Can India build a cloud ecosystem to match 12 GW of data centre capacity projected by 2035?
India is entering a defining phase in its digital infrastructure journey. The country generates nearly 20% of the world’s data, creating a growing need for the computing infrastructure required to store, process, and derive value from it. NITI Aayog projects that India’s data centre capacity could expand from around 1.4 GW today to 10–12 GW by 2035, with the share of GPU-enabled capacity rising from 3–4% to 15–20% as AI workloads accelerate.
The opportunity extends beyond infrastructure, with India’s technology-services sector projected to grow from around $265 billion to $750–850 billion by 2035. Yet the scale of this ambition will depend not simply on how much capacity India builds, but on how effectively that infrastructure is translated into accessible, efficient and commercially viable compute.
Capacity is only the first layer
India’s data centre growth is being driven by cloud adoption, digital services, data localisation and the rise of AI. Installed capacity has increased from around 375 MW in 2020 to approximately 1,500 MW in 2025. At the same time, India has been expanding access to GPU infrastructure through its national AI compute initiatives.
This signals a shift from a data-storage infrastructure story towards a compute infrastructure story. Traditional enterprise workloads are largely CPU-oriented. AI workloads require GPUs, high-bandwidth networking, larger memory pools and faster storage. A data centre designed for conventional workloads cannot automatically deliver efficient AI computing at scale.
What will matter is not only how much capacity India builds, but also the kind of computing infrastructure that capacity can support.
India needs a diverse compute stack
AI workloads have very different requirements. Training a large language model, running inference, developing a digital twin, or undertaking advanced visualisation can require entirely different infrastructure.
The evolution of GPU technology itself illustrates how quickly these requirements are changing. NVIDIA’s B200 and newer B300 GPUs are designed for increasingly demanding AI training and inference workloads. The B200 offers 180 GB of HBM3e memory and up to 8 TB/s of memory bandwidth per GPU, while the B300 increases memory to 288 GB, with bandwidth of up to 8 TB/s. These advances enable higher-density deployments, but they also raise the infrastructure requirements around power, cooling, networking and storage.
For India, access to these accelerators will be as important as the capacity to house them. The global AI hardware supply chain remains complex, spanning advanced semiconductor manufacturing, high-bandwidth memory and specialised packaging. As demand for the latest accelerators grows, GPU availability can become a bottleneck in its own right, affecting deployment timelines and the ability of cloud providers to respond quickly to demand. Infrastructure planning will therefore need to account not only for space and power, but also for hardware procurement and the pace at which new GPU generations can be deployed.
This makes a diverse compute stack important. A model may need high-end infrastructure during training and a more cost-efficient configuration during inference. A manufacturer may need professional GPU infrastructure rather than a large AI-training cluster. India’s cloud ecosystem should therefore focus on matching infrastructure to workload while maintaining flexibility across GPU generations and configurations.
Cloud can turn compute into an operating resource
The economics of AI make accessibility particularly important. An eight-GPU server represents a substantial investment even before power, cooling, networking and specialist talent are considered. For an organisation with consistently high utilisation, ownership can make sense. For a startup experimenting with a model or an enterprise with fluctuating demand, it can create unnecessary capital exposure.
Cloud infrastructure changes this equation by allowing organisations to consume compute when needed and scale it as workloads evolve.
India is already moving in this direction through the national AI compute programme, which provides eligible users access to GPU resources based on defined compute requirements.
The next step is to make this model of flexible compute available more broadly, from developers and startups to enterprises and research institutions.
AI-ready infrastructure requires better engineering
The move towards AI also changes the physical design of data centres. High-density GPU deployments place greater demands on electricity, thermal management, networking and storage. A powerful accelerator is useful only when surrounding infrastructure can move data efficiently and maintain performance.
Utilisation therefore becomes a critical metric. An expensive GPU sitting idle represents underused capital and energy. Intelligent scheduling, workload orchestration and automation can improve utilisation and increase the useful compute generated from every deployed accelerator.
This is where cloud capability becomes distinct from physical infrastructure. Adding another data centre hall increases capacity. Adding systems that dynamically allocate and optimise GPUs creates usable computing capacity.
Data localisation must evolve into data capability
Data localisation has been an important driver of India’s data centre expansion. It is also increasingly connected to low-latency applications and real-time digital services.
The next phase should go beyond where data is stored and focus on where it can be processed, developed and deployed.
For sectors such as BFSI, healthcare and government, the requirement is increasingly to keep sensitive data within appropriate jurisdictions while accessing the compute needed to train, fine-tune and deploy AI systems.
An India-hosted cloud environment can support this model. A financial institution can develop AI applications around sensitive datasets without unnecessarily moving them overseas. An Indian-language AI developer can train models on domestic datasets while retaining greater control over data and infrastructure.
Data residency becomes strategically stronger when combined with domestic compute, storage, networking, and development capabilities.
Cloud economics will shape AI adoption
Cloud adoption will ultimately depend on economics as much as availability.
The lowest hourly GPU price does not necessarily deliver the lowest overall cost. A more powerful accelerator may complete a workload faster, require fewer nodes, or reduce training time.
The more useful metric is therefore the cost of achieving an outcome. For a generative AI application, this could be the cost per million tokens. For a bank, it could be the cost per transaction analysed for fraud. For a manufacturer, it could be the cost of completing a simulation.
This is changing how organisations should evaluate infrastructure. Performance, utilisation, power consumption and workload duration need to be considered alongside the listed price of compute.
Power will become a cloud strategy issue
The expansion of data centres also creates an energy challenge. NITI Aayog expects data centre electricity demand to become increasingly significant as IT loads expand, with AI contributing to that growth. High-density GPU clusters intensify the challenge because the latest accelerators demand significantly more power and increasingly sophisticated cooling.
India will therefore need to treat power availability, energy efficiency, fibre connectivity and geographic resilience as interconnected requirements when developing new data centre clusters.
Established markets such as Mumbai and Bengaluru will remain important, but a more distributed infrastructure footprint can improve resilience and bring compute closer to users.
The ecosystem must extend beyond hyperscalers
India’s cloud ambition ultimately depends on the breadth of its ecosystem.
Data centre operators provide the physical foundation. Cloud providers make infrastructure accessible. Hardware companies supply specialised accelerators, software companies build platforms around them, and startups, researchers and enterprises convert computing into applications.
The potential use cases span healthcare, manufacturing, financial services, agriculture, climate modelling and Indian-language AI. Medical imaging can benefit from accelerated computing, manufacturers can use digital twins, financial institutions can deploy real-time analytics, and researchers can scale compute-intensive work without building dedicated infrastructure.
The objective should therefore not be simply to increase the number of data centres. It should be to ensure that these facilities support a wider range of innovation.
12 GW should be the foundation, not the finish line
India’s projected 10–12 GW of data centre capacity by 2035 represents a major infrastructure milestone. But the more meaningful measure of success will be what organisations can build on top of it.
A strong cloud ecosystem will allow startups to access advanced GPUs without heavy upfront investment, enable researchers to scale compute-intensive projects, help enterprises deploy AI closer to their data and give regulated industries greater control over sensitive workloads.
The physical infrastructure will provide the foundation. The cloud ecosystem will determine its economic and technological value.
India’s opportunity is therefore larger than becoming a major location for data centres. The opportunity is to build an environment where compute is accessible, infrastructure is matched intelligently to workloads, and the next generation of AI applications can be developed and deployed at scale from India.
