By Pankaj Singh | Head – Data Center & Telecom Infra Segment, Delta Electronics India
Google has broken ground on its gigawatt-scale AI hub in Visakhapatnam — its largest investment in India’s digital future, part of a $15 billion commitment through 2030. Reliance has announced a 1 GW data center campus in Jamnagar. Google and Microsoft together have pledged over $32 billion to India’s digital infrastructure in recent months. And the Union Budget 2026-27 just extended a tax holiday to 2047 for foreign cloud providers operating through Indian data centers. The message is unambiguous: India is being built into a global AI infrastructure hub, and it’s happening at an unprecedented speed.
But here’s what those headlines don’t fully capture: behind every gigawatt of announced capacity is a set of deeply unglamorous engineering problems that will determine whether these investments actually deliver. Ask any data center operator what their biggest headache is today and it’s not land or permits — it’s the GPU cluster pulling 80 kW per rack and climbing, with air cooling that was never designed to cope. That shift, from managing IT infrastructure to managing something closer to an industrial power plant, is what AI has done to our industry, faster than almost anyone anticipated.
Each technology wave has reshaped infrastructure priorities — virtualization changed how capacity was managed, cloud changed how it was scaled. AI is doing something different in kind, not just degree. The variables are familiar: power, cooling, compute. What’s shifted is how tightly interdependent they’ve become. A rack density decision is now simultaneously a power architecture decision and a cooling topology decision. There is no longer a clean sequence — everything has to be solved together, from day one.
The Numbers Don’t Lie
A “dense” rack used to mean 10 kW. Today, AI-native racks run at 50–100 kW as standard, and training clusters pull multi-megawatt loads within a single hall. That’s not an upgrade — it’s a complete category shift. And unlike traditional server workloads, which spike and idle, GPUs under AI training run at sustained, near-continuous utilisation for days. The facility never gets a break.
Zoom out and the scale is staggering. Global data center electricity demand is projected to exceed 945 TWh by 2030 — roughly what Japan consumes in a year. India’s own data center market, valued at $3.88 billion in 2025, is on course to nearly double to $7.92 billion by 2032. AI workloads in India typically demand 20–50 kW per rack today, and that ceiling keeps rising as GPU cloud services and large-model inference expand. Facilities designed for 10–15 kW per rack now face four to ten times that load per bay. Most weren’t built for this.
Power First — Everything Else Follows
The old design sequence was: floor space, then cooling, then power. For AI infrastructure, that order has to flip. Power is now the primary constraint — the variable that shapes every other decision downstream.
Legacy power systems were built for a different era. What AI-scale facilities need instead are modular, high-efficiency power architectures with dynamic load-balancing that responds to actual workload demands in real time. Every efficiency gain at this scale compounds — lower energy bills, less cooling load, a smaller carbon footprint. For India, this also means designing for local grid realities — integrating renewables, building in storage buffers, and ensuring resilience from the ground up, not as an afterthought.
Cooling Has to Follow the Heat, Not the Floor Plan
Beyond a certain density threshold, air cooling simply stops working. It’s not an engineering failure — it’s physics. And at the rack densities AI demands, no amount of additional air units will close that gap.
The mental model that needs to change is from cooling a room to cooling a rack. Liquid cooling — whether direct-to-chip, rear-door, or immersion — has moved from specialist territory to mainstream necessity. The next frontier is making cooling intelligent: systems that dynamically respond to workload intensity rather than running at fixed setpoints. For India’s greenfield sites, hybrid approaches that combine targeted liquid cooling with precision air offer a practical path forward, while keeping water consumption in check — a real concern in many of our urban markets.
Modularity Isn’t Optional Anymore
Here’s a tension that every operator in India is feeling right now: AI investment decisions move on 12-month cycles, but traditional data center construction takes 24–36 months. Something has to give. Modular design is the answer.
Rather than over-engineering an entire hall upfront, modular architectures deliver scalable power and cooling in 200–500 kW pods that can be deployed and expanded as AI cluster requirements actually grow. Factory-tested, prefabricated units cut on-site construction risk and compress commissioning timelines. At Delta, we’ve made modularity the default for any facility targeting 40 kW and above — power systems, cooling pods, and monitoring infrastructure are designed to interlock and scale together, so adding capacity doesn’t mean redesigning the facility.
Density and Sustainability Have to Go Together
We tend to hear people talk about AI density and sustainability as if they’re in tension. They don’t have to be — but only if we make the right choices now, at the design stage, not as retrofits later.
With global data center power demand heading past 945 TWh by 2030, every kilowatt of AI density needs to come with a PUE and WUE commitment attached. For AI-optimised facilities, PUE of 1.3 or better should be the baseline, not the aspiration. Waste-heat reuse, renewable-ready architecture, and intelligent load balancing aren’t premium features — they’re how you make high-density AI economically and environmentally viable over a 15-year asset life. India’s digital infrastructure is being built right now, at pace. The decisions made in the next few years will set the energy footprint of our AI economy for decades.
Building for What’s Next
AI isn’t just adding load to our existing infrastructure — it’s rewriting what a data center fundamentally is. The industry is moving from an IT asset to a power-centric, thermally constrained utility where every design choice has to start with density.
The facilities that shape India’s digital decade won’t be retrofitted warehouses. They’ll be platforms built from the first rack for the workloads of the next ten years — modular, efficient, thermally intelligent, and ready to scale. That’s the bar. And honestly, it’s an exciting one to build toward.
