Are We in an A.I. Bubble? I Suspect So (gideons.substack.com)

The author argues we may be in an AI bubble: not because the underlying technology is worthless, but because key physical and economic constraints are already slowing the kind of accelerating returns that justified the recent investment frenzy. Concrete limits include electricity and grid capacity, a growing shortage of high-quality real-world training data (forcing new models to be trained on AI-generated data), and compute demand growing faster than Moore’s Law—creating acute pressure for chips and data centers. Existing AI systems also haven’t yet delivered broad, measurable productivity gains, so continued massive spending looks increasingly driven by sunk investments and chasing earlier timelines rather than by step-change breakthroughs. More worrying is the financial plumbing: large firms have circular commercial ties (e.g., OpenAI–Oracle–Nvidia) that inflate reported activity, and much data-center financing is done via asset-backed loans that are securitized and insured. If asset correlations spike during a downturn—like in the mortgage crisis—banks and insurers could face outsized losses. The practical policy takeaway: bursting the bubble directly is politically fraught, so monetary authorities should be prepared to cushion fallout if it pops, while financial regulators should scrutinize capital adequacy and risk models for banks/insurers exposed to AI infrastructure. Direct AI safety regulation during a bubble risks destabilizing the market and may be politically infeasible.