Nikhil Kamath has a new frame for the AI infrastructure boom, and it has nothing to do with square footage.
In a post on X, the Zerodha co-founder and investor argued that the economics of data centres have flipped. The building was once the asset. Now it is barely a rounding error next to the silicon inside it.
"Every prompt you send ends at a chip," Kamath wrote. "That chip costs more than the building and the land combined."
His numbers back this up. Of every Rs 100 spent on a large AI data centre, roughly Rs 40 goes to chips alone, more than the combined cost of the building, power, cooling, networking and land, according to the breakdown Kamath cited from Bernstein and Epoch AI estimates. Land accounts for barely Re 1 of that Rs 100.
ALSO READ: AI's Volatile Power Demand Is Damaging Its Own Data Centers
That shift, Kamath argued, explains why companies and countries are now making opposite bets on the same infrastructure.
For companies, the logic is about capital efficiency. Kamath's thread notes that global hyperscalers own only around 60 per cent of their data centre capacity, and even that, he said, is "only because renting ran out." The remaining 40 per cent is leased.
"Owning sinks it. Lease improves ROCE," Kamath wrote, framing the calculus in terms of return on capital employed, the metric that governs how listed companies are judged.
ALSO READ: Mumbai Ranks 15th, Hyderabad 27th In Global Data Centre Capacity: Knight Frank
For governments, the calculation runs the other way. Kamath pointed to India's AI programme, where more than 38,000 GPUs are being deployed under agreements that keep the hardware in private hands but the data within Indian jurisdiction.
"They need jurisdiction, not hardware," Kamath wrote, arguing that sovereignty concerns mean countries cannot simply lease their way to strategic control the way a company would.
Kamath's broader point is that the location and form of computing infrastructure are becoming less fixed. His thread cites data centres now running inside quantum-grade refrigeration units, aboard an orbital NVIDIA chip that has already trained an AI model, and on the ocean floor in China, cooled by seawater. A handful of startups, he noted, are experimenting with "biocomputers" built from living human brain cells.
"Not every task needs the same machine, or environment," Kamath wrote.
He closed the thread with the line that has driven much of the reaction: "Rent what depreciates. Own what endures." His closing question was directed squarely at the debate now playing out across boardrooms and finance ministries alike: "So who actually owns AI infrastructure, the companies renting it out, or the countries that can't afford not to control it?"
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.