Goldman warns AI semiconductor boom unsustainable

Semiconductors are the hottest part of the AI trade. That is exactly why Goldman Sachs is getting more careful about them. The firm did not say AI demand is fake or that the build-out is ending tomorrow. It said something more pointed — the spending wave is real, but the stock market may already be paying chip companies as if this pace can run forever. That is the gap. Huge demand does not automatically mean the best returns sit with the most obvious suppliers. ### What did Goldman actually change? The new wrinkle is not just “AI is expensive.” Goldman has been talking for months about how massive the AI build-out could become. On May 1, it published a framework showing total AI infrastructure spending could land in a very wide $4 trillion to $8 trillion range over five years, depending on assumptions like chip replacement cycles, data-center design, and power bottlenecks. But in the past week, Goldman’s equity strategists and analysts pushed the market implication harder — if spending is this extreme, investors should think more about who captures the payoff, not just who sells the hardware. ### Why are chip stocks the pressure point? Because chip stocks have already had the cleanest, easiest narrative. AI demand goes up — GPU and server demand goes up — chip revenues and margins follow. But that simplicity is also the risk. Goldman’s Jim Covello has been arguing that semiconductor valuations have stretched much faster than the underlying certainty around long-term returns, and that hyperscalers may now offer a better risk-reward setup than chipmakers at current prices. (goldmansachs.com) In plain English — the market may have overpaid for the shovel sellers. ### What is the scary number? Capex. Goldman now sees AI hyperscalers spending about $755 billion in 2026, up 83% year over year. Earlier Goldman research had already put 2026 consensus capex around $527 billion, and outside commentary on the latest note says the newer estimate implies spending will absorb roughly 90% or more of expected operating cash flow for the big cloud platforms. That is the kind of number that makes investors stop cheering reflexively, because every extra dollar poured into AI infrastructure is a dollar not going to buybacks, dividends, or near-term free cash flow. (fa-mag.com) ### So is Goldman bearish on AI demand? Not really. That is the important distinction. Goldman’s own research still describes the AI build-out as enormous and highly physical — millions of processors, vast power demand, and data-center costs that keep rising as workloads get denser. The caution is about durability and distribution. Small changes in how often chips get replaced, how quickly power comes online, or how much enterprises are willing to pay can move total spending by hundreds of billions. (finance.yahoo.com) That is not a collapse call. It is a reminder that the boom is conditional. ### Why does enterprise ROI matter so much? Because eventually somebody has to earn money from all this hardware. Goldman’s recent AI skepticism has centered on that missing link — whether enterprise customers can show real returns from generative AI deployments. If cloud providers keep spending aggressively but their customers struggle to monetize AI, the build-out can keep going for a while on competitive pressure alone. But it starts to look less like a self-funding growth loop and more like an arms race. (goldmansachs.com) That is when “unprecedented” starts sounding a lot like “unsustainable.” ### Why prefer hyperscalers over chipmakers? Because hyperscalers have two ways to win. If AI revenue ramps, they monetize the infrastructure directly through cloud and platform services. If spending cools, their cash flow and buybacks can recover. Chipmakers do not have that same cushion when expectations are already sky-high. Goldman’s view, basically, is that the cloud giants now have more balanced upside while semis carry more valuation risk. (goldmansachs.com) ### Is this a bubble call? Not a clean one. Goldman itself has framed AI bubble fears as a real debate, not a settled verdict. The better read is narrower: the semiconductor boom may still be real in revenue terms, but the stock market has concentrated so much hope into a small set of names that even strong demand might not be enough. Deloitte’s 2026 outlook makes the same broader point — AI chips could drive about half of industry revenue while representing a tiny sliver of unit volume, which is great while demand holds and dangerous if it wobbles. (fa-mag.com) ### Bottom line? Goldman is not saying the AI chip boom is over. It is saying the easy trade may be. The build-out is getting bigger, pricier, and more dependent on assumptions that can change fast. When that happens, the winners can shift from the companies selling every extra chip to the platforms that eventually turn all that spending into actual cash. (goldmansachs.com 1) (goldmansachs.com 2)