The short answer: a GPU instance that is stopped or idle is not always a free instance. Depending on the provider, a stopped instance may stop the GPU-hour meter while storage, reserved addresses or the reserved slot itself keep billing. Comparing hourly rates without checking what continues after you press stop can change the real monthly figure. This guide sets out the questions to ask, with hypothetical numbers only; it does not describe any tracked provider's terms or prices.
Three states that are often confused
Providers use different words, so it helps to separate the states by what is actually held for you:
- Running but idle: the GPU is allocated and the hourly meter runs, even if utilisation is near zero. This is the most common source of waste, for example a notebook left open overnight.
- Stopped (compute released): the GPU is returned to the pool and compute billing pauses, but disk volumes, snapshots or images are usually retained and may be charged separately.
- Stopped (slot retained): some plans keep the same GPU reserved for you so that it is available when you return. Whether that reservation is billed at the full hourly rate, a reduced rate or not at all is a term to confirm in writing.
None of these is better or worse in itself. The point is that the label "stopped" does not tell you the billing outcome.
What can keep billing while the GPU is off
- Persistent storage. Block or network volumes are commonly billed per GB per month, often independent of whether an instance is attached. Large model checkpoints and datasets make this meaningful.
- Snapshots and saved images. Some providers bill these separately from the live volume.
- Static or reserved IP addresses. A fixed address may carry a small recurring charge while unattached.
- Minimum commitments. Monthly or reserved plans typically bill for the period regardless of usage, which is covered in the post on reserved versus on-demand pricing.
- Deletion after inactivity. Some terms allow data to be removed after a period of non-payment or inactivity. This is a data-loss risk rather than a charge, but it affects whether stopping is a safe way to save money.
A hypothetical worked example
The figures below are invented for illustration and are not any provider's rates. Suppose a team needs a GPU for 8 hours on each of 20 working days, and keeps a 500 GB volume for checkpoints.
- Instance rate: a hypothetical ₹200 per GPU-hour.
- Storage: a hypothetical ₹5 per GB per month.
Scenario A, stop after each session with compute released: compute is 8 x 20 x 200 = ₹32,000. Storage is 500 x 5 = ₹2,500. Total is ₹34,500 before GST.
Scenario B, leave the instance running between sessions: the instance runs for 24 hours a day over the same 20 days, so compute is 24 x 20 x 200 = ₹96,000, plus the same storage. Total is ₹98,500 before GST.
Scenario C, stopped with the slot retained at a hypothetical 25 percent rate for the 16 idle hours: the idle charge is 16 x 20 x 50 = ₹16,000, giving 32,000 + 16,000 + 2,500 = ₹50,500 before GST.
The same workload spans a wide range depending only on stop behaviour. Whether the slot-retained rate in Scenario C is worth paying depends on how costly a wait for capacity or a rebuild of the environment would be, which is a separate cost discussed in the post on setup time as a cost.
GST and invoicing context
Storage, idle reservation and compute charges generally appear as separate line items on an invoice. For a GST-registered buyer, it is worth confirming that every line, not only the GPU-hour line, appears on the tax invoice so that input credit can be claimed consistently. The GST invoicing guide covers what to check.
DPDP angle on retained data
If stopped instances retain volumes or snapshots containing personal data, that data is still being stored. Under the DPDP framework it is worth knowing where retained volumes sit, how long they persist and how deletion is confirmed. The post on DPDP data residency and region selection expands on this.
Frequently asked questions
Does stopping an instance always stop billing? No. It usually pauses the GPU-hour meter, but storage and other retained resources may continue to bill. Confirm the terms of the specific plan.
Where can I model this? The cost calculator takes an hourly rate and usage hours, and the compare tool shows live INR listings, so the idle and storage lines can be added alongside the rate.
All prices in the worked example are hypothetical. Check current listings and each provider's terms before deciding.