The analyst contends that AI models improve at software speed, with new generations becoming smarter and capable of distilling successors in weeks, unconstrained by physical limits. In contrast, leading-edge chips require hundreds to thousands of process steps and years to qualify and scale into mass production. This creates a widening gap between AI capabilities and the hardware needed to run them, leading to global bidding for un-fabbed compute.
The accelerating pace of AI model development is creating a structural demand-pull for compute, making physical silicon manufacturing the binding constraint on AI progress.