AI "Hyperbole and Silly Numbers"? (www.theregister.com)

Bain & Company’s Technology Report 2025 warns the AI infrastructure buildout driving the current boom may be economically unsustainable: to support an estimated extra 100 GW of U.S. datacenter capacity by 2030 would require roughly $500 billion per year in capex, which Bain says implies the AI sector must reach about $2 trillion in annual revenue to fund it “profitably.” Even if firms moved all on‑prem IT spend to the cloud and plowed AI productivity savings into datacenter builds, Bain calculates the market would still be roughly $800 billion short. The warning comes amid headline projects from OpenAI and Microsoft and a surge in hyperscale spending (Synergy Research reports $127B in hyperscale capex in Q2 2025, up 72% year over year), but also growing skepticism: a recent MIT-linked analysis and other industry voices suggest many AI projects deliver little ROI and some capex estimates are “hyperbole.” Technically, Bain flags four hard constraints: power supply (new generation/transmission can take 4+ years), construction capacity, GPUs/compute enablers, and ancillary gear like switchgear and cooling. London Economics adds that chip supply could be a bottleneck, while analysts put realistic annual AI datacenter capex nearer to $300B. The upshot for AI/ML: unless there are major efficiency or architectural breakthroughs, or substantial public funding/policy shifts, the current hype-driven expansion may hit funding, supply-chain and energy limits — concentrating growth among the biggest players and raising questions about long-term ROI and sector sustainability.