A team in India that keeps an H100 busy all month pays less for it in Mumbai than at the US self-serve on-demand SXM rates we checked on 12 October 2026: ₹1,80,000 a month on ZenoCloud, host included, PCIe or SXM as stated in the quote, against ₹1.88 to 3.0 lakh for 730 hours at $2.69 to $4.29 an hour. The cheaper US routes are a PCIe card at $1.99 an hour (₹1.39 lakh for 730 hours), marketplace offers from $1.73 whose price moves with supply, and spot pricing on eight-GPU enterprise nodes. Add the network distance to a US data centre for a user in India and the dollar invoice, and the monthly node in Mumbai is our recommendation for production work serving Indian users, with GPU Infrastructure Care as an optional service on top. For a few hours of GPU a month, rent from AWS in Mumbai by the hour.
Last verified: 12 October 2026. Market dollar prices are converted at ₹95.78 to the dollar (ECB, 12 October 2026); ZenoCloud's dollar figures are its separately published prices, not conversions. Monthly estimates use 730 hours. Figures marked "our maths" are calculated from the sources, not quoted by them.
Who this comparison is for
Indian startups fine-tuning and serving models for Indian users, enterprises running inference inside the country, and labs or agencies that keep a GPU busy most of the month. If your users are in the US or Europe, or you need a GPU for a weekend, the answer changes, and we say where.
What the US GPU clouds are
Four kinds of provider sit under the label "US GPU cloud":
- Self-serve on-demand clouds. You pick a GPU, pay by the hour in dollars, and get SSH or a container. The two we checked on 12 October 2026 list an H100 SXM at $3.49 to $4.29 an hour on their standard tiers; one also has a lower-priced tier at $2.69.
- GPU marketplaces. Independent hosts list hardware and prices move with supply. The one we checked showed H100 SXM offers from $1.73 an hour with a median of $3.51, A100 PCIe from $0.27 with a median of $0.71, and A100 SXM4 from $0.40 with a median of $0.73. It sells on-demand, interruptible and reserved capacity; the floor and the median describe different points in that market, so we quote both.
- Enterprise neoclouds. Built for training clusters and sold as whole nodes. The one we checked prices an 8× H100 node at $49.24 an hour on demand ($6.16 a GPU-hour, our maths) and $19.71 on spot ($2.46 a GPU-hour), with discounts of up to 60% for committed terms.
- AWS. The hyperscaler sells one H100 in Mumbai (p5.4xlarge) at $8.256 an hour on demand, and an 8× A100 40GB node (p4d.24xlarge) at $26.34.
AWS Mumbai is the only Indian location we could confirm among the four; the other providers' published location information did not establish Indian capacity.
Price: monthly node in India vs hourly in the US
ZenoCloud sells GPU capacity in Mumbai per card per month, host included, with a one-month minimum and no hourly plan; the H100 is listed at one price and the quote states PCIe or SXM. The US clouds sell by the hour. The fair comparison is a card you run all month, in the same form factor.
| GPU | ZenoCloud, Mumbai, per card per month | US on-demand, per GPU-hour (12 Oct 2026) | US on-demand for 730 hours, our maths | AWS Mumbai, on demand |
|---|---|---|---|---|
| H100 80GB, compared with US SXM rates | ₹1,80,000 ($2,199), PCIe or SXM per the quote | $2.69 (lower-priced tier) to $4.29 | $1,964 to $3,132 (₹1.88 to 3.0 lakh) | p5.4xlarge, one H100: $8.256/hr, $6,027 a month (₹5.77 lakh) |
| H100 80GB, compared with US PCIe rates | ₹1,80,000 ($2,199) | $1.99 to $3.29 | $1,453 to $2,402 (₹1.39 to 2.3 lakh) | not sold as a single-card instance |
| A100 80GB | from ₹99,000 ($1,099) | $1.19 (lower-priced tier) to $2.79 | $869 to $2,037 (₹83,000 to 1.95 lakh) | p4d.24xlarge, 8× A100 40GB: $26.34/hr ($3.29 a GPU-hour, our maths) |
| L40S 48GB | ₹55,000 ($649) | $0.79 (lower-priced tier) to $1.09 | $577 to $796 (₹55,000 to 76,000) | g6e.xlarge, one L40S: $2.235/hr, $1,632 a month |
| L4 24GB | from ₹32,000 ($349) | $0.44 to $0.59 | $321 to $431 (₹31,000 to 41,000) | g6.xlarge, one L4: $0.9664/hr, $705 a month |
Sources: the ZenoCloud GPU hosting price list and public price book; two US on-demand clouds' price pages (where a page showed two figures for one card, the range covers both), the marketplace's live price page, the neocloud's price page, and the AWS EC2 on-demand price list for Asia Pacific (Mumbai), all read on 12 October 2026. ZenoCloud prices exclude GST; the US pages add sales tax or VAT where applicable.
What the table says:
- H100 against SXM rates: ₹1,80,000 is below every self-serve on-demand SXM rate we checked for a card that runs the month ($2.69 to $4.29, ₹1.88 to 3.0 lakh). The marketplace floor ($1.73) is lower; its median ($3.51, ₹2.45 lakh) is not. The neocloud's on-demand node works out at $6.16 a GPU-hour; its spot node, $19.71 an hour for eight cards, is about ₹1.72 lakh a card for 730 hours (our maths), lower than ours, if you can use eight cards and live with spot availability. AWS Mumbai is $8.256.
- H100 PCIe: the US is cheaper. $1.99 an hour is ₹1.39 lakh for 730 hours. If a PCIe card suits your model, say so in the quote request, and compare like with like.
- A100: inside the US range. The $1.19 lower-priced tier is about ₹83,000 for 730 hours against our from-₹99,000; the standard tiers ($1.59 to $2.79) cost more.
- L40S: level with the lowest self-serve rate. $0.79 for 730 hours is ₹55,236 (our maths) against our ₹55,000; the $1.09 tier is ₹76,000.
- L4: level. Pick on latency and billing, not price.
- AWS costs the most to run a GPU all month, and is the easiest way to run one for a few hours. More on that below.
Two costs do not appear on any US price page. Indian cards typically carry a foreign-currency markup of 2 to 3.5% on dollar invoices, and GST on imported services is usually self-assessed under reverse charge; ask your accountant how it applies to you. A rupee invoice from an Indian company carries GST on the invoice.
Measured connection time: Mumbai vs the US
We measured TCP connection time from our office in Noida on 12 October 2026, three samples each, to public AWS endpoints in each region:
| Region | TCP connection time from Noida |
|---|---|
| Mumbai (ap-south-1) | 31 to 45 ms |
| Singapore (ap-southeast-1) | 76 to 224 ms (one slow sample; the other two were 76 and 97 ms) |
| Frankfurt (eu-central-1) | 169 to 209 ms |
| US East, Virginia (us-east-1) | 277 to 347 ms |
These are three TCP-connect samples from one office network to AWS endpoints, not a benchmark of any GPU cloud or of application latency, and your numbers will differ by ISP and city. They show the order of magnitude of the network distance: for an interactive application serving Indian users, a US data centre is several times further away in network time than Mumbai. Batch training does not care. Chatbots and voice assistants do.
Billing and support
| ZenoCloud, Mumbai | US self-serve clouds and marketplaces | |
|---|---|---|
| Currency | Rupee invoice with GST | Dollars |
| Term | One month minimum; monthly after that | Per hour or per second; some sell 1 to 6 month reservations |
| What you get | A dedicated card on a host we run, in Mumbai; NVIDIA driver installed | A GPU instance or container; you run the stack |
| Managed option | GPU Infrastructure Care: host and runtime watched 24/7, P1 response under 15 minutes, from ₹25,000 a node a month (₹50,000 account minimum) | Not listed on the price pages we read |
| Try first | A short paid trial node; if you keep it, it becomes your production node | Pay as you go from the first minute |
The care plan has a ₹50,000 account minimum, so an account with a single H100 node pays ₹1,80,000 plus ₹50,000, ₹2,30,000 a month before taxes; accounts with two or more nodes pay ₹25,000 a node (₹2,05,000 for one H100 node within a larger account). Without it you get the node with root access and the driver installed.
Where the US clouds win
- Bursty or short work. At the $3.49 standard rate, ₹1,80,000 buys about 538 hours (our maths at ₹95.78); below that, hourly wins. At AWS Mumbai's $8.256 the crossover is about 228 hours.
- Selected lower-priced offers: H100 PCIe at $1.99 an hour, the $1.19 A100 tier, the $0.44 L4 tier, marketplace floors and eight-card spot nodes all undercut our monthly prices at 730 hours, as the table shows, if your workload fits them.
- Per-request inference endpoints, where a provider offers them. We sell nodes, and a managed serving layer on top of them, not scale-to-zero endpoints.
- Large clusters. The neocloud node we checked is 8 GPUs and its pricing is built for many of them; AWS has the inventory. We supply single nodes and small multi-node setups in Mumbai; H100 is on request, and L40S, RTX PRO 6000 and H200 are ready now.
- Users in the US or Europe. Host near your users. Our published GPU prices are for Mumbai.
Need a GPU for a few hours? Use AWS in Mumbai
For occasional use, AWS in Mumbai keeps the latency and data-residency story intact and bills by the second for Linux instances. The p5.4xlarge (one H100 80GB, 16 vCPU, 256 GiB RAM) is $8.256 an hour on demand, the g6.xlarge (one L4) is $0.9664, and the g6e.xlarge (one L40S) is $2.235, all in Asia Pacific (Mumbai) on 12 October 2026. Watch the bill once the hours add up: at 730 hours the H100 costs ₹5.77 lakh, three times our monthly card. If your AWS account is already the problem, we run AWS accounts too: Managed AWS hosting.
Our recommendation
- Production inference or training for Indian users, card busy most of the month: a ZenoCloud node in Mumbai. For an H100 it is below every self-serve on-demand SXM rate we checked; only marketplace floors, PCIe cards and eight-card spot nodes come in lower. It comes with a rupee invoice and an optional care plan run from India.
- Occasional GPU hours, or an AWS-native stack: AWS in Mumbai, by the hour.
- Interruptible batch jobs that tolerate restarts: interruptible marketplace offers or spot nodes, with checkpointing.
- Large training clusters or users abroad: an enterprise neocloud or AWS in the region your users are in.
Ask for a trial node with the card you plan to run. Measure the latency your users will see, run your workload for the agreed trial period, and compare the invoice with your current dollar bill.
Frequently Asked Questions
Is GPU cloud cheaper in India than in the US?
For an H100 that runs all month, against SXM rates, yes: ₹1,80,000 on ZenoCloud against ₹1.88 to 3.0 lakh for 730 US self-serve hours at $2.69 to $4.29 (12 Oct 2026). PCIe cards, marketplace floors and eight-card spot nodes cost less. A100 sits inside the US range; L40S is level.
Can I rent a GPU by the hour from ZenoCloud?
No. ZenoCloud sells GPU capacity per card per month with a one-month minimum. For hourly work use AWS in Mumbai: one H100 (p5.4xlarge) costs $8.256 an hour on demand as of 12 October 2026.
What latency should I expect from a US GPU cloud in India?
Our 12 October 2026 TCP-connect samples from Noida to AWS endpoints: US East 277 to 347 ms, Frankfurt 169 to 209 ms, Singapore 76 to 224 ms, Mumbai 31 to 45 ms. They measure connection setup, not application latency, but interactive applications feel that distance; batch jobs do not.
Do US GPU clouds bill in rupees?
The ones we checked bill in US dollars. Indian cards typically add a foreign-currency markup, and GST on imported services is usually self-assessed under reverse charge; check with your accountant. ZenoCloud issues a rupee invoice with GST on it.
Does ZenoCloud manage the GPU server for me?
As an option. GPU Infrastructure Care watches the host and runtime 24/7 with a P1 response under 15 minutes, from ₹25,000 a node a month with a ₹50,000 account minimum. Without it you get the node with the NVIDIA driver installed and root access.
Which GPUs are available in Mumbai right now?
As of 12 October 2026: L40S (₹55,000), RTX PRO 6000 96GB (₹1,10,000) and H200 141GB (₹2,20,000) are ready now. L4 from ₹32,000, A100 80GB from ₹99,000, H100 80GB at ₹1,80,000 and B200 from ₹4,59,000 are sourced on request, with a written quote within one business day.