A better way of thinking about the AI bubble (techcrunch.com)

Think of the “AI bubble” not as an apocalypse but as a series of very large bets that can easily outpace real demand or the physical systems that make AI possible. Recent reporting highlights the scale: an Oracle-linked New Mexico data‑center campus pulled as much as $18 billion in credit, Oracle has committed $300 billion in cloud services to OpenAI, and together with SoftBank is tied to a $500 billion “Stargate” infrastructure push—while Meta has pledged roughly $600 billion over three years. Those headline numbers matter because datacenters take years to build, and between design and go‑live the software, hardware and energy landscapes can change dramatically. The core technical risk is a timeline and systems mismatch: AI software and model capabilities evolve at breakneck speed, but power grids, supply chains, semiconductor roadmaps and physical data‑center shells move slowly. Surveys (McKinsey) show most firms use AI in pockets but few at scale, so demand growth is uncertain. Even where demand exists, operators face real constraints—Satya Nadella warned the bottleneck is “warm shells to plug into,” and many facilities sit idle because local power capacity can’t support the latest GPUs. For practitioners and infra planners this means heightened risk of stranded assets, bigger returns to energy efficiency, modular/shorter‑build deployments, and a renewed focus on aligning procurement, power strategy and realistic adoption timelines.