Analysis
I've spent this issue writing up Microsoft's plan to triple its data-center footprint to 38 gigawatts, Nvidia's bet on 2 gigawatts of new Australian AI capacity, and a funding week where The Boring Company, Positron AI and Cognition all re-rated at multiples that assume the AI capex cycle keeps compounding indefinitely. Here's my actual read: nobody involved -- not the hyperscalers, not the chip makers, not the application-layer companies burning that compute -- has fully underwritten what happens if AI revenue growth decelerates even modestly before this capex comes fully online.
Microsoft's own admission is the tell: the company is turning away cloud and AI customers TODAY, at $175 billion a year in spending, and its answer is more capex, not better allocation of what it already has. That's rational if demand keeps outrunning supply indefinitely. It's a genuinely bad position if demand growth merely normalizes to a still-healthy-but-less-explosive rate, because gigawatts of committed data-center capacity don't un-commit the way a software company can cut a hiring plan.
“That's rational if demand keeps outrunning supply indefinitely.”
The application layer makes this worse, not better. Cognition, Harvey and Clay are all re-rating on usage and ARR growth that assumes the compute underneath stays cheap and available -- but every dollar Microsoft, Nvidia and the hyperscalers spend chasing 38 gigawatts is a dollar betting THOSE application companies' growth curves hold. If AI-application revenue growth cools before 2032, the infrastructure layer is left holding gigawatts of capacity built for demand that never fully showed up, and the application layer is left with compute costs that don't fall as fast as competition compresses their own pricing.
Room for disagreement: the counter-case is real and I take it seriously. Every one of the last three AI capex cycles -- the initial ChatGPT-driven buildout, the 2025 inference-cost wars, and this year's agentic-coding wave -- has been met with demand that arrived faster than infrastructure could scale, not slower. If that pattern holds a fourth time, Microsoft's 38-gigawatt bet looks conservative in hindsight, not reckless, and the hyperscalers currently rationing compute will look prescient rather than overextended. The honest answer is nobody, including me, actually knows which pattern wins this time -- but the size of this week's combined infrastructure and application-layer bets means the cost of being wrong just got a lot bigger than it was even six months ago.
