AI & ML
India’s data centre capacity is expected to more than double by 2027 — and AI workloads are already making that target feel conservative.
That’s not hype. It’s a supply-demand gap that operators, policymakers, and enterprises are scrambling to close right now. AI data centres in India aren’t just bigger versions of what exists today; they’re a fundamentally different class of infrastructure, demanding more power, smarter cooling, scarce hardware, and a policy environment that hasn’t fully caught up.
Four forces are shaping this race. Understanding all of them — and which carries the most weight — is the only way to predict where India’s AI infrastructure story goes next.
Why AI Data Centres in India Need to Be Built Differently
The gap between a standard data centre rack and an AI-ready one is enormous — and most people underestimate it.
An AI data centre is a facility purpose-built to handle power densities, thermal loads, and compute architectures that standard enterprise infrastructure wasn’t designed for. A typical server rack draws somewhere between 5 and 10 kilowatts. A rack dense with NVIDIA H100 GPUs? You’re looking at 40 to 100 kilowatts, sometimes more. That’s not a modest upgrade. It’s a complete rethink of how you design the floor, route the power, and remove the heat.
This matters because the applications driving demand — AI-powered threat detection systems, AI-powered software testing platforms, large language model inference at scale — don’t tolerate the kind of infrastructure compromises that a standard enterprise workload might shrug off.
The GPU shortage is real, and it’s biting India harder than most.
Global allocation of NVIDIA H100 and H200 chips has been dominated by US hyperscalers. Indian operators — Yotta, NxtGen, CtrlS — are competing for the same constrained supply. The IndiaAI Mission’s push for sovereign GPU procurement is a direct response to this, though industry watchers are split on how much that changes the picture near-term. Nobody really knows yet, though early signs suggest domestic demand will outpace sovereign supply for at least another 18 months.
There’s also a geography angle that’s easy to miss. AI inference for real-time applications — fraud detection, live threat analysis, customer-facing recommendation engines — has strict latency requirements. That’s quietly pushing operators to look beyond Mumbai and Chennai toward tier-2 cities. Hyderabad, Pune, and Ahmedabad are showing up in site selection conversations in ways they weren’t two years ago.
Powering the AI Boom: India’s Energy Equation
Here’s where the infrastructure story gets genuinely complicated.
A single GPU cluster running a large model training job can consume as much electricity as several thousand homes — for weeks at a stretch. Contrast that with a conventional batch processing workload, and you start to understand why power is the single biggest constraint facing AI data centres in India right now.
India’s grid isn’t uniformly ready for this. Maharashtra, Tamil Nadu, and Telangana have the strongest existing infrastructure, which is why they’ve attracted the bulk of investment so far. But Rajasthan and Gujarat are positioning aggressively, partly because they sit on enormous renewable energy potential.
The operators moving fastest on this aren’t waiting for grid upgrades. They’re locking in captive renewable energy — signing long-term Power Purchase Agreements with solar and wind developers to guarantee 24/7 clean power. Adani Green’s partnerships with data centre operators are a visible example of this trend. It’s not cheap, but it de-risks the long-term capacity question in a way that grid dependence simply doesn’t.
The tension, though, is real. Data centres draw enormous loads. State governments that want the investment also have to balance it against residential and industrial grid demand. That negotiation is happening right now in boardrooms and state secretariats, and how it resolves will shape the map of AI infrastructure in India for the next decade.
Keeping AI Clusters Cool: The Technology Shaping Facility Design
Cooling is the problem most people outside the industry don’t think about until it’s too late.
Air cooling — the default for most facilities built before 2020 — starts breaking down around 30 kilowatts per rack. AI racks blow past that threshold easily. When you’re pushing 60 or 80 kilowatts through a rack, air cooling doesn’t just become inefficient; it becomes physically inadequate.
The two technologies gaining real traction are Direct Liquid Cooling (DLC) and immersion cooling. DLC runs coolant directly to the chip level; immersion submerges servers in a dielectric fluid that absorbs heat far more effectively than air. Both approaches dramatically improve Power Usage Effectiveness (PUE) — a standard efficiency metric where lower is better.
| Cooling Method | Max Power Density | Typical PUE | Best Suited For |
|---|---|---|---|
| Air Cooling | ~30 kW/rack | 1.5–1.8 | Standard enterprise workloads |
| Direct Liquid Cooling (DLC) | 60–100 kW/rack | 1.1–1.3 | High-density GPU clusters |
| Immersion Cooling | 100+ kW/rack | 1.02–1.1 | Maximum-density AI training |
STT GDC and Nxtra by Airtel are among the Indian operators already deploying liquid cooling at scale. But retrofitting existing facilities is expensive and technically complex. Most operators building new capacity today are designing for liquid cooling from the ground up — which is the right call, even if it increases upfront capital cost.
India adds a layer of difficulty that European or Canadian operators don’t face: ambient temperature.
Running a cooling system in Chennai or Pune when outdoor temperatures regularly hit 38–40°C is a harder engineering problem than doing the same in Frankfurt. Water scarcity compounds this — data centres near water-stressed cities can’t rely on evaporative cooling approaches that work well elsewhere. Closed-loop systems and dry coolers are becoming more common as a result.
Policy as Infrastructure: How Regulation Will Decide Which States Lead
I’d argue that policy is the most underrated variable in this race — more than hardware or even power.
India granted data centres infrastructure status several years ago, which helped unlock financing. But a coherent national framework with single-window clearance, standardised approvals, and clear data localisation rules is still a work in progress. That ambiguity has a direct cost: foreign investors making long-term capital allocation decisions need regulatory certainty, and India’s patchwork of state-level policies creates friction.
At the state level, the competition is intense:
| State | Primary Advantage | Position |
|---|---|---|
| Maharashtra | Subsea cable connectivity, established capacity | Leading |
| Telangana | Dedicated data centre zones, hyperscaler MoUs | Leading |
| Tamil Nadu | Port access, relatively stable power supply | Leading |
| Uttar Pradesh | Lower land costs, scale potential | Emerging |
| Gujarat | Renewable energy corridors, policy incentives | Emerging |
The Digital Personal Data Protection Act (DPDPA) adds another layer. Data localisation requirements affect where operators can physically store certain categories of data, shaping facility decisions in ways that go beyond simple cost calculations.
And then there’s security. As AI cybersecurity threats in 2026 grow more sophisticated — with adversarial attacks targeting AI systems themselves — facility design is starting to incorporate physical and logical security requirements together. Operators hosting AI workloads can’t treat physical and software security as separate problems anymore.
Software-layer vulnerabilities, including PHP security vulnerabilities that security teams have flagged through 2026, affect the applications running inside these facilities. A thorough website security checklist isn’t purely a developer concern; it’s part of an operator’s broader risk posture.
The Road Ahead: Three Scenarios for India’s AI Infrastructure
The honest answer is that nobody knows exactly how this plays out. But the shape of the uncertainty is worth understanding.
If things go well: Renewable energy capacity scales, liquid cooling expertise deepens in the domestic market, GPU supply gradually normalises, and the national policy framework gets the consolidation it needs. In that scenario, India captures a meaningful share of Asia-Pacific AI infrastructure investment — probably 15–20% of regional capacity additions by 2028.
If the bottlenecks persist: Grid upgrades stall, cooling talent remains scarce, and policy fragmentation keeps deterring long-horizon capital. Operators redirect AI-critical workloads to Singapore, Malaysia, or Japan, and India ends up building mostly conventional capacity while the AI layer sits offshore.
The wildcard is domestic demand. The IndiaAI Mission, enterprise adoption of AI-powered software testing and AI-powered threat detection platforms, and a startup ecosystem generating real AI applications — these are demand-side forces that could accelerate supply regardless of how the policy and infrastructure sides move. India’s domestic market is large enough to pull investment even when the external environment is difficult.
The AI software development trends of 2026 point squarely at India-first applications in fintech, healthcare, and government services. That demand won’t wait for perfect infrastructure. And that pressure, more than any policy announcement, might be what ultimately drives the investment.
Frequently Asked Questions
What makes an AI data centre different from a regular one?
An AI data centre is a facility purpose-built to handle power densities of 40 to 100+ kilowatts per rack, compared to 5 to 10 kilowatts in conventional setups. This requires specialised cooling systems (often liquid-based), reinforced power infrastructure, and hardware like GPU clusters that standard facilities weren’t designed to support. The difference isn’t cosmetic; it’s architectural.
Which Indian cities are best positioned for AI data centre investment?
Mumbai, Chennai, and Hyderabad currently lead based on existing capacity, connectivity, and policy support. But Pune, Ahmedabad, and Noida are gaining ground as alternatives, driven by lower land costs, improving power infrastructure, and proximity to major enterprise demand centres. State-level policy incentives are shifting this map in real time, so the picture 18 months from now may look quite different.
How do AI cybersecurity threats affect data centre design decisions?
AI cybersecurity threats are pushing facility operators to merge physical and digital security design from the outset, rather than treating them as separate functions. Facilities hosting AI workloads need to account for threats that target AI systems directly — model poisoning, adversarial inputs, inference attacks — alongside conventional perimeter security. Operators are also being asked to maintain software-layer hygiene as part of their overall risk management posture, which includes staying current on application-level vulnerabilities affecting the workloads they host.
Is India’s power grid ready for large-scale AI workloads?
Not uniformly. Some states — Maharashtra, Tamil Nadu, Telangana — have stronger grid infrastructure, but even they face reliability challenges at the scale AI clusters demand. Most serious operators are supplementing or replacing grid power with captive renewable sources, using long-term Power Purchase Agreements to guarantee clean, stable electricity. The grid will improve, but operators aren’t waiting for it.
The Stakes Are High Enough to Get This Right
India’s position in the global AI infrastructure race is genuinely uncertain — and that’s what makes the next 24 months so consequential.
The demand signal is strong. The policy intent is there. The talent to build and operate these facilities exists. What’s still being resolved is whether the power grid, the cooling expertise, the GPU supply chains, and the regulatory environment can align fast enough to capture the investment that’s currently making decisions.
If you’re evaluating data centre infrastructure — as an enterprise, an investor, or a policymaker — track four things: compute access, power reliability, cooling capability, and regulatory clarity. Where those align, the facilities will get built. Where they don’t, the capital will go elsewhere — and it will be harder to bring back than most people expect.
