Ming Lu: AI bet at 2% of U.S. GDP won't pay off soon

Lu grounds his skepticism in a pattern he calls the Innovation Paradox: five decades of digital innovation achieved broad consumer adoption yet came with a decline in aggregate productivity growth. Technology diffusion follows a double K-shaped pattern, he writes—consumers adopt almost immediately, while enterprises move at a glacial pace, and within enterprises a handful of superstar firms embrace new tools quickly while the rest struggle to make use of them. The structure of the U.S. economy compounds this, he argues, having shifted from vertically-linked industrial sectors, where innovations could travel through supply chains, to horizontally spread service sectors that resist transformation because they depend on human judgment. These "stagnant" sectors, which the economist William Baumol described, now account for roughly half of GDP in advanced economies, according to Lu.

Lu also points to a financing loop that ties the economy's growth drivers together. Hyperscalers will spend $800 billion this year on chips, data centers, and the power to run them, propping up GDP growth. Investors are bidding up AI stocks, fattening the portfolios of the richest 20%, whose spending boosts consumption. As the buildout outgrows hyperscalers' cash flow, new investment is increasingly financed by debt and off-balance-sheet partnerships—the dynamic the BIS flags as "circular financing." Lu adds that concentration of capital in AI-linked companies starves the rest of the economy of investment needed to be AI-ready, and that firms are saddled with administrative and regulatory compliance costs he estimates now strangle innovation.

He concedes AI could still defy the paradox, since unlike previous digital waves it is a cognitive technology that can reason and make judgments—precisely what's needed to unlock productivity in labor-intensive Baumol sectors. But that would require deep adoption across a broad spectrum of the economy, with firms allocating capital to workers and workflows, not just chips and data centers, he writes. History offers a possible silver lining if the bubble bursts: canals in the 1830s, railroads in the 1880s, electrification in the 1920s, and telecoms and the internet in the late 1990s each ended in painful collapse, yet the physical infrastructure created in those speculative booms laid the foundations for decades of later prosperity. Data centers, semiconductors, and AI models may prove equally long-lasting, Lu suggests, but that process will take decades, not months or years—and in the short term, the more concentrated the bet, the more severe the potential reckoning.