For years, running AI on the cloud worked like living in a rented apartment. Someone else owned the building and handled the upkeep, and you paid for exactly what you used. No commitment, no maintenance, move out whenever you like.
That arrangement made sense when AI was a guest in most businesses. It looks different now that AI has become a permanent resident, and increasingly, one that wants a home of its own.
A couple of years ago, AI in most companies meant a handful of pilots. A chatbot here, a forecasting tool there, quietly running on someone else's servers, folded into a bill alongside email hosting and file storage. Nobody thought much about where the model actually sat.
That's no longer the picture. AI now sits inside core banking systems, hospital scheduling, factory lines and legal review, the parts of a business that can't afford to go down. Once a model is doing work that critical, the questions around it change. Where is the data processed? What happens if a vendor changes its terms mid-contract? Who else, in theory, has a line of sight into what's running?
Security is one answer to those questions, but only one. Cost predictability matters just as much once usage scales up, and so does a simple, growing discomfort with depending entirely on infrastructure a company doesn't control. Banks, hospitals, manufacturers and governments are arriving at the same instinct from different directions. The reasoning varies. The direction doesn't.
What's easy to miss is that "AI infrastructure" was never really one thing, and it's splitting further as adoption deepens.
Training still leans on massive, elastic capacity, the kind only a handful of hyperscale campuses can offer, because building a model from scratch is a short, enormous burst of compute.
Inference, running the model day to day once it's built, behaves more like a steady utility, and companies are increasingly moving it closer to home for speed and cost control.
Edge deployments are showing up inside factories, hospitals and retail stores, small local clusters built to sense and respond in real time without waiting on a distant data centre.
Sovereign and national clouds are emerging as their own category entirely, built specifically to keep a country's or a company's data inside a border, or a boardroom, that everyone agrees on.
The upshot is that the future isn't purely cloud or purely on-premises. It's a set of different builds for different jobs, each shaped by what it actually needs to do.
None of this is only a change in appetite. It's changing what gets built and where. What was once a small number of enormous, largely interchangeable data halls is giving way to smaller, purpose-built facilities located closer to where companies actually operate. Fewer decisions made from a handful of hyperscale hubs, more decisions made city by city, closer to where the demand actually sits.
None of this would matter much if owning your AI infrastructure meant settling for a weaker model. That's the part that's changed fastest.
Open-weight models, the kind companies can inspect, adapt and run without sending anything back to whoever built them, have closed much of the gap with the most advanced closed systems over the past year. For companies that wanted to own their AI infrastructure but couldn't justify building a competitive model from scratch, that excuse has largely disappeared.
Every one of these threads, ownership, cost, latency, use case, ends up pointing at the same physical outcome: more AI infrastructure built and held locally rather than rented from a distant hyperscaler.
We explored the technology side of this shift in our take on how AI is reshaping the real estate industry. What's changing now extends that story from the buyer's side of a property transaction to the infrastructure itself. Cities with a strong base of technology companies and Global Capability Centres, Chennai among them, are well placed for a wave of demand that looks less like a handful of giant campuses and more like a wider spread of smaller, well-located facilities.
AI infrastructure is being pulled out of a small number of distant, shared facilities and rebuilt closer to the businesses that depend on it.
For real estate, that's not a distant, futuristic idea. It's a present one. The next decade of AI will be shaped as much by where companies choose to build as by what their models can do, and that decision is already being made, city by city, plot by plot.
