By Mukesh Bansal, Founder and CEO, NuPlay AI
India’s AI race is increasingly becoming an infrastructure race.
India’s data-centre sector is entering a period of rapid capacity expansion, with AI emerging as an important driver of new, high-density infrastructure. KPMG estimates that India’s data-centre market could grow from approximately $1.7 billion in FY26 to nearly $6.8 billion by FY30. At the same time, CBRE says India had around 1.75 GW of operational data-centre capacity by the end of H1 2026, with AI adoption driving higher-density deployments and specialised compute requirements.
The implication is bigger than the construction of more server rooms.
AI workloads require a different physical infrastructure. Training large models demands enormous amounts of compute. Inference—the process of running those models for users—requires speed, reliability and increasingly low latency. Both can require significantly higher power and cooling density than many traditional data-centre environments were designed for.
That makes India’s AI opportunity partly a race to build the physical systems underneath the software.
The companies—and ultimately the countries—that can secure reliable power, predictable compute, high-speed connectivity and trusted infrastructure will have a structural advantage. Building the best AI application will not be enough if the infrastructure underneath it cannot scale.
The demand has changed
For years, Indian data centres were primarily built around conventional cloud and enterprise workloads: hosting applications, databases, websites and storing data.
AI is now adding a fundamentally different source of demand.
Training requires large pools of compute running intensively for extended periods. Inference creates another challenge: the system has to respond quickly and reliably whenever a user or enterprise calls the model.
These workloads are more power-intensive, more sensitive to latency and increasingly dependent on high-density computing.
This is why India’s AI story is an infrastructure story first.
Applications are the visible layer. Power, compute, cooling, connectivity and jurisdiction are the layers that determine whether those applications can run reliably and profitably at scale.
Five bottlenecks, none of them glamorous
Power: AI-ready data centres need substantially more power density than many traditional facilities. A data centre cannot simply install more GPUs once demand arrives. Grid connectivity, substation capacity, backup systems and round-the-clock reliability have to exist first.
This makes power availability a strategic constraint. Every uptime commitment an AI company makes to an enterprise customer ultimately depends on infrastructure that can keep the underlying systems running.
Compute : GPUs and other specialised accelerators are becoming critical inputs for AI. India has recognised this through initiatives such as the IndiaAI Mission, which is building access to shared compute capacity. But the challenge is not simply how many GPUs exist on paper. It is whether a startup, researcher or enterprise can access the right amount of compute when it needs it, at a predictable price and with sufficient reliability. In AI, unavailable compute is effectively unavailable capacity.
Cooling: Higher-density computing also means more heat. As AI infrastructure moves towards increasingly dense GPU deployments, data centres are adopting more advanced air-cooling, liquid-cooling and hybrid approaches. This creates a new engineering challenge: how to remove heat efficiently without allowing energy and water requirements to undermine the economics of the facility.
Cooling may not appear in an AI product’s user interface, but it can materially influence the cost and scalability of the infrastructure underneath it.
Connectivity: Compute is only useful if data can reach it and responses can return quickly. High-speed fibre, network capacity and low latency become particularly important for inference workloads, where delays can directly affect the user experience. CBRE’s latest assessment of India’s data-centre market identifies network connectivity, power availability, renewable-energy sourcing and market access among the factors influencing where new facilities are developed.
This makes geography increasingly important. The ideal data-centre location is not simply where land or electricity is cheapest. It has to balance power, fibre connectivity, cooling, land availability and proximity to customers.
Sovereign compute: For banks, insurers, healthcare providers and government organisations, where an AI workload runs is increasingly a question of compliance, security and trust.
India’s data-protection framework, sector-specific requirements and enterprise policies can create incentives to keep sensitive workloads within Indian jurisdiction. KPMG identifies data localisation, AI workloads and regulatory developments among the structural forces shaping domestic data-centre demand.
Sovereign compute is therefore not simply about keeping data inside the country. It is about giving regulated organisations the control, security and assurance required to deploy AI at scale.
Infrastructure shows up in unit economics
These constraints do not stay inside the data centre.
They eventually show up in the economics of every AI product built on top of it.
For companies deploying enterprise AI agents, the product is not just the model. It is reliability: an agent that responds in seconds, remains available through multi-step workflows and costs little enough per interaction for the business case to work at scale.
Every layer beneath the application eventually surfaces in the P&L.
Latency can be influenced by where compute and users are located. Margins are affected by GPU pricing and electricity costs. Reliability depends on redundancy, network architecture and the resilience of the underlying facility.
This is why AI infrastructure is becoming a business issue, not merely an engineering issue.
A better model may win a benchmark. But an AI product has to deliver predictable performance at a sustainable cost to win customers.
What needs to happen next
Plan power alongside capacity: Data-centre clusters, grid upgrades, transmission infrastructure and renewable-energy procurement need to be planned together rather than on separate timelines. Capacity that cannot be powered reliably is capacity on paper only.
Make compute access predictable: India does not just need more GPUs. It needs transparent pricing, shorter provisioning times and dependable access for startups, enterprises and researchers.
Invest in unglamorous engineering: Cooling efficiency, power usage effectiveness, network latency, redundancy and operational resilience are where long-term competitiveness is built. These are not features that make headlines, but they determine whether an AI business can scale.
Build infrastructure around geography: The next generation of data-centre clusters will need to consider power availability, fibre connectivity, land, cooling requirements and proximity to users together. There is no single “best” location if one of those constraints becomes a bottleneck.
Make sovereign infrastructure competitive: Enterprises will adopt sovereign infrastructure at scale only if it can match global alternatives on performance, reliability and price—not merely on compliance.
The real race
Every technology cycle has a visible layer and a foundational one.
The visible layer gets the attention. The foundation determines how far the ecosystem can go.
India’s digital public infrastructure offers a useful precedent. UPI and Aadhaar created common digital rails on which thousands of applications and businesses could build.
AI will require a similar foundation, although the infrastructure will look very different.
The winning ecosystem will be the one where a builder in Bengaluru, Delhi, Hyderabad—or a smaller Indian city—can access reliable, affordable and secure compute without having to become an infrastructure expert first.
The ambition should not simply be to build more data centres or deploy more GPUs. It should be to make compute as accessible as a utility: available when needed, priced predictably and reliable enough that the infrastructure disappears from the user’s mind.
That is the boring problem.
It is also the one worth solving.