AI Infrastructure

Nvidia’s $164.5 Billion Backstop Is Demand Finance

RRahul RanjanOctober 27, 2026 7 min read

Featured essay

Environmental data make the trade-off starker. Delhi receives 66 out of 100 on the Data Center Environmental Impact Index, whereas Kochi gets 27.6 and Guwahati 27.9. Delhi's water table declines by 1-2 metres annually. India's major data-centre hubs are also the ones with the greatest environmental exposure. Heat adds to the cooling burden. Most of the operational data centres packed into India's major hubs already see temperatures above 35°C for more than 90 days a year, and by 2040 nearly 90 per cent may face similar heat stress. Water demand is severe: a 1 MW facility requires roughly 26 million litres for cooling, which runs directly into Delhi-NCR's water constraints. Fiscal sweeteners such as Haryana's 20-year electricity duty waiver or Uttar Pradesh's revised 2026 goal cannot manufacture water or firm power. They can only lift the financial return on a project that still has to get through the physical bottleneck.

4. The second-order risk is vendor-credit contagion and a shrinking electricity market

Nvidia's guarantees tie chip suppliers, cloud providers, and model labs into a single balance-sheet chain. A CIO entering a multi-year GPU capacity contract is therefore partly exposed to Nvidia's credit, not only the operator's. This matters to Indian founders because the local financing stack is lopsided: Indian financiers will lend against land and power assets but not against GPUs, so early overseas backers realise returns while Indian banks hold back, leaving domestic operators reliant on offshore capital cycles. When those cycles turn, the buildout does not decelerate uniformly. It halts wherever the weakest credit sits.

Concentration is an equally serious problem. About 70 per cent of India's operational data centres are in Mumbai, Delhi-NCR, Chennai, Bengaluru, Hyderabad, and Kolkata, placing resource strain on the same urban systems already facing heat and water stress. Private distribution licences for data-centre parks can shrink the public discoms' customer base and transfer fixed costs to the customers left behind; the public-interest question is how much smaller the electricity market can become. Policy should combine targeted fiscal, siting, and water-conservation conditions with market rules that deliver firm power and renewable supply where load actually appears. Without that, the buildout will stall on local bottlenecks well before it stalls on GPU supply. A GPU order amounts to a promise, while a node is a physical fact.