Crawford Del Prete: AI task gains not yet in the numbers

The conversation situates today's AI spending against the personal computer and internet eras, when productivity gains took years to appear. Del Prete recalled that it wasn't until 13 or 14 years after the PC was introduced that productivity gains started growing, during the connectivity period between 1996 and 2002. He contrasted that with the current buildout, where every GPU that gets put out today is lit up and at 100% capacity as soon as it comes online, unlike the dark fiber that sat unused after the dot-com boom.

Del Prete described a split among enterprise customers. Some quantitative-oriented firms say they cannot get enough compute capacity and will take all they can from OEMs like Lenovo, HPE, and Dell. Others are starting to manage token spend by matching models to workloads, for example using Opus 4.8 rather than the more token-intensive 5.0 on extra effort. He said he is not seeing many customers pull back because competitiveness demands AI investment.

The productivity J-curve, based on work by Brynjolfsson, Rock and Syverson, offers a framework for the delay. Individual productivity gains can coincide with negative cumulative cash flow as companies invest in new tools, skills, and data before redesigning workflows. Del Prete warned that organizations which simply hand out tools like Claude or Copilot will get frustrated when they don't see institutional change, and he suggested leading-edge firms are instead driving bottom-up projects and codifying them for reuse across the organization.