There’s a funny thing about India’s AI revolution. It has a lot to do with… err… rotation. Of cooling fans. Of power meters. And of the hype cycle itself.

You see, the recent revolution began with one headline across the board. “India reaches $20B in new cumulative AI investment commitments”. Peel that back and you will find that most of those billions aren’t going to young startups training foundation models in co-working spaces. They are going to those who pour concrete and install chillers.

Google wants to build a $15 billion AI data hub (its largest outside the US). AWS will spend $13 billion soon. Microsoft promises $3 billion. TCS has plans for another $5–7 billion. The government, too, has thrown in $1.2 billion for GPU farms and LLMs.

In other words, most of India’s AI “investment” is going into racks, not reasoning. And that makes Sify Infinit Spaces, the company behind India’s first data-centre IPO, a rather honest mascot for this moment.

Sify doesn’t build chips or train LLMs; it rents out space to those who do. Colocation, as the industry calls it, is essentially high-security, industrial-grade real estate for servers. Clients bring their own machines; Sify supplies the power, cooling, and connectivity.

The company is the data-centre arm of Sify Technologies—the same Sify that gave India dial-up internet in the 1990s. Now, it operates 14 data centres across six cities with 188 megawatts of built IT-power capacity—roughly 15% of India’s total, according to its IPO documents.  

But India’s “total” is modest. The US and China measure data-centre capacity in gigawatts, not megawatts. Though India generates nearly a fifth of global data traffic, it holds only 3% of capacity, according to Cushman & Wakefield. The gap between bytes and buildings is vast.

Source: Kotak Mutual Fund

View in Browser

Elsewhere, the market has raced ahead. The global colocation industry grew about 10% a year for the past five years and could speed up to 15% as AI workloads soar, says JLL. Capacity worldwide has more than doubled—from 20 GW in 2019 to 42 GW in 2024—and could reach 65 GW by 2027.

The problem is that AI is an energy glutton. Training large models is 3–5X more power-intensive than regular computing.