India’s AI infrastructure race is entering a phase in which adding compute may be easier than sustaining it. Enterprises are moving from AI experimentation towards production workloads, driving demand for increasingly dense GPU systems. But every increase in compute density brings a corresponding infrastructure question. How much power can a rack draw, how much heat can it generate, and how efficiently can that heat be removed without creating another resource constraint?
The investment appetite is unlikely to be the immediate problem. Lenovo’s CIO Playbook 2026 says 59% of organisations in India are already piloting or systematically adopting AI, while 99% plan to increase AI investments over the next 12 months. India is also projected to record average year-on-year AI budget growth of 19%, the highest among the Asia-Pacific markets covered by the study.
The harder question is what happens when those investments translate into high-density production infrastructure.
Srinivas Rao, Managing Director of Infrastructure Solutions Group at Lenovo India, argues that the traditional approach of treating servers, power and cooling as separate layers is becoming increasingly difficult to sustain.
“If you talk about a big GPU platform, there is a different LPM which is required,” says Rao, referring to liquid flow requirements in cooling systems. “It depends upon the kind of system which you need to put in place.”
That dependence on system configuration is becoming more important as AI platforms evolve. The thermal requirement is not simply a function of how many GPUs are installed. It is also influenced by the workload and how intensively the systems are being used.
“When you are running at a 50% efficiency, it doesn’t really produce so much of heat,” says Rao. “But when you are running at a 90% kind of efficiency, it produces huge amount of heat.”
That changes the cooling equation. Direct liquid cooling is emerging as one response to the rising thermal density of AI infrastructure. Liquid can remove heat from high-density systems more efficiently than conventional air-based approaches, reducing the energy required for cooling. Rao said the technology can offer significant savings in cooling power compared with air cooling.
But this is where the sustainability argument becomes more complicated for India. Liquid cooling reduces one infrastructure burden while potentially bringing another resource question into sharper focus. Water availability is uneven across the country, and the sustainability of a cooling system cannot be judged simply by describing it as liquid-based or by looking at its instantaneous consumption.
Rao does not offer a single figure for water consumption. Instead, he says the requirement varies according to the platform, workload and cooling architecture. “It is completely variable depending upon the load and all those things,” he shares.
That qualification is important. It means that broad comparisons between air and liquid cooling can conceal differences in system design and operating conditions. The efficiency of the cooling technology is therefore only part of the equation.
Rao points to closed-loop systems as offering better efficiency than approaches that require greater liquid consumption over time. He also distinguishes direct liquid cooling from immersion cooling, saying immersion brings its own maintenance challenges and that direct liquid cooling has emerged as the more practical approach.
The consequence is that the location of AI infrastructure could become a more strategic decision. “Choosing the right locations with the right environmental ambience plays a very, very important role,” says Rao.
That consideration becomes more significant as the size of India’s data centre industry expands. Lenovo expects India’s data centre capacity to more than quadruple by 2030.
For AI infrastructure, however, capacity alone will not tell the full story. The economics of a facility will increasingly be shaped by its thermal environment, access to resources and the cooling architecture required to maintain its compute density.
This also means India has less room for a one-size-fits-all approach. A cooling design that performs well in one geography may have a very different resource or operating profile in another. The country’s data centre expansion will therefore increasingly have to account for environmental conditions alongside power, land and connectivity.
Rao says Lenovo’s experience across different geographies provides operational lessons for these deployments, particularly because systems have been implemented in both very hot and very cold environments. “Those learnings are like a goldmine, actually, because there will be operational challenges,” he notes.
The larger issue for India is that AI infrastructure is moving beyond a capacity race. The press note itself points to the growing importance of rack-scale systems and liquid cooling as chip generations push rack density and power requirements higher. That makes cooling a strategic infrastructure decision rather than an engineering afterthought.
India can add GPUs. It can add data centre capacity. But as AI workloads become denser and more continuous, the ability to operate that compute efficiently will increasingly depend on decisions made around the rack, the cooling system and, ultimately, the location of the data centre.