SemiAnalysis: AI’s biggest scientific payoff may come from engineering improvements, not new discoveries

In the interview, a SemiAnalysis representative pushed back on the idea that AI investment in materials science would quickly run out of ideas and be constrained by the capital available to test them. He said science involves both discovery and engineering: most spending goes toward the engineering needed to bring discoveries to market. As an example, he described Periodic Labs as focusing mainly on increasing the pace of materials engineering.

Small improvements can still carry substantial economic value, he argued: a 5% increase in battery capacity, a more efficient semiconductor deposition process or better manufacturing yields. He said yields for some CPU substrates are around 40–60%, and invited listeners to consider the value of raising them to 95%. He also said Intel generates about 100 terabytes of data for each wafer but does not use most of it. In his view, AI, simulation and data analysis could accelerate engineering loops and improve yields; this was a prospective opportunity, not a report of an achieved result.

As another example, he said Nvidia takes 2.5 years to go from design concept to shipping a chip in volume production, and suggested that faster simulations and engineering could potentially cut that to a year. That was a hypothetical scenario, not a claim that the production cycle has already changed.