Why is India attracting more AI infrastructure lately?
Every funding round, every investor, and who is really making money from AI infrastructure, in one document. Have a look ↓
Summary
India is attracting more AI infrastructure because AI companies now see India as a live demand market with enough users, enterprise adoption, local partners, power-linked sites, and policy support to justify building compute inside the country.
The story has moved from pledges to execution. In August 2026, Microsoft opened its fourth Indian cloud region in Hyderabad, and L&T won an order to build a 10,000-GPU NVIDIA B300 cluster in Chennai for U.S.-based Together AI. In September 2026, Tata’s HyperVault secured 264 acres in Hyderabad for an AI campus planned at up to 1GW.
The evidence now comes in three stages, and they should not be mixed up. Some capacity is live, like Microsoft’s new region and the government-reported 1,575MW of data centers already running. Some is ordered, like the Together AI cluster. Much of the headline gigawatt pipeline is still planned, from Google’s $15B Vizag hub to AirTrunk’s 5GW target.
Hyperscalers are not acting alone. OpenAI chose Tata, Meta chose Reliance, Google is working with AdaniConneX and Airtel, and L&T is now hosting a U.S. AI cloud, which shows India’s local conglomerates are becoming execution partners for AI infrastructure. Conglomerates pairing up with hyperscalers raise the question of who is funding whom, which we follow in our AI infrastructure funding report.
The user base explains the urgency, but enterprise adoption explains the monetization. OpenAI’s reported 100 million weekly ChatGPT users in India and the 300,000-plus Microsoft 365 Copilot seats at Infosys, TCS, and Wipro show demand from both consumers and workplaces.
“Cheap compute” is still not proven. India subsidizes GPU access for eligible users through IndiaAI, and 237 projects had used about 9.3 million subsidized GPU-hours by August 2026. But no like-for-like comparison shows that unsubsidized AI compute in India is cheaper than in the U.S., Singapore, or the Gulf.
Data control is a real driver, but it is not enough by itself. Payment data localization and data protection rules make local AI infrastructure more useful, and Microsoft now sells its Indian regions on data residency and sector compliance.
Power and land are now the main test. India’s appeal is the ability of large local groups to package land, power, renewables, connectivity, and permits. L&T’s Chennai site has 150 MVA of power ready for a first phase designed for 250MW.
India is starting to win training work, not just inference. The Together AI cluster is designed for training, fine-tuning, and large-scale inference for a global customer. Still, the deepest frontier training remains concentrated in the U.S., and India’s clearest edge today is deployment, sovereign compute, and enterprise AI.
Policy is acting like a confidence layer. IndiaAI’s shared pool passed 45,000 GPUs as of June 2026, and in October 2026 the government asked providers for 10,000 more GPUs on five-year assured use. That is a procurement request, not a purchase, but it gives private builders multi-year visibility.
The overall pattern is convergence: demand, data pressure, power-linked land, conglomerates, and policy are all becoming valuable at once, and some of that capacity is now running.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us
We keep an up-to-date report on AI infrastructure. Have a look ↓
Sample pages from an earlier edition. The report has been updated with new data since.
What are the actual signals that AI infrastructure is moving to India?
India’s AI infrastructure boom is now visible in live cloud regions, signed GPU orders, secured land, and government usage data.
Contrary to what many think, this is no longer just an “India has lots of AI users” story. The useful way to read the signals is by stage: what is running, what is ordered, and what is only planned.
The first proof is capacity that actually went live.
In August 2026, Microsoft made its India South Central region in Hyderabad generally available, its fourth cloud region in the country after Pune, Chennai, and Mumbai. Microsoft said its India Azure business had grown at a double-digit rate over the past two years, without giving the exact figure. In September 2026, it said its Indian regions were now AI-enabled and named Hyderabad a hub for Asia and the Global South. It sits under Microsoft’s $3B (January 2025) and $17.5B (December 2025) pledges, which are commitments, not cash spent.
The second proof is orders, not just announcements.
In August 2026, L&T’s Vyoma unit won an order to build a single cluster of 10,000 NVIDIA B300 GPUs at its Chennai campus for Together AI, a U.S. AI cloud company. L&T called it a “mega” order, a category it uses for ₹10,000–15,000 crore, and Reuters put that at roughly $1.05–1.57B.
The third proof is land and the GPU layer.
In September 2026, Tata’s HyperVault secured 264 acres in Hyderabad for an AI data-center campus planned at up to 1GW, built in phases as demand comes. Telangana’s investment agency put the planned investment at up to ₹70,000 crore. Earlier commitments still stand as plans: Google’s $15B Vizag AI hub broke ground in April 2026, OpenAI and Tata plan 100MW with a path to 1GW, Meta’s Reliance deal covers a 168MW Jamnagar facility, and AirTrunk targets more than 5GW by 2030. Yotta said it would deploy 20,736 NVIDIA Blackwell Ultra GPUs in Noida. On the public side, the government said in August 2026 that India’s shared IndiaAI compute pool had passed 45,000 GPUs as of June 2026, up from more than 38,000 in March 2026.
Did Microsoft, Google and Meta just see a huge user base for AI in India?
Yes, and the user base is still the first reason India became impossible to ignore.
OpenAI said India had around 100M weekly ChatGPT users in early 2026, making it one of its largest markets globally. A country with 100M weekly users of one AI product is no longer a “future market.” It already generates the kind of traffic that makes latency, reliability, pricing, and local infrastructure matter.
The same pattern shows up in enterprise AI. Infosys, TCS, and Wipro each scaled Microsoft 365 Copilot to more than 100,000 employees, reaching more than 300,000 seats across three Indian IT companies in under six months. Microsoft’s Hyderabad region launched in August 2026 with named early users including Adani Group, Bajaj Finserv, HDFC Bank, and PB Pay. That is large Indian companies asking for local AI capacity, not pilots.
Meta’s wording also matters. Its India deal with Reliance was framed around bringing AI-enabled infrastructure closer to one of its largest communities. The logic is not “India is cheap.” It is more direct: the users are already there, and the compute needs to move closer to them.
One caution. Google’s Vizag plan and Meta’s Reliance deal were both announced before mid-2026, and neither company has since published a new user figure tied to fresh data-center spending. So the user base explains why these bets were made, but the newest proof of demand comes from Microsoft’s live region and its enterprise customers.
These pages are excerpts from our report. See the rest ↓
Sample pages from an earlier edition. The report has been updated with new data since.
Is India attractive because AI compute is cheaper there?
Partly, for some users, but “cheap India” is still too lazy as an explanation, and it is not proven.
What is clear is that India is trying to make compute cheaper through structure, not wages. IndiaAI says eligible users can access GPUs at up to 40% lower cost, and the government said in August 2026 that 237 projects had used about 9.3 million subsidized GPU-hours. Yotta is building a pay-per-use sovereign AI cloud around more than 20,000 Blackwell GPUs. Tata, Reliance, and L&T are trying to put infrastructure, power, engineering, and enterprise demand under one roof. A pay-per-use sovereign AI cloud is a different way to sell compute; we compare which business models win in our report on AI infrastructure.
What is not clear is the price outside the subsidy. There is no public, like-for-like comparison showing that an unsubsidized GPU-hour, a delivered megawatt, or a full training run costs less in India than in the U.S., Singapore, or the Gulf. Subsidized access for startups and researchers is real; a national cost advantage is a claim nobody has shown with matched data.
That matters because AI infrastructure is constrained by GPUs, electricity, cooling, land, fiber, financing, and delivery timelines. Lower salaries do not solve those bottlenecks. What counts is whether India can coordinate enough of the physical stack to make AI capacity available at a price that startups, enterprises, and public-sector users will pay. If local compute gets cheaper in practice, it can unlock more Indian AI usage, which then justifies more infrastructure. That loop is the bet, not a settled fact.
Is it because India wants more control over its data?
Yes, but this is more of a pressure point than the full explanation.
India already has clear examples of data staying local. The Reserve Bank of India requires payment system data to be stored only in India. The Digital Personal Data Protection Act also makes data governance a much bigger compliance topic for companies operating in the country. For banks, payments, public services, telecom, healthcare, and large enterprises, local cloud and AI infrastructure reduces friction.
Companies now sell this openly. In September 2026, Microsoft pitched its AI-enabled Indian regions on local data residency and sector-specific compliance for frontier workloads. OpenAI and Tata’s planned sovereign compute capacity is built on the same idea. That said, there is no rule that all AI data must legally stay in India; the pressure comes from specific sectors and from procurement habits.
As companies use AI across customer service, payments, HR, and internal knowledge bases, more sensitive data touches the compute layer. If that compute sits abroad, legal, latency, and audit questions get harder.
Still, data control alone would not explain $15B, $17.5B, or 5GW-scale plans. Many countries want more control over their data without attracting hyperscaler-scale AI infrastructure. India has data-control pressure plus enough demand to monetize the local buildout.
Does India have enough power for its gigawatt AI campuses?
Not yet for all of them, and that is now the real test of the India story.
The recent announcements are full of power language. Google’s Vizag AI hub is described as gigawatt-scale and backed by clean energy. AirTrunk plans more than 5GW of digital infrastructure by 2030. HyperVault’s Hyderabad campus is planned at up to 1GW. Meta’s Reliance deal is tied to a 168MW Jamnagar facility, while Meta is also expanding renewable capacity in India with CleanMax to more than 900MW.
The gap between those plans and what is switched on is large. The government said in August 2026 that India had about 1,575MW of data-center capacity in total, across all data centers, not only AI. Industry forecasts cited by Adani and Jabil in October 2026 put India at 5–8GW by 2030. L&T’s Chennai site shows what a real step looks like: 150 MVA of power ready for a first phase designed for 250MW. A 1GW plan is power-plant scale, and most of these gigawatts are designs, not delivered power.
India has been adding renewable capacity fast. FY26 saw record renewable additions, with solar doing most of the work. That does not solve the grid problem. It lets India tell hyperscalers a more credible story: large AI sites can be paired with large clean-power procurement, especially when groups like Reliance, Adani-linked infrastructure, Tata, and L&T are involved.
This is where India’s conglomerates become an advantage. The winning pitch is “we can assemble land, power, connectivity, construction, permits, and operations into one buildable package.” For AI infrastructure, that package is often the difference between an announcement and a campus that actually turns on. Our judgment: power is the bottleneck to watch, and sites with power already in place, like L&T’s Chennai campus, are likely to move first.
Prefer verified numbers to headlines? Our report is here ↓
Sample pages from an earlier edition. The report has been updated with new data since.
Is India becoming a serious data-center market anyway?
Yes, India is becoming a serious data-center market, and that makes the AI wave easier to absorb.
India’s data center market was already scaling before the newest AI deals. JLL put India’s data center inventory at 1,123MW of IT load in H1 2025, up 48%. CBRE said India crossed around 1,700MW in 2025 after adding 440MW of supply, and projected about 500MW of new supply in 2026, roughly 30% year-on-year growth. CBRE also expects capacity to exceed 3GW by 2028. Adding 440MW of supply in a single year is a big move, and we cover the past year in our AI infrastructure market report.
The official number is a bit lower. In August 2026, the government put India’s data-center capacity at about 1,575MW, up from 375MW in 2020, a more than four-fold rise; the gap with CBRE likely reflects different definitions and dates. Either way, this is total data-center capacity, not installed high-end AI GPU power.
That base matters because hyperscalers prefer markets where the ecosystem already exists. You need experienced operators, fiber routes, power procurement teams, cooling vendors, construction capacity, and local permitting knowledge.
India is now in a more credible zone. It is still far behind the U.S. in AI infrastructure depth, but it is no longer starting from zero. Microsoft bringing a fourth region online in August 2026 shows that the market can absorb new hyperscale capacity, not just announce it.
Is this really about Indian conglomerates?
Yes, and this may be the least obvious reason global AI companies trust India now.
The recent deals are not just “foreign tech company enters India.” They are “foreign tech company enters India through a local infrastructure machine.” OpenAI chose Tata. Meta chose Reliance. Google is working with AdaniConneX and Airtel. Together AI chose L&T’s Vyoma to host its 10,000-GPU cluster in Chennai. Yotta, part of the Hiranandani group ecosystem, is building one of Asia’s largest NVIDIA Blackwell deployments.
The role of these groups is also widening. Tata’s HyperVault secured its own Hyderabad land in September 2026 and is pitching it to global AI companies. In October 2026, Adani Enterprises and Jabil signed a non-binding MoU for a proposed 50:50 joint venture to make AI racks, servers, networking, power, and liquid-cooling equipment in India. That is a plan, not a factory, but it shows conglomerates moving from land and power into hardware.
AI companies do not want to spend years learning every local bottleneck themselves. They want partners that can handle land, power, data centers, connectivity, government relationships, construction, and enterprise distribution.
So global AI companies seem to trust India because it has local actors that make AI infrastructure feel executable. India is not only offering demand. It is offering counterparties.
Want to know who is really ahead? Our report is here ↓
Sample pages from an earlier edition. The report has been updated with new data since.
Can India win global AI training work, like L&T’s Together AI deal?
It has started to, but India is not yet a frontier training hub.
The L&T order is the clearest evidence so far. In August 2026, Together AI, a U.S. AI cloud company, picked Chennai for a single 10,000-GPU NVIDIA B300 cluster that L&T says will support training, fine-tuning, and large-scale inference. HyperVault is also marketing its Hyderabad campus to global frontier-AI companies. Neither the full Together AI cluster nor any 1GW training campus is running yet.
That distinction matters. Frontier model training still clusters around places with the deepest chip access, mature power contracts, specialized engineering talent, and existing hyperscale AI campuses. The U.S. keeps a massive lead there. India’s immediate advantage is different: a giant local user base, a huge enterprise-services sector, multilingual demand, government use cases, and a need for local AI capacity.
Our verdict: India has now won an order for training-capable capacity for a foreign customer, at a scale it did not have before. But it is not winning the frontier-training race, and its strongest position remains deployment, where infrastructure can become just as strategic once usage explodes.
Is government policy actually changing company behavior?
Yes, but policy is acting more like a confidence layer than the main engine.
The IndiaAI Mission gives the market a clear signal: India wants domestic compute capacity, not just AI apps running on foreign infrastructure. The government said in August 2026 that the shared compute pool had passed 45,000 GPUs as of June 2026, and that 237 projects had used about 9.3 million subsidized GPU-hours. In October 2026, it issued a request for 10,000 more GPUs on five-year assured use, Moneycontrol reported. That is a procurement request, not a completed purchase, and plans for a domestic inference GPU and an AI rules consultation are still proposals.
That is not enough by itself to explain Microsoft, Google, Meta, or AirTrunk. Their investments are much larger than the public program. But policy reduces uncertainty. It tells investors that AI infrastructure has government support, that domestic compute is politically desirable, and that public-sector demand can be multi-year, as the five-year GPU request shows. Government support lowers the risk for investors, and we look at where the openings are in our report covering AI infrastructure.
Policy also makes India more legible as an AI infrastructure market. There is data protection law, public compute, AI missions, state-level data center incentives, and national messaging around AI capacity. For capital-intensive infrastructure, that visibility matters.
So private demand is pulling the market, while policy is making the pull safer to act on.
We made a report out of all this. Take a look ↓
Sample pages from an earlier edition. The report has been updated with new data since.
So why is India attracting more AI infrastructure lately?
India is attracting more AI infrastructure because global AI companies see something specific there: huge live usage, rising enterprise adoption, local data pressure, power-linked sites, and domestic partners that can actually build.
The real shift is that some of it is now running. Microsoft opened its fourth Indian region in August 2026. L&T has an order to build a 10,000-GPU cluster for Together AI in Chennai. Tata’s HyperVault has its Hyderabad land. Subsidized public compute is being used, not just counted. ChatGPT has massive weekly usage, and Indian IT firms are rolling out Copilot at scale.
The tension is the gap between plans and power. India has about 1,575MW of data-center capacity today, while the announced pipeline runs to several gigawatts. Google’s $15B hub, AirTrunk’s 5GW, and the 1GW campuses are still plans. And the claim that India offers cheaper compute holds only for subsidized users so far.
Put together, India is not attracting AI infrastructure because of one magic advantage. Several constraints in the global AI race are converging there: users need lower latency, enterprises need local deployment, regulators want more control, hyperscalers need new regions, and AI labs need partners that can deliver megawatts. Our judgment is that India has moved from promise to early execution, and the next proof will be how many of those gigawatts get grid power.
| Criteria | Answer | Comment |
|---|---|---|
| Is India already a massive AI user market? | Yes | ChatGPT’s 100M weekly users and 300,000-plus Copilot seats at Indian IT firms turn the market into current infrastructure demand. |
| Is local AI compute becoming cheaper? | Only for subsidized users | IndiaAI subsidies cut access costs for eligible users; no like-for-like data shows cheaper unsubsidized compute. |
| Does India want more control over data? | Yes | Payment data localization, data protection rules, and data-residency pitches make local infrastructure more useful for regulated workloads. |
| Is power a real advantage? | Partly | Groups can bundle land, power, and renewables, but most gigawatt plans are designs, not delivered power. |
| Is the data-center market already scaling? | Yes | The government reports about 1,575MW in 2026 versus 375MW in 2020; CBRE put 2025 at about 1.7GW. |
| Are conglomerates the hidden unlock? | Yes | Tata, Reliance, L&T, Adani, and Airtel host, build, and now plan to manufacture AI infrastructure. |
| Is India becoming a frontier training hub? | Not yet | The Together AI order will bring training-capable capacity, but India is stronger in inference, enterprise AI, and sovereign compute. |
| Is policy making investors more confident? | Yes | A 45,000-plus GPU public pool, real subsidized usage, and a five-year GPU request give builders visibility. |
If this is useful, the full report goes much further. Have a look ↓
Sample pages from an earlier edition. The report has been updated with new data since.
Our methodology
This analysis tests why India is attracting more AI infrastructure lately based on the evidence available today. We compare the headline explanation of “India has a big AI market” with concrete signals across user demand, enterprise adoption, compute supply, data control, power availability, data-center capacity, domestic infrastructure partners, training versus deployment, and policy support.
We treated the question as an evidence problem rather than a narrative one. The reason AI infrastructure is moving toward India is not obvious if we look only at broad assumptions about market size, low costs, regulation, or government ambition, so we broke the question into separate analytical dimensions and tested each one against recent signals.
We prioritized fresh and concrete signals: live cloud regions, signed GPU orders, secured land, announced megawatts, user numbers, enterprise rollouts, public compute usage, official policy updates, and market-capacity figures from institutional sources.
We separated three stages that are often mixed up: capacity that is operating, capacity that is ordered or under construction, and capacity that is only planned. Pledges, MoUs, procurement requests, and design capacity are labeled as such and never counted as delivered infrastructure.
When we refer to India’s AI infrastructure buildout, we mean the physical and operational stack needed to run AI workloads locally: data centers, GPUs, cloud regions, power procurement, fiber connectivity, sovereign compute, AI-ready campuses, and domestic infrastructure partnerships.
We separated frontier training from AI deployment because they are not the same infrastructure question. India can become strategically important for inference, enterprise AI, sovereign compute, and local deployment even if the deepest frontier model training remains concentrated elsewhere for now.
We treated government policy as a confidence layer rather than the main demand engine. IndiaAI, public compute, data governance, and state-level data-center incentives matter because they reduce uncertainty, but the largest private investments appear to be driven by demand, execution capacity, and infrastructure economics.
We used direct company and government sources where available, and tier-1 institutional sources for data-center capacity and renewable-energy context. Where official and industry capacity figures differ, we show both.
Key sources used for this analysis include: L&T on its 10,000-GPU NVIDIA B300 AI factory order for Together AI, W.Media on Microsoft’s India South Central region going live in Hyderabad, W.Media on HyperVault securing 264 acres for its Hyderabad AI campus, The Tribune on government data for India’s 1,575MW data-center capacity and IndiaAI usage, Adani on its MoU with Jabil for a data-center hardware joint venture, Microsoft on its $17.5B India investment over 2026–2029, Google on its $15B AI hub in Visakhapatnam, AirTrunk on its $30B and 5GW India plan, OpenAI on India usage and its Tata infrastructure partnership, Meta on its first AI-enabled data center deal in India with Reliance, Yotta on its 20,736 NVIDIA Blackwell Ultra GPU deployment, PIB on IndiaAI compute cost support, Microsoft on Infosys, TCS, and Wipro scaling Microsoft 365 Copilot to more than 300,000 employees, RBI on payment system data localization, India Code on the Digital Personal Data Protection Act, 2023, JLL on India data-center inventory and H1 2025 growth, CBRE on India’s data-center market outlook, and MNRE on renewable-energy capacity data. We also used Microsoft’s September 2026 note on its AI-enabled India regions, Reuters on the value range of the L&T order, Invest Telangana on HyperVault’s planned investment, and Moneycontrol on the October 2026 request for 10,000 GPUs.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us