Reserved vs. on-demand GPU pricing in India: what the actual numbers show

Reserved capacity is supposed to be the cheaper option. Checked against real pricing across the providers we track, that's not always true — here's what actually determines the right billing model for an agent product.

20 July 2026 · compute-economics

The conventional assumption in cloud computing is that committing to reserved or monthly capacity gets you a discount over paying hourly. Checked against the real, verified pricing we track across nine GPU cloud providers, that assumption doesn't hold up cleanly — the two providers in our dataset that only offer monthly billing are priced in the middle-to-expensive end of the market, not the cheap end. If you're deciding how to pay for GPU capacity behind an agent product, the honest answer is: check the actual on-demand alternatives before assuming a commitment saves you money.

What our own H100 data actually shows

Across the providers we track, on-demand H100 pricing currently ranges from ₹191/hr (NeevCloud) to ₹422/hr (Neysa) — more than a 2x spread between the cheapest and most expensive on-demand option alone. AceCloud, which only offers monthly billing for H100, works out to roughly ₹247/hr once you divide their monthly rate by 730 hours — cheaper than four of the seven on-demand options we track, but more expensive than three. Leapswitch, also monthly-only, works out to roughly ₹375/hr — more expensive than six of the seven on-demand alternatives. Neither monthly provider is winning purely on the strength of being a commitment-based product; both sit inside the range that on-demand pricing alone already covers.

The same pattern shows up on A100 80GB: Leapswitch's monthly rate (≈₹214/hr) lands almost exactly in the middle of the on-demand range we track (₹126–269/hr), not below it.

An important caveat on this comparison

This isn't a true apples-to-apples test — no single provider in our tracked set currently offers both an on-demand rate and a reserved rate for the same GPU, so we can't show you "the same provider is X% cheaper if you commit." What we can show is where monthly-only providers land relative to the broader on-demand market, and right now, that's not obviously cheaper. If you have a quote from a provider offering both tiers directly, compare that specific pair — don't assume the general pattern holds for every provider.

Why this matters more for agent products specifically

An agentic workload's GPU utilization pattern looks different from a typical training job. Training runs are often bursty and finite — you rent a large cluster for days or weeks, then release it. A production agent serving live user requests, by contrast, tends to need capacity available continuously, even if request volume itself is uneven through the day. That continuous-availability requirement is exactly the scenario where a reserved/monthly commitment is conventionally expected to pay off — you're trading flexibility for a lower steady-state rate. Our data doesn't dispute that logic in general; it disputes whether any specific reserved offer is actually delivering on it. The only way to know is to check the real on-demand alternative for the same GPU before signing a monthly commitment.

A practical framework for choosing

Pricing figures above reflect our tracked dataset as verified on 2026-07-20 — check the current H100 and A100 pages for live numbers, since on-demand and monthly rates both change. This is a pricing comparison, not financial advice — confirm current rates directly with any provider before committing to a monthly or reserved plan.