GPU Rental vs Buy in India (2026): When ₹/Hour Beats a ₹35 Lakh H100 | GPUIndia
GPU Rental vs Buy in India (2026): When ₹/Hour Beats a ₹35 Lakh H100
Rent beats buy in India until you sustain near-24/7 utilization for 3+ years. At ₹209/hr (E2E Networks), renting an H100 for the typical 8-hr/5-day week costs about ₹4.4 lakh a year; buying one costs ₹30–40 lakh upfront. Buy only above ~70% sustained utilization.
Last verified: 2026-08-06. Live rental rates from the GPUIndia pricing feed (Parallel.ai, updated daily). Buy prices are India street ranges from dated 2026 primary guides, linked below. All ₹ figures round to the nearest lakh.
Is it cheaper to rent or buy a GPU in India?
Yes — renting is cheaper in India for almost every team below ~70% sustained utilization. At today's live rates, an H100 on-demand at ₹209/hr costs about ₹18.3 lakh per year if run 24/7, and only ₹4.4 lakh per year if used the way most Indian teams actually use a GPU (8 hours a day, 5 days a week). Buying the same card costs ₹30–40 lakh upfront before you add power, cooling, and maintenance that roughly double the number in practice. Unless your GPUs are busy around the clock for three years or more, rental wins on cost alone — and that is before you count the fact that a rented GPU can be swapped for a newer one, while a bought H100 quietly depreciates.
This is a decision between CAPEX (buying) and OPEX (renting). Buying converts a big one-time hardware purchase into an asset that needs power, cooling, uptime, and a resale exit. Renting converts it into a per-hour operating cost with zero upfront cash, no maintenance, and instant scale.
Rent vs Buy: Full Cost Comparison (single H100 80GB, India, 3-year horizon)
| Scenario | Rent on-demand (₹209/hr) | Rent spot (~₹94/hr) | Buy (₹35L + running costs) |
|---|---|---|---|
| 24/7 utilization | ₹54.9 L / 3 yr | ₹24.7 L / 3 yr | ≈ ₹52 L / 3 yr (1.5× sticker) — toss-up |
| 50% utilization | ₹27.5 L / 3 yr | ₹12.3 L / 3 yr | ≈ ₹52 L / 3 yr — rent wins |
| 24% utilization (8 hr × 5 day/wk) | ₹13.0 L / 3 yr | ₹2.0 L / year | ≈ ₹52 L / 3 yr — rent wins decisively |
| Upfront cash | ₹0 | ₹0 | ₹30–40 L (add import duty + GST) |
| Power + cooling | Included | Included | ~700 W/GPU + cooling + infra |
| Scales up/down | Instant | Instant | No — capacity is fixed |
| Stale-ware risk | None (use newest) | None | H100 vs B200: depreciates |
Rent columns use E2E Networks ₹209/hr on-demand and Vast.ai/Spheron spot ≈₹94/hr (see live feed). Buy TCO = ₹35L sticker × 1.5 for power/cooling/maintenance over 3 years — a labeled estimate consistent with industry TCO research (hardware is ~35% of a 5-year GPU TCO; see sources). FX ₹95/USD for the spot column, labelled approximation (BookMyForex, Aug 5 2026).
The table is the whole argument in one view. At 24% utilization — the reality for most Indian startups, research labs, and freelancers — renting on-demand costs roughly a quarter of what buying costs, and renting spot costs less than a tenth.
How much does an H100 80GB cost to buy in India?
An H100 80GB PCIe costs roughly ₹30–40 lakh to buy in India in 2026; the SXM variant runs ₹40–50 lakh. An 8-GPU H100 server slips into the ₹3.5–5 crore range. Import duties and GST add about 20–30% to the already high global sticker of roughly US$25,000 per GPU, which is why Indian street prices land so far above the US list price.
The ₹30–40 lakh band is confirmed across multiple 2026 India guides: Cyfuture's H100 India price guide lists H100 80GB PCIe at ₹30–40 lakh and SXM at ₹40–50 lakh; JarvisLabs' H100 India breakdown puts PCIe at ₹25–30 lakh with a DGX H100 at ₹3–4 crore. Both are 2026-dated primary sources we verified on 2026-08-06.
The sticker is not the true cost. A single H100 draws ~700 W; a small cluster needs real power delivery, cooling, rack space, networking (InfiniBand for multi-GPU), and a person to run it. Industry TCO analysis finds hardware is only about 35% of a 5-year total cost of ownership — power, cooling, space, staff, and maintenance push real ownership cost to roughly 2.5–3× the sticker (Introl GPU TCO model, Dec 2025 update). For our single-H100 comparison we conservatively used 1.5× the ₹35L sticker over 3 years (≈ ₹52L real TCO), which already pushes the buy-vs-rent break-even in renting's favour.
How much does renting an H100 cost per hour in India?
Indian and global providers rent H100s from ₹209/hr on-demand to $0.99/hr spot (~₹94/hr). E2E Networks, an Indian provider, lists H100 on-demand at ₹209/hr. On global marketplaces, H100 on-demand runs $2.50/hr (Vast.ai, Spheron) up to $8.50/hr on AWS; spot drops to $0.99/hr on Vast.ai and Spheron.
These are live rates from the GPUIndia pricing feed (source: Parallel.ai, last_updated 2026-08-06, accessed 2026-08-06). The same feed gives you the wider menu if H100 is more power than you need:
| GPU | On-demand (low→high) | Spot (low→high) |
|---|---|---|
| H100 80GB | $2.50 (Vast/Spheron) → $8.50 (AWS) | $0.99 (Vast/Spheron) → $5.00 (AWS) |
| A100 80GB | $1.67 (Vast/Spheron) → $4.00 (AWS) | $0.61 (Vast/Spheron) → $2.40 (AWS) |
| RTX 5090 | $0.58 (Spheron) / $0.76 (Vast) / $1.10 (RunPod) | — |
| B200 (Blackwell) | $4.50 (Spheron) / $5.00 (Nebius) | $2.12 (Spheron) |
The ₹209/hr E2E rate is the direct INR anchor for the buy-vs-rent math here — it means renting an H100 on-demand for one year of continuous 24/7 use costs ₹18.3 lakh, less than a single H100's ₹35L purchase price. See our other guides for a deeper dive on spot vs on-demand H100 pricing and the August 2026 H100 price trend.
When does buying a GPU beat renting it?
Buying wins only when you sustain more than roughly 70% average utilization for 3+ years. At that level the billed-on-demand hours finally exceed what you would have spent owning the card, including its power and maintenance. In our single-H100 table, on-demand renting at 24/7 utilization (₹54.9L over 3 years) ends up essentially neck-and-neck with the buy total (≈₹52L), and the buy only clearly wins if the card stays near-maximally busy the whole time and has a usable resale value at the end.
This matches independent industry analysis: total-cost-of-ownership modeling shifted the buy-vs-cloud break-even in favour of cloud below 60–70% utilization as GPU rental prices fell (Introl TCO model, Dec 2025). Below that utilization, a rented GPU is simply cheaper per hour of actual useful work.
What is the GPU utilization trap?
The utilization trap is buying expensive silicon you only occasionally use — the idle hours cost you money every month. If you buy an H100 for ₹35L and run it 8 hours a day, 5 days a week (about 24% utilization), you get 2,080 useful compute hours a year. Renting those same hours on-demand costs just ₹4.4 lakh a year. Buying commits ₹35L for the right to use a fraction of that capacity, and every idle hour still drags on the asset's value.
Because GPUs depreciate fast — hyperscalers have moved to 6-year depreciation schedules, and used-H100 street value fell sharply once supply caught up (Hashrate Index, used GPU market, 2026) — the cost of idle silicon is a real loss, not just an opportunity-cost footnote. A GPU that is 76% idle is, economically, 76% wasted CAPEX. Rental never has this problem: you stop paying the moment you stop using it.
How long does it take for a bought GPU to pay for itself versus renting?
A bought H100 only pays for itself against on-demand rental after roughly 3 years of 24/7 use — and against spot rental it never does within the card's useful life. Here is the arithmetic, all from live rates:
1. Sticker parity (on-demand): paying ₹35L ÷ ₹209/hr = 16,746 rental hours ≈ 1.9 years of continuous 24/7 rental to equal the sticker price alone.
2. Real ownership cost: add power (700 W), cooling, maintenance, and admin over 3 years — the buy total rises to ≈ ₹52L (our 1.5× estimate). On-demand rental at 24/7 hits ₹52L at about 2.9 years. So ownership break-even lands near 3 years of non-stop use — right at the edge of, or beyond, the card's economic life.
3. The realistic case: at 24% utilization, buying takes ~8 years to break even against on-demand rental — long past the point the H100 is superseded.
4. Spot rental: at $0.99/hr (≈₹94/hr), continuous 24/7 rental is only ₹24.7L over 3 years — cheaper than owning even at maximum utilization. Ownership never beats spot rental within the card's usable life.
Rule of thumb: if your H100 would sit below ~70% utilization, rent. Only buy when you can commit it to near-24/7 training for 3+ years and have a resale plan.
Should a startup rent or buy GPUs in India?
Startups should rent — the CAPEX is too large, the utilization too low, and the flexibility too valuable. Tying up ₹30–40+ lakh of a young company's cash in one depreciating card is a poor use of scarce capital when the identical H100 is available on-demand at ₹209/hr with zero commitment. Rental also lets you use the newest silicon (A100, H100, then B200) instead of being locked to the generation you bought (explore the H100 vs H200 vs B200 comparison).
The one exception is an established firm with a guaranteed, predictable, near-continuous training load — a product that genuinely runs GPUs 24/7 for years. That team can amortise ownership below rental cost. But that describes almost no Indian startup; it describes big labs and inference-hosting companies with proven demand. If your GPU budget is uncertain, is funded by a single round, or would starve salary or marketing, rent.
Is buying an H100 in India even worth it in 2026?
For most buyers, no — Blackwell (B200) is superseding Hopper (H100), and rental has made ownership hard to justify. A ₹35L H100 bought today faces a shrinking resale market as B200 supply grows and its own successor (B300/GB300) arrives. The live pricing feed already shows B200 available on-demand at $4.50/hr (Spheron) and spot at $2.12/hr (Spheron), and our B200 pricing guide for India breaks down those ₹ rates. Every month, a bought H100 is worth less at resale while a rented one simply moves on to whatever is current.
Buy for these specific reasons only: you need deterministic, low-latency local inference with strict data-residency controls (your data never leaves the building); you are building a business of renting that GPU to others; or your workload is so constant that the 3+ year payback genuinely fits, and you have the power, cooling, and talent to run it. Otherwise, rental is the rational 2026 choice — see our DGX-style buy-vs-cloud TCO deep dive for the same conclusion at a larger scale.
Frequently Asked Questions
Is it cheaper to rent a GPU or buy one in India?
Renting is cheaper for almost every India-based team except those that sustain near-24/7 utilization for 3+ years. At ₹209/hr on-demand, renting an H100 for a typical 8-hr/5-day work week costs about ₹4.4 lakh a year; buying the same card costs ₹30–40 lakh upfront plus power, cooling, and maintenance.
How much does it cost to rent an H100 per hour in India 2026?
E2E Networks rents H100 on-demand at ₹209/hr. Global rates range from $2.50/hr on-demand (Vast.ai, Spheron) to $8.50/hr on AWS, with spot as low as $0.99/hr on Vast.ai and Spheron (live feed, updated 2026-08-06).
Is buying an H100 worth it in 2026?
Only if you can keep it above roughly 70% utilization for 3+ years. Against on-demand rental a bought H100 pays for itself in about 3 years of 24/7 use; against spot rental it never pays back within its usable life. Blackwell (B200) supersession also shrinks the H100 resale market.
What is the GPU utilization trap?
It is buying expensive silicon you only occasionally use. At 24% utilization a ₹35L H100 delivers just ₹4.4L-worth of useful on-demand compute a year — idle hours still depreciate the asset, so the majority of your CAPEX sits wasted.
How long does it take for a GPU to pay for itself?
At ₹209/hr on-demand it takes about 1.9 years of continuous 24/7 rental to equal the ₹35L sticker, and nearly 3 years to match the full ownership TCO including power and maintenance. At realistic 24% utilization, break-even stretches to ~8 years — beyond useful life.
Should I rent or buy GPUs for an AI startup in India?
Rent. ₹30–40L of startup capital in one depreciating card is a poor use of cash when the identical H100 is available on-demand at ₹209/hr with instant scale-up/down, no maintenance, and access to newer silicon. Only established firms with proven 24/7 demand should consider buying.
Are GPU rental rates in India fixed?
No. Spot prices move hourly and on-demand rates shift with supply and market cycles. Check the live feed before provisioning — the figures here are a 2026-08-06 snapshot.
Sources
- Live rental rates (H100 $2.50/$0.99; A100 $1.67/$0.61; RTX 5090 $0.58; B200 $4.50/$2.12; E2E H100 ₹209/hr): GPUIndia pricing feed, `pricing-data.json`, source Parallel.ai, `last_updated 2026-08-06T09:41:29Z`, accessed 2026-08-06.
- H100 India purchase ₹30–40L (PCIe), ₹40–50L (SXM): Cyfuture — NVIDIA H100 GPU Price in India (2026) — https://cyfuture.ai/blog/nvidia-h100-gpu-price-in-india (accessed 2026-08-06).
- H100 India purchase ₹25–35L PCIe, ₹3–4Cr DGX, import duty +20–30%, 700 W/GPU: JarvisLabs — NVIDIA H100 price in India (2026) — https://docs.jarvislabs.ai/blog/h100-price-india and https://jarvislabs.ai/blog/h100-price (accessed 2026-08-06).
- Hardware ≈35% of 5-yr GPU TCO; cloud break-even below 60–70% utilization: Introl — GPU Infrastructure TCO Model (Dec 2025 update) — https://introl.com/blog/gpu-infrastructure-tco-5-year-cost-model (accessed 2026-08-06).
- Power draw (>5.6 kW for 8× H100), idle-silicon depreciation: Lyceum Technology — Total Cost of Ownership for a GPU Cluster (2026) — https://lyceum.technology/magazine/total-cost-ownership-gpu-cluster-2026/index.html (accessed 2026-08-06).
- GPU depreciation / used-market value fall: Hashrate Index — Inside the Used GPU Market (2026) — https://hashrateindex.com/blog/used-gpu-market-pricing-deprecation-secondary-ai/ (accessed 2026-08-06).
- FX ₹95.2/USD (spot column conversion): BookMyForex — USD to INR forecast, Aug 5 2026 — https://www.bookmyforex.com/currency-converter/usd-to-inr/forecast/ (accessed 2026-08-06).
Disclosure: GPUIndia is an independent GPU price comparison site. We earn commissions through affiliate links at no extra cost to you. Rental rates are a live snapshot and change daily; always confirm against the live feed before provisioning. Converted ₹ figures on global USD rates use a labelled ₹95/USD approximation.