scheduled training campaign may value an allocation window even when it offers little discount. A research team with uncertain experiments may value the ability to stop. State which problem the commitment solves.

Confirm whether capacity can be replaced after a failure and whether the replacement preserves the required topology. A reservation for several independent devices may not satisfy a tightly coupled multi-node job.

Commit in stages when the usage is new

Start with a pilot that measures actual runtime and utilization. Reserve only the baseline you can justify, then use flexible capacity for uncertain work. Revisit the split when the experiment schedule becomes more predictable.

Assign someone to monitor the commitment end date and remaining obligations. Include retained storage in the closeout plan so finishing the reservation also finishes the intended spending.

Questions readers ask

Does reserved always mean guaranteed GPU capacity?

No. The product’s capacity terms determine that; some commitments primarily change billing.

When does a discount become a bad purchase?

When unused committed time, cancellation costs or mismatched capacity exceed the saving on useful work.

Sources and editorial notes

Primary documentation reviewed October 9, 2026. Provider descriptions reflect published documentation. Recommendations are our analysis; worked figures are illustrative unless explicitly stated otherwise. No independent provider benchmark or uptime ranking is claimed.

Read the editorial policy. NeuraMesh is a prototype, not live verified GPU inventory.