The semiconductor industry has two problems to solve to keep pace with AI, Bill Wiseman, senior partner at McKinsey & Company, told the SEMI Market Symposium at SEMICON West: control the power it takes to process tokens, and build faster. “Demand’s on an exponential,” he said. “Everything that we have built in this industry is all linear.” Inference already accounts for roughly 60% to 70% of tokens processed in AI data centers, and agentic workloads run about 100 times more compute-intensive than chat. Wiseman called joules per token “the new Moore’s Law”: NVIDIA’s roadmap has driven it from 3.2 J/token on Hopper toward 0.5 on Rubin, but he put the target closer to 0.1. Meanwhile, supply measures such as leading-edge fab capacity and DRAM bit shipments grow far more slowly than demand, and a fab that takes about 19 months from permit to first wafer in Taiwan takes about 38 in the U.S. Power is an even longer pole, with grid connections now averaging 61 months. Drawing on Henry Kaiser, Adm. Hyman Rickover and Gen. George Marshall, Wiseman urged the industry to start thinking at wartime scale, beginning with fab productivity. Click here to read the full article in the SEMICON West Show Daily. The post McKinsey’s Wiseman: AI Demand Is Exponential, but the Industry Still Builds Linearly appeared first on Semiconductor Digest .