The short answer: an uptime percentage on a GPU cloud is a promise about a month, while a training run or serving workload experiences individual outages, and the remedy is usually a service credit rather than cash or lost work recovered. Two providers with different SLA terms can therefore cost different amounts per useful GPU-hour even at the same listed rate. This guide shows how to translate an uptime figure into hours, what to read in the credit clause, and how to compare terms using hypothetical numbers only.
What an uptime percentage means in hours
The arithmetic is simple. A 30-day month has 720 hours, or 43,200 minutes. A 99.9 percent monthly commitment allows about 43.2 minutes of downtime. At 99.5 percent the allowance is 216 minutes (3.6 hours), and at 99 percent it is 432 minutes (7.2 hours). Whether a given outage counts also depends on the definition in the agreement, so the percentage alone does not settle the comparison.
Five clauses that decide what the SLA is worth
- What is measured. Some agreements measure availability of the control plane or the host, not of the specific GPU instance you rented. A GPU fault on a running node may or may not be counted.
- Exclusions. Scheduled maintenance, network issues outside the provider's boundary and customer-caused faults are commonly carved out.
- Form of remedy. Credits are typically applied to future invoices, may be capped at a share of the monthly charge, and may require a claim within a set period.
- Applicability by plan. Spot or preemptible capacity, free tiers and promotional credit may sit outside the SLA entirely. See the spot and preemptible pricing guide for how interruption risk is priced separately.
- Support response. A response-time target is not a resolution-time target. Check which one the agreement states.
A worked example with hypothetical numbers
Imagine two providers, A and B, both listing INR 100 per GPU-hour. These are not real prices and do not describe any tracked provider. A team runs a 4-GPU job for 72 hours, a compute bill of INR 28,800 (4 GPUs x 72 hours x INR 100).
Suppose an outage hits at hour 60, and the job's last checkpoint was at hour 54. The 6 hours of work since the checkpoint must be repeated, which is 4 x 6 x INR 100 = INR 2,400 of compute paid for twice, before counting engineer time spent restarting. If the team checkpoints every hour instead, the repeated work falls to 1 hour or less, about INR 400.
Now apply the remedy. Suppose Provider A's agreement credits 10 percent of that month's charge when availability falls below its commitment, and the team's monthly bill is INR 50,000. The credit is INR 5,000, usable on future invoices. Suppose Provider B's agreement offers a smaller credit, or excludes the failure type that occurred. The credit may exceed the repeated compute, or fall short of it; it does not repay a missed deadline, and it only has value if the team keeps buying from that provider. The listed rate is identical in both cases. The effective cost per useful hour is not.
How to compare SLA terms across providers
- Convert each commitment into minutes per month using the arithmetic above.
- Read the measurement definition and exclusions before the percentage.
- Note the remedy form, cap and claim window.
- Check whether your intended billing model, such as spot or monthly commitment, is covered. The billing granularity guide covers how rounding interacts with restarts.
- Decide the cost of an interruption for your workload, then compare it with the credit on offer. Checkpoint frequency usually moves this number more than the SLA does.
Compliance and invoicing notes
An SLA credit is an adjustment to what you are billed, so ask how it appears on the tax invoice and whether it is issued as a credit note, particularly for GST-registered buyers; the GST invoicing guide lists the invoice details to confirm. Availability terms are also separate from data protection duties. An outage clause says nothing about where data is stored or how incidents are reported, which the DPDP data residency guide addresses.
Hourly rates and billing types for tracked providers are on the compare page. SLA terms are not part of that table, so read each provider's current agreement directly.
FAQ
Does a higher uptime percentage mean a more reliable GPU? Not by itself. The percentage depends on what is measured and excluded. Compare definitions and remedies, not only the headline figure.
Is an SLA credit the same as a refund? Usually not. Credits are commonly applied against future usage, while refunds return money. The agreement states which applies.
This is a general framework, not legal, tax or financial advice. All provider terms and figures in the worked example are hypothetical. Agreements change, so verify the current SLA with each provider.