In October 2025, Google announced a $15 billion investment to build Asia’s first gigawatt-scale AI hub in Visakhapatnam, Andhra Pradesh — a landmark moment for India’s digital ambitions. The facility, when operational, will place Vizag alongside the world’s most important AI infrastructure nodes.

But a natural question arises: why does an AI data centre need a so much of power — Why can’t we just use normal office electricity?

The answer traces back to two concepts from computer chip design: Dennard Scaling and the Power Wall. You don’t need an engineering degree to understand them. And once you do, you’ll see why building energy infrastructure for data centres is one of the most important economic challenges of the 2020s — and why India’s proactive approach to meeting it is exactly the right strategy.

The Golden Era: When Chips Got Faster for Free

From roughly the 1970s to the early 2000s, the computer industry enjoyed an extraordinary free lunch. Every couple of years, engineers figured out how to make the tiny switches on a computer chip (called transistors) smaller. And here’s the magical part: when you made a transistor smaller, it didn’t just fit more of them on a chip — it also used less power per transistor.

This observation was formalised in 1974 by a researcher named Robert Dennard, and it became known as Dennard Scaling:

As transistors shrink, the voltage and current they need shrink proportionally. So even though you pack more transistors onto a chip, the total power consumption stays roughly the same.

Imagine you run a restaurant kitchen. Dennard Scaling is like discovering that every time you hire more cooks (transistors), each cook needs proportionally less food (power) to work. So your grocery bill stays flat even though you’re cooking twice as many dishes. More output, same energy bill.

For three decades, this was the engine of the computing revolution. Your laptop today is millions of times more powerful than the computers that sent astronauts to the moon, and for most of that journey, each leap in performance came without a proportional leap in electricity bills.

The Wall: When the Free Lunch Ended

Around 2005–2006, Dennard Scaling broke down. Transistors had become so tiny — just a few dozen atoms wide — that electrical current began to leak through the transistor walls, even when the transistor was supposed to be “off.” Think of it like a water tap you can never fully close: no matter how hard you turn the handle, it still drips.

This leakage meant that shrinking transistors no longer reduced their power consumption. Every additional transistor now added to the power bill. This breakdown is what engineers call the Power Wall.

Before the Power Wall, chip clock speeds doubled every few years — from 100 MHz to 200, to 500, to 1 GHz, to 3 GHz. Then the climb stopped. Today, twenty years later, most chips still run at 3–5 GHz. The wall held.

So What Did the Industry Do? It Went Wide.

When you can’t make a single engine faster, you bolt on more engines. Instead of one super-fast processor core, modern chips have thousands of cores running in parallel — this is what GPUs (Graphics Processing Units) do, and they are the workhorses of AI.

This shift solved the performance problem but made the power problem much worse. Since Dennard Scaling no longer helps, every additional chip is an additional power consumer.

Inside a 5 MW Data Centre: Where Does All the Power Go?

Let’s make this concrete with a 5 MW AI data centre — a modest facility by global standards. The workhorse GPU chip for AI today is NVIDIA’s H100, drawing about 700 watts each — roughly the power of a large microwave oven running continuously, 24/7/365.

Only about 50% of total facility power goes to compute. The rest supports cooling, storage, networking, and power conversion:

Number of H100 chips = 0.5 × 5,000,000W ÷ 700W ≈ 3,571 GPUs

A full quarter of all electricity — 1.25 MW — goes purely to cooling. That’s 1.25 million watts doing nothing computationally useful. It exists solely to keep the chips from overheating. This is the Power Wall made tangible: because Dennard Scaling broke down, every watt of useful computation generates waste heat that demands additional energy to manage.

The PUE Number: Measuring Overhead

Engineers use Power Usage Effectiveness (PUE) to measure data centre overhead. Indian data centres in 2024 range from PUE 1.3 to 1.6. At PUE 1.6, for every watt used by IT equipment, the facility consumes 1.6 watts total — the extra 0.6 watts goes to cooling, power conversion, and support systems.

India’s Proactive Response: Meeting the Power Challenge Head-On

What’s remarkable about India’s approach is that the Government has recognised this challenge early and is building a comprehensive energy ecosystem to support data centre growth — rather than waiting for a crisis to force action.

The Causal Chain: Why This All Matters

The Bottom Line

When you hear about data centres needing the power, it’s not waste or poor planning. It’s physics:

• Dennard Scaling — the principle that let chips get faster without more power — broke down around 2005.
• The Power Wall forced the industry to use thousands of parallel chips instead of faster individual ones.
• A modest 5 MW data centre packs 3,571 GPUs, each drawing 700 watts non-stop. Only half the electricity does computation; a quarter goes to cooling.

What matters is how a nation responds to this challenge. India’s integrated strategy — combining a ₹20,000 crore Nuclear Energy Mission with SMR development, 250+ GW of installed renewables, 43 GWh of VGF-supported battery storage, domestic manufacturing through Make in India, and grid modernisation — positions the country not merely to cope with AI’s energy demands but to lead the global transition to sustainable, AI-ready infrastructure.

The Google AI hub in Visakhapatnam — with its $15 billion investment and $2 billion dedicated to renewable energy — is a testament to this approach. It demonstrates that when policy, infrastructure, and private investment align, the immense power needs of modern AI become not a burden but a catalyst for national development.

The next time your AI assistant answers a question in two seconds, remember: somewhere, thousands of chips are running at full speed, backed by power plants, solar farms, battery banks, and soon, small modular reactors — an entire national energy ecosystem working in concert, all because a scaling law named after Robert Dennard stopped working twenty years ago. The physics created the challenge. India is building the answer.

Data centre power benchmarks from Takshashila Institution Report 2025-27.
Government of India policy data from Union Budget 2025–26, PIB, and Ministry of Power announcements.
Google AI Hub details from official Google Cloud press release, October 2025.

Disclaimer : This article is made using Claude AI.