A GPU on your balance sheet feels like an asset. A GPU on your invoice feels like a cost. For most Indian startups, the numbers say the invoice is the cheaper way to hold it.
But "most" is not "all," and the break-even point is more specific than the usual rent-versus-buy slogans admit. Here are real 2026 Indian price points, a three-year comparison you can rerun with your own numbers, and the conditions under which buying does win.
What does GPU as a service mean, and what is the alternative?
GPU as a service simply means renting GPU time from a provider, by the hour or by the month, instead of owning the hardware. The alternative is buying the card or server yourself and running it in your own space or in a colocation facility.
In practice there are four routes in India, for the same H100:
Route
Upfront cost
Typical H100 cost
Worth knowing
IndiaAI subsidised pool
None
About ₹67 to ₹92 per GPU-hour
Approval-gated, for eligible startups and researchers
Indian GPU cloud, on-demand
None
About ₹217 to ₹250 per hour
No commitment, billed per second, minute or hour
Indian GPU cloud, reserved
None
One provider lists ₹219 per hour on a 12-month term, against ₹329 on-demand
A term in exchange for a lower rate
Buy outright
₹28 to ₹40 lakh per card
Plus server, power and operations
You own it, and you carry the risk
Why is the purchase price in India higher than abroad?
Because of the way the hardware reaches you. The card sells for roughly $25,000 to $40,000 abroad, and the Indian price sits higher because of duties and taxes, channel partners such as Dell, HP, Lenovo or Supermicro, and currency exposure. NVIDIA does not sell the H100 at retail in India.
What does buying a GPU actually cost in India?
More than the card. The card is the visible part, and several other items arrive with it.
The card: about ₹28 to ₹40 lakh per H100, with the SXM variant at the upper end
A server to host it: commonly ₹15 to ₹25 lakh
Power and backup: each H100 draws up to 700W, and redundant power with UPS runs about ₹8 to ₹15 lakh per rack
Cooling and space: colocation is commonly quoted at ₹5 to ₹20 lakh a year per rack
Networking for clusters: an InfiniBand fabric starts around ₹50 lakh
People: someone has to run it
A working 8-GPU H100 server lands at about ₹3.5 to ₹5 crore once you add the rest.
What about import duty and GST?
Sources disagree, and the difference is large. Several Indian price guides add 20 to 30 percent to the base price for duties and taxes. At least one analysis says basic customs duty on this class of hardware is 0 percent, with 18 percent IGST charged at clearance that a registered business can claim back as credit.
Which applies depends on classification and your own tax position, so ask a customs adviser before you build a budget around either figure. The swing alone can move a purchase by lakhs.
How long does it take to get hardware?
Often months. Procurement, import clearance and installation are slow compared with renting capacity the same day. One provider guide puts it at three to six months, though that is a vendor estimate. Whatever the real number is for you, it is time your team is not training models.
How does the three-year cost compare?
Here is an illustrative comparison for one H100, using figures from 2026 Indian guides. It assumes ₹33 lakh for the card and ₹80,000 a month for power, space and operations. It ignores resale value and the cost of the host server, both of which would move the result.
Option, over 36 months
25% utilised
50% utilised
100% utilised
Rent at ₹220 per hour
₹14.5 lakh
₹28.9 lakh
₹57.8 lakh
Rent at ₹264 per hour (+20%)
₹17.3 lakh
₹34.7 lakh
₹69.4 lakh
Buy
₹61.8 lakh
₹61.8 lakh
₹61.8 lakh
Read it this way. At today's rates, renting is cheaper than buying even if the GPU runs all day, every day, for three years. If rental rates rise 20 percent, buying overtakes renting only above about 89 percent utilisation.
Stretch the horizon to five years, and the break-even falls to about 84 percent at today's rates, and about 70 percent if rates are 20 percent higher.
What does this comparison leave out?
A lot, and it cuts both ways.
Resale value: unpredictable, and it falls as newer generations arrive, though 2026 rental rates for older cards have held up better than many expected
Obsolescence: a purchased card cannot be swapped for a newer one without reselling it
Tax treatment: depreciation and the GST credit question can change the real cost
The host server: it adds ₹15 to ₹25 lakh, which makes buying look worse than the table shows
Rate changes: rental rates could fall as well as rise
When does buying actually make sense for an Indian startup?
Under a narrow set of conditions:
Your GPU will run above roughly 85 to 90 percent of the time for several years
Your data cannot leave your own premises
You already have people who can run the hardware
Your workload is steady and predictable, so you can size it with confidence
You can wait for procurement without hurting your roadmap
When does GPU as a service win?
For most early-stage teams.
Your usage is spiky, uncertain or still being worked out
You want newer GPUs without being locked to the one you bought
You need capacity this week, not this quarter
You may qualify for IndiaAI subsidised access, which at about ₹80 an hour for a year of round-the-clock use is roughly ₹7 lakh, far below any purchase
**
Can you combine the two approaches?**
Yes, and many teams do. Cover your steady baseline with a reserved term, run bursts and experiments on on-demand, and use subsidised capacity where you qualify. You get most of the discount that ownership would give, without the capital outlay or the lock-in.
Conclusion
Buy only if you can keep the card busy nearly all the time for years, and rent for everything else. Run the three-year sum with your own utilisation, ask your adviser about duty and GST, and measure your real usage for a few months before you commit lakhs to hardware. Prices and rates change quickly, so treat the figures here as a starting point, not a quote.
Frequently asked questions
Is it ever cheaper to buy a GPU in India?
Yes, if utilisation stays very high for several years and you avoid the hidden costs. On the figures above, the break-even sits around 84 to 90 percent utilisation, depending on the time horizon and where rental rates go.
Can any startup use IndiaAI subsidised GPUs?
No. Access is approval-gated, so you need to apply and meet the eligibility criteria. For teams that qualify, it is the cheapest legitimate route at roughly ₹67 to ₹92 per GPU-hour.
Should I wait for GPU prices to fall before deciding?
Usually not. Rental rates rose through much of 2026, and buying at a high point does not protect you from further movement. If you start on rented capacity, you can switch the decision later with real utilisation data in hand.
Top comments (0)