Across boardrooms in Silicon Valley and server halls from Virginia to Visakhapatnam, one reality is becoming impossible to ignore: artificial intelligence is growing faster than the physical systems that can handle it. The chips, power grids, cooling systems, permits, fibre—none can match the velocity of AI adoption.
Amazon, Alphabet, Meta, and Microsoft confirmed in their latest earnings calls what analysts had long suspected: capital expenditure is rising in lockstep with AI-driven revenue, but the gap between demand and available capacity is only widening.
According to an analysis by Data Center Knowledge, hyperscalers are now effectively selling into the capacity they are still building. Microsoft’s Azure grew 40% year-on-year, even as its capital expenditure declined sequentially from $37.5 billion to $31.9 billion. Meanwhile, Google Cloud’s backlog has swelled to more than $460 billion, reflecting a pipeline of demand that has yet to be delivered.
And nowhere is this tension more acutely felt than in India, which is racing to become a global AI hub but also grappling with the limits of its own digital backbone. According to the Institute for Energy Economics and Financial Analysis, the country’s data centre capacity is expected to grow exponentially to 9 GW by 2030, consuming about 3% of total electricity.
“We are in a phase where infrastructure is the limiter, and it is not close. Demand is outstripping available capacity across every hyperscaler. The constraint has already shifted decisively. It is no longer about silicon; it is about power, cooling, and permitting,” Steven Dickens, CEO and Principal Analyst at HyperFRAME Research, tells AIM.
Though there has been considerable activity when it comes to AI infrastructure deals in India. Google recently broke ground to establish a 1 GW data centre in Andhra Pradesh. Adani Group committed $100 billion to renewable-powered data centres, while Reliance upped that number with a $110 billion pledge to develop AI infrastructure.
However, data centres take time to build. Google’s facility is expected to be fully operational by July 2028.
“New data centre capacity takes 18–36 months to bring online at scale. Demand is accelerating on a six- to 12-month cycle. Power generation is harder to spin up than chips. That is the binding constraint. Google Cloud grew 63% year-on-year, with its backlog expanding to more than $460 billion. That is Google taking real share. And yet even they cannot build fast enough to satisfy what enterprises want today,” Dickens comments.
The ERP Tipping Point
If the current wave of AI infrastructure spending is driven largely by foundational models and cloud platforms, analysts warn that the next wave could be far larger and harder to accommodate. The trigger, according to Prabhu Pandian, Chief AI Architect of informatics company Excelra, will be enterprise software.
“We are only at the tip of the iceberg when it comes to AI in the enterprise. The durability of this infrastructure build-out will depend on demand and supply growing in harmony. A lot will depend on how easily the ERP vendors build agents. If Oracle and SAP can replicate the kind of adoption curve that Salesforce achieved with CRM, then demand will outstrip supply in ways that make the current crunch look manageable,” he notes.
Enterprise AI workloads span ERP systems, agentic AI, fraud detection, and operational automation.
In India, enterprise AI infrastructure buildout is still in early stages. Yotta Data Services recently partnered with IBM to support enterprises and government organisations in meeting data residency mandates and operating AI workloads within sovereign boundaries. Bhavish Aggarwal’s AI startup Krutrim also shifted focus to offering AI infrastructure, including cloud-based compute and tools.
“Enterprise adoption is still early, but it is beginning to accelerate sharply. The infrastructure being built today may not be sufficient for the workloads that arrive in the next 24–36 months,” Pandian says.
Racing Against Ambitions
India’s AI infrastructure buildout is also focused on access to chips.
Larsen & Toubro has partnered with NVIDIA to establish a sovereign AI factory as part of the IndiaAI Mission, with a 30 MW GPU compute cluster in Chennai and a new 40 MW AI-ready facility in Mumbai. Yotta Data Services has committed to investing more than $2 billion to deploy over 20,000 liquid-cooled NVIDIA Blackwell Ultra GPUs, creating one of Asia’s largest AI computing superclusters at its Noida campus.
However, timely access to infrastructure is now critical to sustaining the momentum across the ecosystem.
“My target for this year is almost 80,000 GPUs. My capacity to hold the GPUs is already there. Between Mumbai and Delhi, the buildings are ready,” Sunil Gupta, Co-founder, CEO, and Managing Director of Yotta Data Services, earlier told AIM.
India is in the middle of its largest-ever data centre expansion cycle, but the gap between announced and operational capacity remains significant.
Power: The Hidden Bottleneck
If there is one thread that runs through every conversation about AI infrastructure, it is power, or the lack of it.
In India, this challenge is particularly acute. The country’s five major data centre hubs, Mumbai, Hyderabad, Delhi-NCR, Bengaluru, and Chennai, are already straining local power networks. Water consumption, another critical constraint given the cooling demands of high-density GPU infrastructure, is projected to exceed 358 billion litres annually by 2030, per an India AI Impact Summit projection.
Patrick Moorhead, Founder and Chief Executive of Moor Insights & Strategy, who tracks AI infrastructure investment across global markets, believes companies that define AI infrastructure over the next five years will be those that solve the power problem.
“Whether you are building in Ohio or Hyderabad, the fundamental constraint is the same: you cannot run a 100-megawatt GPU cluster without a dependable, affordable, and increasingly green source of electricity. The hyperscalers know this. The question is whether the grid operators, the regulators, and the capital markets can keep pace,” he posted on LinkedIn.
This is where the government’s 500 GW renewable energy target fits well. “The two ambitions must be built together, or neither will fully succeed,” he noted.
India’s data centre expansion is increasingly a two-speed story. A small set of AI-first mega-projects, GPU superclusters, sovereign AI factories, and hyperscale campuses are still years from full delivery. A broader base of phased hyperscale campuses is being implemented more gradually, constrained by power availability, partner capabilities, and regulatory timelines.
What stands out in this cycle is that nearly half of all announced Indian capacity now markets itself as AI-focused, but these labels obscure large differences in technical maturity. True AI infrastructure, with power redundancy, thermal efficiency, GPU lifecycle management, and sub-10-millisecond latency, is far more demanding.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.