The short answer: environment setup time is a real line item in GPU cloud cost, and it does not appear on any rate card. Every minute spent installing drivers, matching CUDA and framework versions, pulling container images or downloading model weights is billed at the same hourly rate as productive compute on most instances. This guide shows how to put a rupee figure on that time so two providers can be compared on effective cost, not only on the listed rate.
Where setup time comes from
Setup time is not one thing. It is a set of steps that recur each time an instance is started from scratch, and each step varies by provider and image.
- Base image and drivers. Some providers offer images with a GPU driver, CUDA toolkit and common frameworks preinstalled. Others hand over a bare operating system. The difference can be a few minutes or a longer session of troubleshooting version mismatches.
- Framework and dependency installs. Pulling Python packages and compiling extensions repeats on every fresh instance unless a custom image or persistent volume is used.
- Container image pulls. Large images take time that depends on the provider's registry proximity and network, which you can only learn by measuring.
- Model weight and dataset downloads. Multi-gigabyte downloads add time, and if the source is outside India they can also interact with the data-transfer charges discussed in our egress cost guide.
- Warm-up and health checks. Loading weights into GPU memory and running a first test request is work you pay for before any useful output.
The formula
Effective cost of a job equals the hourly rate multiplied by the sum of setup hours and productive hours. The setup share is setup hours divided by total billed hours. Two providers with identical productive speed can therefore differ in effective cost purely because of setup overhead.
The figures below are hypothetical round numbers for illustration. They are not any provider's price and not a measurement of any image. Substitute live rates from the compare table and your own timed setup runs.
- Scenario A. A hypothetical rate of ₹100/hour, 0.25 hours of setup, 4 hours of productive work. Billed hours: 4.25. Effective cost: ₹425, of which setup is about 6 percent.
- Scenario B. A hypothetical rate of ₹90/hour, 1.5 hours of setup, same 4 hours of work. Billed hours: 5.5. Effective cost: ₹495, of which setup is about 27 percent.
In this illustration the lower listed rate produces the higher effective cost. Change the inputs and the ranking can reverse, which is the point: the rate card alone does not settle it. The same logic applies when an instance is started and stopped many times in a week, because setup repeats each time unless the environment persists.
How billing granularity interacts with setup
Setup time is billed at whatever granularity the provider uses. Under per-hour billing, a short failed start may still be charged as a full hour; under per-second or per-minute billing the same attempt costs far less. See our per-second versus per-hour billing comparison for how billing increments work. Whether the clock runs while an instance is provisioning or only once it is reachable is a provider-specific term worth checking in the documentation.
Reducing setup cost without changing provider
- Build a custom image or container once, so installs do not repeat.
- Keep weights and datasets on a persistent volume, then compare the storage charge per month against the hourly cost of re-downloading. Persistent storage is usually billed even when no GPU is attached, so this is a trade-off, not a free fix.
- Do dependency installation and data preparation on a cheaper CPU instance where the workflow allows it.
- Time a full cold start on each shortlisted provider before committing a large job, and record it.
Invoicing and compliance notes
Setup hours are billable usage and appear on the tax invoice like any other compute hours. For GST treatment across Indian and foreign providers, see our GST invoicing overview. If setup involves downloading datasets that contain personal data, the DPDP considerations in our data residency article apply to where that data is stored and processed during setup as well.
FAQ
Is setup time always billed?
It depends on the provider's billing terms. Many bill from the moment an instance is running, which includes time you spend installing software. Check each provider's documentation for when the meter starts and stops.
Does a preinstalled image always reduce cost?
Not automatically. It reduces setup time only if its driver and framework versions match your workload. A mismatch can cost more time than a clean install, so measure rather than assume.