The short answer: an hourly rate only matters if the GPU is available when you need it. Capacity limits, quota approvals and waitlists can add days between choosing a provider and starting a job, and that waiting time has a cost that never appears on a rate card. This guide sets out how to compare availability alongside price, using hypothetical numbers only. It does not describe any tracked provider's capacity, quota rules or prices.
Four ways availability shows up
- Account or quota limits. Some providers cap the number or type of GPUs a new account can launch until it is reviewed. Approval may involve business verification and GST details, so lead time can vary from minutes to days.
- Stock-outs on a given GPU model. A listed GPU is not always launchable in a given region at a given moment. Popular memory tiers and multi-GPU nodes tend to be the most constrained, but this varies by provider and week.
- Waitlists and reservations. Larger clusters are sometimes allocated on request rather than on demand, with a booking window and a minimum term.
- Preemption and capacity reclaim. Lower-priced interruptible capacity can be reclaimed. The spot and preemptible pricing guide covers that trade-off.
None of these is a defect. They are terms to read before deciding.
A way to price the wait
Waiting is a cost only if something is idle because of it. A simple framing is to compare two options on total cost to completion, not on rate:
- Cost of the compute: hourly rate multiplied by GPU-hours needed.
- Cost of the delay: hours of waiting multiplied by what the delay actually costs you. For a team of salaried engineers this is the time they cannot use productively. For a customer-facing launch it may be a missed date instead.
Hypothetical example, for illustration only. Option A lists INR 100 per GPU-hour and needs a two-day quota review. Option B lists INR 120 per GPU-hour and can launch the same day. The job needs 200 GPU-hours.
- Compute on A: 200 x INR 100 = INR 20,000.
- Compute on B: 200 x INR 120 = INR 24,000.
- Rate difference: INR 4,000 in favour of A.
Now suppose three people would otherwise be blocked for two working days, and the loaded cost of their time is a hypothetical INR 1,500 per person per hour. Sixteen working hours of blocked time across three people is 48 person-hours, or INR 72,000. In that scenario the delay costs far more than the rate gap. If instead the team has other work to do while the quota is approved, the delay cost may be close to zero and the lower rate wins. The numbers change the answer, which is why they are worth writing down for your own situation.
What to ask before you commit
- Is the listed GPU launchable now, in which region? A listing and live capacity are different things. Region also matters for data residency, as covered in the DPDP region selection guide.
- What are the default account limits, and what does a limit increase require and how long does it usually take?
- Is capacity guaranteed or best effort? On-demand wording rarely promises that a specific GPU will be free at a specific time.
- What is the minimum term for larger or reserved capacity, and what happens if you need less than planned? The reserved vs on-demand analysis looks at that trade-off.
- How long do onboarding and invoicing set-up take? GST-registered buyers usually need the provider to hold their GSTIN before the first invoice. See the GST invoicing guide.
Planning around availability
Teams can reduce exposure without paying a higher rate. Requesting limit increases before a project starts, rather than on the day, removes most approval delay. Flexible job design helps too: a workload that can run on either of two GPU memory tiers, as discussed in the VRAM sizing guide, is less likely to be blocked by one tier being full.
Frequently asked questions
Does a lower hourly rate mean a cheaper project? Not necessarily. Total cost to completion includes any idle time caused by quota approval, stock-outs or queueing. Whether that time is costly depends on what your team would otherwise be doing.
Where can I model this? The cost calculator takes an hourly rate and usage hours, and the compare tool shows live INR listings. Add the delay cost as a separate line next to the compute cost.
All figures in the worked example are hypothetical. Availability and quota terms change often, so check each provider's current terms and live listings before deciding.