The capital markets are pricing AI as a solved infrastructure problem while the actual builders are discovering it isn't. Ford admitted this week that swapping experience for algorithms produced worse cars, forcing the company to rehire the engineers it thought it could replace. Meanwhile, Wall Street is hunting for the next Nvidia in Micron, Samsung and SK Hynix are committing $600 billion to memory expansion, and the US power sector is in a $200 billion M&A sprint to feed data center electricity demands. This is the sound of capital flowing toward the commodity layer of AI, not the intelligence layer. The BIS is warning that exuberance could trigger a lengthy bust, but that warning assumes the bottleneck is compute or funding. The actual constraint appears to be execution: Flexion Robotics, built by ex-Nvidia engineers, is training robots to do real work. Liquid AI shipped a 230-million-parameter model that runs on a phone. DeepSeek open-sourced a speculative decoding framework that cuts inference time 57 to 85 percent. These are products solving actual problems, not frameworks solving theoretical ones. The gap between what investors think they're buying and what actually works is widening. Chips and power are being treated as the scarce resource when the scarcity is talent that knows how to build things that function in production. Ford's rehiring of gray beards is the canary. When the capital stops flowing to the infrastructure vendors and starts flowing to the people who can make the infrastructure useful, the hierarchy of returns shifts. We're not there yet, but the pressure is building.
Sloane Duvall