Hosting your GPUs in India: latency, power, support and skills
Short answer: yes, India is a good place to host GPUs, for the right workloads and the right building. It is close to users in India, the Middle East and South-East Asia. Building a data centre costs about half as much as in the US. Grid power for large users costs about the same as in Virginia, and less than in Germany or Ireland. The country is not short of power, but your site may be. And quality varies by building, so check certificates and SLAs (service level agreements: the uptime and support terms) in writing.
More European and US companies are asking us whether to colocate their GPUs in an Indian data centre. Here are honest answers to the questions they ask most, with the numbers and the catches. For the market itself, see our India data centre guide.
The short answers
Question | Short answer |
|---|---|
Are European and US companies really coming? | Yes, but it is early. US GPU cloud companies signed large India deals in 2026. European firms show up so far mainly as data centre developers. Most of the money is still pledged, not spent. |
What is the latency to Europe and the US? | On Microsoft's network, about 135 to 150 ms (milliseconds) round trip to Western Europe and about 200 ms to the US East coast. Dubai is 32 ms from Pune, and Singapore 37 ms from Chennai. Public internet routes vary more. |
How fast and reliable is the internet? | Large and growing, with about 15 international subsea cables live. But most routes to Europe cross the Red Sea, and repairs can take months. |
Is India short of power? | Not nationally. India met a record 270.8 GW peak in May 2026, only 12 MW short. The real question is power at your site. |
Are support and SLAs good? | At the best sites, yes. But independent proof is thin, and no major operator publishes its SLA terms. Put them in your contract. |
Has India produced global players? | Yes, in IT services, telecom and software, such as TCS, Infosys, Tata Communications, Airtel and Zoho. Not yet in data centres or AI hardware. |
Labour is cheaper, but is the quality there? | Building costs about half the US level, and the software talent pool is huge. Hands-on skills for dense GPU halls are thinner, so check the team. |
Can India meet cutting-edge AI standards? | At a few new campuses, increasingly yes. Across the market, not yet. Choose the building, not just the country. |
Why are European and US companies looking at India?
Three things pull them in: fast-growing Indian demand for AI, heavy cloud investment, and lower building and space costs. The interest is real but early. These are the biggest signals since 2025.
What happened | Size | When |
|---|---|---|
A B300 cluster at L&T's Chennai campus, mostly for a US GPU cloud company | 10,000 B300 GPUs; an order worth ₹10,000 crore to ₹15,000 crore (a crore is 10 million rupees; about US$1.05 billion to US$1.57 billion) | Announced August 2026; go-live expected by November 2026 |
A US GPU cloud company signed for a whole AdaniConneX campus at Taloja, Navi Mumbai, to run NVIDIA's next-generation Vera Rubin systems | 240 MW, with an option for 240 MW more | Signed October 2026; first phase due mid-2028 |
OpenAI became the first customer of TCS's new AI data centre platform in India | 100 MW to start, with an option to grow to 1 GW | February 2026 (a plan) |
NVIDIA agreed to run one of Asia-Pacific's largest DGX Cloud clusters (NVIDIA's own AI cloud service) inside an Indian operator's planned B300 supercluster | Over US$1 billion across four years | Agreed February 2026; not yet confirmed live |
Amazon, Microsoft and Google, the biggest cloud providers (hyperscalers), pledged money for AI and cloud in India | Amazon US$21 billion or more, Microsoft US$20.5 billion, Google about US$15 billion: more than US$56 billion in all | Pledged in 2025 and 2026, for spending to 2030 |
Who built India's new data centre space | US developers 28%, European 24%, Indian 48% | First half of 2026 |
Demand is local too. The government's IndiaAI Mission had onboarded 38,000 GPUs by February 2026 for startups, researchers and public bodies. Some sectors must keep data in India: payment data has had to stay in the country since 2018.
Be clear about what is proven. Most of this money is pledged for 2026 to 2030, not spent. It is also better proven that foreign firms buy capacity in India than that they bring their own GPUs. In Chennai, L&T says it will deploy the GPUs and does not say who owns them. Only the Taloja deal is clearly a foreign company bringing its own platform.
The two US deals are described as serving customers in India, and neither says it will serve other countries from there. IndiaAI sells subsidised GPU time to eligible users, which sets a low price for that part of the market. India's data centre tax break runs to 2047. It is unlikely to suit a foreign company that owns its GPUs in India, so check with your tax adviser. More on all of this in our pages on why India and hyperscaler investment in India.
What is the latency from India to Europe, the US and Asia?
On Microsoft's cloud network in July 2026, India was about 135 to 150 ms round trip from Western Europe. It was about 200 ms from the US East coast. It is much closer to the Gulf and South-East Asia: 32 ms from Pune to Dubai and 37 ms from Chennai to Singapore. A round trip is the time a message takes to get there and back.
The table shows Microsoft's published figures from Pune (its Central India region, near Mumbai) and Chennai.
From India to | From Pune (near Mumbai) | From Chennai |
|---|---|---|
London | 141 ms | 150 ms |
Frankfurt | 140 ms | 151 ms |
Amsterdam | 139 ms | 152 ms |
Paris | 135 ms | 147 ms |
US East coast (Virginia) | 201 ms | 220 ms |
US West coast (California) | 217 ms | 201 ms |
Dubai | 32 ms | 50 ms |
Singapore | 53 ms | 37 ms |
Tokyo | 122 ms | 106 ms |
Sydney | 145 ms | 128 ms |
For comparison, Singapore is 169 ms from Western Europe, so India is closer to Europe than Singapore is. London to the US East coast is 78 ms. Inside India, the big cities are close to Mumbai on the public internet: Bengaluru 13 ms, Hyderabad 17, Chennai 22 and Delhi 27.
What that means for your workload
Workload | Does the distance matter? |
|---|---|
Training and fine-tuning | Hardly. The work happens inside the data centre. Only uploading data and downloading results cross the distance. |
Batch jobs, such as overnight document or video processing | No. |
Chat, search and AI agents for users in India, the Middle East or South-East Asia | India is a good fit: 13 to 27 ms inside India, and 32 ms to Dubai and 37 ms to Singapore on Microsoft's network. Public internet routes can be much slower, so check yours. |
Chat for users in Europe | Each round trip takes about a seventh of a second. Streaming text hides most of it. |
Live voice and real-time apps for users in Europe or the US | Users will feel it, especially from the US at about 200 ms. Serve them from a site nearer to them. |
Moving large datasets between continents | Plan for bandwidth and transfer time, not latency. |
One caution: these figures are for Microsoft's own private network. On the public internet, some routes from India take long detours. In one October 2026 test, Mumbai to Dubai took 224 ms, against 32 ms from Pune on Microsoft's network. A few routes to Europe were faster than Microsoft's, and others switched between fast and slow paths during the day.
So ask carriers for their 95th percentile latency (the delay that 95% of traffic stays under), not just the average. Put it in the contract. Our page on connectivity and interconnection in India goes deeper.
How fast and reliable is India's internet connectivity?
International capacity is large and growing, but it is concentrated. International cables land mostly in Mumbai and Chennai, and most routes to Europe cross the Red Sea.
Item | What India has (October 2026) |
|---|---|
Subsea cables | About 15 international cables live and 5 planned, plus 2 domestic cables to India's islands. TeleGeography lists 18 in service because it also counts 2Africa, whose India branch is not yet confirmed live. |
Where international cables land | Mostly Mumbai (11, plus 2Africa Pearls) and Chennai (6) |
New cables due | Three more to Mumbai in 2027: SEA-ME-WE 6, India Europe Xpress and Raman. Raman's Red Sea section was reported delayed in January 2026, so treat 2027 as a target. 2Africa Pearls has landed but is not yet confirmed live. |
A large new cable to Singapore | MIST, over 216 Tbps (terabits per second), ready for service between December 2024 and June 2025 (sources differ) |
Internet exchange ports (where networks connect to swap traffic directly) | Published prices of about ₹50,000 (US$525) a month for a 10G (10 gigabits per second) port, and ₹250,000 (US$2,625) for 100G |
Direct links to cloud providers | AWS Direct Connect at 8 Indian locations, up to 100G |
Internet transit (paid access to the rest of the internet), bought in bulk | About US$0.97 per Mbps a month on a 100G port in Mumbai, before the local link to your data centre. That is 3.7 times Singapore's price, down from 9.2 times, after falling about 35% a year since 2023. |
Home broadband | Median 64 Mbps. That is the line to homes, not data centre links. |
Most cables from Mumbai to Europe pass through the Red Sea. Cuts near Jeddah in September 2025 slowed traffic in India, Pakistan and the UAE. Repairs can be slow. In a 2023 submission to TRAI, India's telecom regulator, one industry respondent put repair-ship response times at 3 to 5 months. The ships are based abroad, and customs and permit clearances add time. So ask your data centre and carriers for two physically separate paths, and which cables each one uses.
No Indian operator publishes a price list for cross-connects (the cable that links your racks to a carrier). Negotiate them into the contract. For the full list of cables and landing stations, see subsea cables and landing stations in India.
Is India really short of power?
No, not at the national level. India met a record 270.8 GW peak in May 2026, only 12 MW short. The real question is power at the site you choose.
Measure | Figure |
|---|---|
Record peak demand | 270.8 GW in May 2026, met with only 12 MW short |
Electricity demand not met, FY2025-26 (April 2025 to March 2026) | 0.03% |
Electricity demand not met, April to August 2026 (a hot El Niño summer) | About 0.1% |
Installed generating capacity | 558.9 GW (September 2026), 55% of it non-fossil |
Where the electricity actually comes from | About 70% from coal and lignite (April to June 2026) |
Average power supply in cities | 23.4 hours a day (FY2025-26) |
Two notes. Most capacity is non-fossil, but most of the electricity still comes from coal. So a green power claim needs a renewable power contract, not just grid supply. And the tightest hour is the evening, after solar output falls away.
The real question: power at your site
National numbers do not get you a connection. For a GPU hall, three things decide it:
- the sanctioned load: the power the utility has approved for the site
- the substation that feeds it
- a firm date when the power will be switched on
India publishes no figure for how long a new grid connection takes. In Europe's main hubs (Frankfurt, London, Amsterdam, Paris and Dublin), it can take 7 to 10 years. So in India, ask for evidence, not promises. City supply averages just under 24 hours a day, so check for two independent grid feeds and on-site backup power. Several states offer two grid feeds to large data centres.
What power and colocation cost in India
Cost | India | For comparison |
|---|---|---|
Grid power for large users, per kWh | ₹6.2 to ₹10.9 (US$0.065 to US$0.114) in the six main data centre states | Virginia US$0.10, Texas US$0.07; Germany US$0.15, Netherlands US$0.14, Ireland US$0.20 |
Colocation rent (renting data centre space for your own servers), per kW a month | Mumbai US$73 to US$105 for 250 kW to 1 MW (US$63 to US$79 at 5 MW or more) | Northern Virginia US$190 to US$235, Frankfurt US$235 to US$265, Singapore US$330 to US$475 |
Cost to build, per MW, standard build (excluding land and IT) | US$7.4 million | USA US$15.2 million, Japan US$16.0 million |
These figures use different bases, so treat them as a guide. Indian power is our estimate from FY2026-27 state tariffs (Tamil Nadu's are FY2025-26) for a steady round-the-clock load. It leaves out electricity duty, fuel adjustment and open access charges. US figures are July 2026 state averages for industry. EU figures are for large users in the first half of 2026, before VAT but including other levies.
Mumbai rents are from 2025 and the others from early 2026. Northern Virginia and Frankfurt rents are for 250 to 500 kW, and Singapore's is an asking rent. Whether power is included in rent varies by market. Build costs are from Cushman & Wakefield for 2026, which puts AI-ready capacity in India at about US$9.0 million per MW.
One trap: the tariff category. In Maharashtra, a data centre billed as commercial pays about 54% more than one billed as industrial on a high-voltage (HT) connection. On the highest-voltage (EHV) connections, it pays about 76% more. Get the category in writing. State-by-state prices are in our India data centre cost comparison.
Are Indian data centres good on support and SLAs?
The best ones are. But independent proof is thinner than the marketing, and quality varies by building, not just by brand. Here is what the certificates tell you.
Certificate | What it proves | In India (October 2026) |
|---|---|---|
Uptime Institute Tier IV, finished building | The finished building was tested to the highest design standard | 4 sites, and only 1 is a commercial colocation campus (in Navi Mumbai) |
Uptime Institute Tier III, finished building | The finished building was tested at the level below Tier IV | 23 sites |
Uptime Institute Operational Sustainability award | How well the site is run day to day | 3 sites. One more, SBI's Hyderabad data centre, holds Uptime's separate operations stamp. |
Uptime Institute Tier, design only | The drawings meet the standard. It says nothing about the finished building. | Common. Ask whether the building itself is certified. |
TIA-942 "Rated" | A different certification scheme | Used by several large operators. Check who issued it. |
SOC 2 Type 2, ISO 27001, PCI DSS | Independent audits of security and controls | Common at large operators. Ask for the report that covers your building. |
These counts come from Uptime's public list, which may not show every certificate.
No major Indian operator publishes its SLA schedule (the terms for uptime, response times and service credits). Lines like "100% uptime" on a website are not contract terms. Put these in writing:
Put in the contract | Why |
|---|---|
Separate power and cooling SLAs | A GPU hall needs both, and one without the other leaves a gap |
How downtime is defined and measured | It decides whether you get service credits |
Inlet temperature and humidity range | GPUs slow down or shut off when they run too hot |
Liquid-cooling SLAs (the coolant units and water loops) | Liquid-cooled GPU racks, such as NVIDIA's GB300 NVL72, depend on them |
Service credit levels and caps | What you get back when the SLA is missed |
Remote-hands response time (how fast on-site staff act on your request) | Large operators offer 24x7 remote hands, but rarely publish response times |
A SOC 2 Type 2 report and incident history for your building | Shows how the site performs in practice, not just on paper |
India's cyber security agency, CERT-In, has rules that apply to foreign companies too. Cyber incidents must be reported within 6 hours, and logs kept for 180 days. This is general information, not legal advice, so check how it applies to you with your counsel. Our guide to data centre operators in India lists who runs what, and where.
What global players has India produced?
Plenty, mostly in IT services, telecom and software. All figures are the companies' own.
Company | What it does | Scale |
|---|---|---|
TCS | IT services | Over US$30 billion revenue (FY2025-26); 598,000+ people in 56 countries |
Infosys | IT services | US$20.3 billion revenue (last 12 months); 328,000+ people in 59 countries |
HCLTech | IT services | US$14.8 billion revenue (12 months to September 2026); 230,000+ people in 60 countries |
Wipro | IT services | US$10.5 billion revenue (FY2025-26); 240,000+ people in 65 countries |
Tata Communications | Global network and subsea cables | Says over 35% of the world's internet traffic runs on its network |
Bharti Airtel | Telecom | 681 million customers in 15 countries, in India and Africa (June 2026) |
Zoho | Business software | 150 million users and its own data centres. It has never taken outside investment. |
Global companies also run 2,117 capability centres in India, their own engineering and operations hubs. Together they employ 2.36 million people (Zinnov and Nasscom, FY2025-26).
The honest gap: India has no global data centre champion yet, and no GPU or chip champion. For a GPU owner, the most relevant names are the networks: Tata Communications and Airtel carry traffic around the world.
Labour is cheaper, but is the quality there?
For software and AI, yes. For hands-on data centre work, it depends on the team.
Where India is strong | The figure |
|---|---|
Cost to build a data centre | US$7.4 million per MW for a standard build, about half the US cost |
Tech workforce | 5.95 million people (Nasscom estimate, FY2025-26) |
Developers on GitHub | 21.9 million, second only to the US |
AI skills | First in the world on LinkedIn's AI skills measure: AI skills appear on Indian members' profiles about 3 times as often as the global average (Stanford AI Index 2026) |
Where to check | What the data shows |
|---|---|
Construction labour | An average 30 to 35% labour shortage across 71 Indian construction projects of all kinds, not only data centres (Turner & Townsend, first half of 2026) |
Specialist skills | Experienced electrical and cooling engineers, and commissioning teams (who test a new hall before it goes live), are scarce |
Data centre training | New and small. Microsoft's academy at IIIT Hyderabad started with 30 students. Vertiv plans to train 1,500 students in 2026. |
Proven day-to-day operations | Only 3 Indian sites hold Uptime's Operational Sustainability award for how a site is run |
Cheap does not mean low quality. But dense, liquid-cooled GPU halls are new everywhere, and in India that experience is still thin in places. So meet the team. Ask for the named site team and their shift pattern. Ask about their liquid-cooling experience, their training and certifications from equipment makers, and their staff turnover.
Can India deliver global standards to cutting-edge AI companies?
At a few new campuses, increasingly yes. Across the market, not yet.
Signal | Status (October 2026) |
|---|---|
A liquid-cooled B300 supercluster at Greater Noida | 20,736 GPUs planned and due by August 2026; go-live not yet announced |
The B300 cluster in Chennai | 10,000 GPUs; go-live expected by November 2026 |
Reliance's AI data centres at Jamnagar | More than 120 MW expected in the second half of 2026; not yet announced live |
Dense racks | Operators say new builds can take 60 to 140 kW or more per rack with liquid cooling. That is what they are built for, not halls already running. |
Leading AI companies committing | OpenAI (100 MW with TCS), NVIDIA (DGX Cloud, over US$1 billion) and a US GPU cloud's 240 MW Vera Rubin campus from mid-2028 |
Hyperscalers | Microsoft's new region in Hyderabad, with three separate zones, went live in August 2026 |
Top independent certification | Only 1 commercial colocation site holds both Uptime's Tier IV certificate for the finished building and its Operational Sustainability award |
What is still to come | 76% of future data centre commitments are early stage or land waiting to be built on |
Two practical points for GPU owners. First, US export rules. India is in the US's Country Group A:5, a group of close US partners. As of October 2026, most shipments of NVIDIA GPUs to India need no US licence. The exception is an end user headquartered in, or owned from, China, Macau or another country under a US arms embargo.
US checks on who uses the GPUs, and for what, still apply, including the US Entity List of restricted companies. A paused US rule on companies owned by listed firms is due back after 9 November 2026, unless it is extended. The rules change often.
Second, Indian import rules. Servers need a DGFT import authorisation (a government import licence) each calendar year. Confirm both points, and the customs duty, with your export counsel and a customs broker.
Before you ship GPUs to India: a checklist
Area | Ask for |
|---|---|
Power | The sanctioned load for the site, a firm date to switch the power on, and the tariff category in writing |
Cooling | Liquid-cooled capacity that is built, powered and tested, per hall, in writing |
Network | Two physically separate paths, which cables each uses, 95th percentile latency, and cross-connect prices |
Certificates | Certificates for the finished building, not just the design, and a SOC 2 Type 2 report for your building |
SLA | Power and cooling SLAs, service credits, remote-hands response times and the site's incident history |
Team | The named site team, their liquid-cooling experience and staff turnover |
Rules | A CERT-In contact, data rules for your customers' sector, US export screening and the DGFT import authorisation |
Our guide for GPU clouds entering India covers the entry options, the rules and the costs in more detail.
How FlexiCloud can help
We are based in Kochi, and we are in regular contact with Indian data centre operators, carriers and GPU buyers.
Bringing your own GPUs? Our data centre (DC) advisory helps you choose a site and partners, and plan power, connectivity and the compliance path. It also covers commercial terms and introductions to Indian customers.
Would you rather rent first? We source dedicated NVIDIA GPU nodes, such as B200 and B300, through more than ten partner providers in India, the EU and the US. Before you sign with any provider, see our questions to ask a GPU provider.
We do not own data centres, so we have no site of our own to sell you. If any party we recommend will pay us, we tell you in writing before you sign. Book a call, or start with our India data centre guide and its tracker of projects across the country.
Where these figures come from
Latency: Microsoft's Azure network latency statistics (30 days to 30 July 2026) and WonderNetwork public pings (9 to 23 October 2026). Power: India's Central Electricity Authority, the Ministry of Power's answer in Parliament of 23 July 2026, our own estimate from state tariff orders, the US Energy Information Administration and Eurostat. Cables: TeleGeography and TRAI. Certificates: Uptime Institute's public list, read on 23 October 2026. Costs and skills: Cushman & Wakefield, CBRE and Turner & Townsend. Company figures come from each company's own pages. US$ figures use ₹95.25 and €1 = US$1.119 (13 October 2026).