KKR's Ming Lu Says AI Spending Keeps Outrunning Any Payoff, and History Says That's Normal

Lu points to a longer-running structural problem that predates AI entirely. Looking across S&P 500 companies by sector and by decade, he finds that 73 percent of the time, overhead costs, selling, general and administrative expenses, rise even as revenue shrinks. He bundles that corporate overhead together with government spending and compliance costs into what he calls a Combined Overhead Ratio, which he estimates crossed 47 percent of U.S. GDP around the year 2000. In his telling, GDP growth hasn't cleared 2.5 percent since, and no amount of new software fixes a cost structure problem that large on its own.

A faster engine bolted onto the same old process

The core of Lu's argument is that most companies are buying AI the way they'd buy a tool, not the way they'd rebuild a factory. He describes the typical deployment as dropping a faster engine into a horse-drawn carriage: the new capability sits on top of workflows, approval chains and org charts that were never redesigned around it, so the horse, not the engine, still sets the pace. Getting the economic payoff, in his view, requires rebuilding the carriage, not just swapping the engine, which is a slower and far less visible project than buying the software in the first place.

Why marketers and media buyers should read this as a timeline, not a verdict

For an audience that sells and buys media against AI-driven efficiency claims, Lu's numbers are a useful check on how those claims get pitched. Vendor decks tend to describe AI's return on investment as a near-term, almost automatic outcome. Lu's data says the honest version is closer to a multi-decade adoption curve, gated less by model capability than by how much of an organization is actually willing to rewire itself around the tool. Deep penetration inside individual companies and broad adoption across entire sectors both have to happen before the growth numbers move, and Lu's own estimate is that this is a process measured in decades, not product cycles.