The $800B AI Capex Race Who Wins When Big Tech Spends Saurabh Mukherjea x Ishaan Chitale

Description

In 2026, six American tech giants — Alphabet, Amazon, Anthropic, Apple, Meta, and Microsoft, are on track to spend over $800 billion on AI infrastructure. That sum is roughly a quarter of India's entire annual GDP poured into a single category of building.But what actually goes into these multi-billion dollar facilities, and who is capturing the true value?In this episode of the Biggest Questions in Finance, Ishaan Chitale and Saurabh Mukherjea dissect the unprecedented global AI build-out. They break down the staggering costs, the physical anatomy of modern data centres, the severe supply-chain chokepoints, and what this technological revolution means for India's economy and investors.--------------------------------------------------------------------💡Key Takeaways• The Anatomy of an AI Facility: An AI data centre relies on six highly constrained systems: compute (brain), power (heart), cooling, networking (nervous system), storage (memory), and the physical building.• Nvidia's Monopoly: Building a 1-gigawatt data centre costs approximately $50 billion. Out of this massive budget, roughly 60% to 65% ($30 to $35 billion) goes straight to Nvidia for GPUs, servers, and HBM memory.• The Cooling Crisis: Power density per rack has surged from 5–15 kilowatts to over 300 kilowatts for next-gen GPUs, making traditional air cooling obsolete and forcing expensive liquid cooling redesigns.• Soaring Construction Costs: The cost to physically build a data centre has skyrocketed nearly six-fold in just five years, jumping from $183 per square foot in 2020 to $1,033 by 2025. • India's Position: While India is securing mega-campuses like the Vizag hub, 80 to 90 percent of the high-value technology content is imported, threatening the current account deficit and raising sovereignty concerns. • The "Shovel-Seller" Strategy: To win in the AI gold rush, investors need to own the companies supplying the essential tools like American hyperscalers and Taiwanese/Korean chipmakers rather than just the companies building the models.-----