Teams starting AI work assume the question is cost. Often the first real obstacle is availability: the instance type you planned around is not available in the region you need, and the wait is measured in weeks.
Separate training from inference before deciding anything. Training is bursty, tolerant of interruption, and largely indifferent to where it happens. Inference is continuous, latency-sensitive, and frequently subject to data residency expectations. Optimising both with one procurement decision produces a compromise that suits neither.
Training is the better candidate for spot or preemptible capacity, which is dramatically cheaper and can be reclaimed at any moment. That is workable if your training loop checkpoints frequently and resumes cleanly — and building that in from the start is far easier than retrofitting it after a long run is lost.
For inference, the honest first question is whether you need a GPU at all. Smaller models, quantisation and modern CPU inference cover a wider range of production workloads than the discourse suggests, and a CPU deployment that meets your latency target removes an entire supply problem.
Where a hosted model API meets the requirement, it usually beats self-hosting on total cost until volume is substantial, because you pay only for what you use and none of the idle time. The reasons to self-host are typically data residency, latency, or a model you have adapted — not price at low volume.
If data must stay in India, verify the specific region and instance type rather than the provider's country-level claim, and get it in writing. This is also where Indian providers and colocation become worth evaluating, particularly for steady inference workloads where the hyperscaler premium is hardest to justify.
