Spot vs On-Demand GPUs: Price, Capacity and Risk

What is the difference between Spot and on-demand GPUs?

Spot rents interruptible capacity under the provider’s Spot rules. On-demand generally charges for ordinary usage without a long commitment; it still requires successful allocation and sufficient quota. A reservation or contract adds separate financial and capacity terms. Read the actual offer to determine what is guaranteed.

SiliconTicker labels Spot separately from standard listings. Compare Spot offers or compare standard offers, then verify the provider’s allocation and billing conditions.

Interruption risk is part of the price

Microsoft says Azure Spot capacity varies by VM size, region and time, has no availability SLA and can be reclaimed. Notifications are best effort and may arrive up to 30 seconds before eviction. A published tariff alone does not prove that a VM can be allocated. Azure Spot documentation.

Interruption rules are provider-specific. Check the notice mechanism, checkpoint storage and restart behavior for the exact service. The Azure behavior above should not be assumed for every cloud.

Questions to ask when comparing pricing models
Model	Main tradeoff	Confirm before renting
Spot	Potentially lower tariff; interruptions and allocation risk	Eviction rules, restart path, quota and checkpointing
On-demand	Flexible duration; ordinary usage rate	Region, quota, actual availability and billing interval
Reserved / contracted	Commitment can change price and capacity terms	Term, minimum spend, cancellation and explicit delivery obligations

Calculate effective job cost

Effective compute cost = billed GPU plan-hours, including lost work and restarts, × the applicable plan rate. Add checkpoint storage, transfer and operational costs. With changing rates, calculate each billed interval separately.

ILLUSTRATIVE EXAMPLE · CONSTANT RATES

A job taking ten hours on a $4-per-hour standard plan costs $40 in compute. At a $2 Spot rate, it costs $20 if uninterrupted. If lost work, checkpointing and restarts consume another twelve billed hours, it costs $44. The lower tariff no longer gives the cheaper completed job.

For this example the Spot job must finish in fewer than twenty billed hours to beat $40, before extra costs. Replace the assumptions with measured restart overhead and observed billing terms. See GPU rental cost calculations.

Choose the terms your workload can tolerate

Restartable batch jobs: evaluate Spot when tasks are independent or can resume from durable checkpoints.
Long training runs: include lost progress, checkpoint frequency and reallocation time in the budget.
Latency-sensitive inference: confirm a capacity and failover design that meets the service requirement.
Fixed deadlines: price the cost of missed completion alongside the hourly rate.

A discount is useful only when the completed result still meets your requirements. A reserved price should not be described as a capacity guarantee unless the contract explicitly provides one.

Common Spot GPU questions

Does a Spot price mean capacity is available now?

No. Our Azure Spot tape publishes official retail tariffs with region and effective dates. Availability and quota require provider confirmation.

Does setting a maximum Azure Spot price prevent eviction?

No. A price setting can control price-related eviction, but it does not prevent Azure from reclaiming capacity.

Does on-demand guarantee uninterrupted service?

Check the service’s SLA and allocation terms. The pricing label alone does not specify reliability, delivery or capacity commitments.

Primary sources

Technical and billing details were reviewed on 7 October 2026. Provider terms and capacity can change; price tables show their own observation dates.

Compare GPU rental offers →