AI Growth Outpaces Semiconductor Manufacturing Capacity, McKinsey Executive Warns
At SEMICON West, McKinsey senior partner Bill Wiseman said AI demand is rising much faster than the semiconductor supply chain can expand. He identified token-processing power use and slow fab construction as key constraints, noting U.S. fabs can take about twice as long as those in Taiwan to reach first wafer. He also said grid connections now average 61 months and urged the industry to plan at wartime scale.
At SEMICON West’s SEMI Market Symposium, McKinsey’s Bill Wiseman described a mismatch: AI usage is climbing rapidly while the industry’s capacity additions follow slower, incremental paths. He noted inference represents about 60–70% of tokens processed in AI data centers, and agentic tasks can require roughly 100 times the compute of chat.
He framed energy per token as a key efficiency benchmark. NVIDIA’s progression has moved from 3.2 joules per token on Hopper toward 0.5 on Rubin, though Wiseman sees about 0.1 as the goal. He also contrasted fab timelines—about 19 months from permit to first wafer in Taiwan versus 38 in the U.S.—and cited 61-month average grid connections.
If AI demand keeps outpacing fab and grid expansion, access to advanced AI services could become more constrained or costly, affecting businesses, researchers, and consumers who rely on cloud tools. Longer U.S. fab timelines and multiyear grid connections may shape where compute capacity grows and how quickly. Energy efficiency gains per token may help, but local power systems and semiconductor workers could face pressure to adapt.