LLMs Are Dying - Long Live SLMs!
The AI revolution is real. But the trillion-dollar infrastructure boom supporting it is in a financial bubble approaching its peak.
In his excellent October 12 Wall Street Journal column, “How the AI Boom Ends,” Andy Kessler argues that AI technology is advancing rapidly, but its economics are changing even faster.
Four forces could bring the boom crashing down in 2027.
1. Small Language Models (SLMs) are coming. Smaller, increasingly capable models running locally on smartphones and computers will eliminate much of the need for expensive cloud processing. In my book Three Black Swans (www.rcwoodbridge.com), a humanoid robot with an SLM brain plays a main character
2. The 2025 One Big Beautiful Bill has accelerated tax deductions. Immediate expensing of costs juice after-tax results. But deductions taken today cannot be taken again tomorrow, reducing future tax benefits.
3. Data-center overinvestment is accelerating. Technology giants are committing hundreds of billions to LLM facilities, GPUs, NPUs, and power infrastructure whose economic lives may prove shorter than anticipated. Will dwindling supplies of Helium hurt chip production?
4. Political incentives are weakening. Next year the current President will be a lame duck, constitutionally barred from another term. Without pressure of reelection, his administration will have less incentive to pursue politically difficult measures to sustain economic growth.
History offers warnings: railroads, AM radio, and the internet all experienced investment booms followed by painful corrections.
My prediction: By 2017, collapsing processing costs, smaller models, fading tax advantages, and weakening political incentives will expose massive overcapacity and trigger a sharp correction in AI infrastructure valuations.
AI will flourish. But today's infrastructure investors probably not.
Comments