Goldman Sachs strategists say the AI trade is climbing the stack, moving from chips to the software and services that put models to work inside businesses and everyday gadgets.
It’s not a retreat from silicon. Pete Callahan, the firm’s TMT specialist, notes that investors have settled on an outlook for hyperscaler capex through 2027, giving a year or so of forward visibility, while the picture past that stays fuzzy.
That uncertainty has squeezed semiconductor valuations, with investors arguing over whether the group is over-earning or under-earning, Callahan said. At the same time, excitement has shifted to what Goldman calls the inference economy, which includes security tools and data pipes that let businesses and consumers run AI. Ben Schneider, the firm’s chief U.S. equity strategist, puts it bluntly: the micro is driving the macro, and fresh product wins like Muse have rekindled interest after the first half was all about capex winners.
The inference bid is already showing up in prices elsewhere. Nebius rallied 9% to $254 on a small AI inference deal while trading at ~197x trailing P/E, showing inference pricing power and stretched valuations beyond Goldman's stack discussion. That kind of move illustrates the stack shift Goldman Sachs described, where deployment, not just silicon supply, commands the premium.
Schneider warns breadth is razor thin. The median S&P 500 stock sits more than 15% below its peak, and one breadth gauge is the tightest since the dot-com era. He says narrow breadth usually means a catch-up, not a catch-down, though timing and direction are hard to call. On earnings, sustainability is the focus; consensus expects roughly 30% growth, with a slowdown likely by 2027 as macro pressure mounts and capex growth eases, even if spending continues to rise.
Looking ahead to third-quarter earnings, Callahan says the two key tests are return on invested capital at the biggest hyperscalers and any guidance firms can give for 2028 to clear the semis overhang. Beyond that, he adds, the market is eager for the next wave beyond coding, and the open question is how quickly personal agents will scale and what that