A question overhanging Wall Street is whether the $19 trillion surge in AI-linked market value already pushes against the ceiling of its plausible economic payoff. That’s after Goldman Sachs issued a pointed warning that valuations in the artificial intelligence sector may be running ahead of near‑term earnings reality.
Ryan Hammond, U.S. equity strategist at Goldman Sachs, said in the note to clients, “Our discussions with investors and recent equity performance reveal limited appetite for companies with potential AI-enabled revenues as investors grapple with whether AI is a threat or opportunity for many companies… investors will likely require evidence of a tangible impact on near-term earnings to embrace these stocks.” His warnings come as high‑profile AI names show signs of losing momentum. Nvidia shares have fallen about 6% over the past five sessions after investors reassessed its quarterly outlook, while Salesforce and Figma both sold off sharply when earnings failed to match the narrative enthusiasm on their calls.
Goldman’s macro team has quantified that disconnect: Their baseline estimate for the present discounted value of US capital revenues from generative AI is about $8 trillion-with an upper bound near $19 trillion. Yet since the launch of ChatGPT in late 2022, the combined market value of companies directly tied to the AI boom-from semiconductor leaders to hyperscalers and the top private model providers-has already risen by more than $19 trillion. That puts current valuations at the top end of the plausible macro benefit range, even before much of the projected productivity gains materialize.
This is compounded by what Goldman describes as the “fallacy of aggregation” extrapolating the exceptional growth of winners into the whole ecosystem-and “fallacy of extrapolation,” while markets treat temporary spikes in profitability as permanent. Similar dynamics during the late 1920s and the tech boom of the 1990s were followed by sharp corrections.
Meanwhile, investment in AI infrastructure has not let up: Bank of America estimates global hyperscale capex will leap 67% in 2025 and another 31% in 2026 to $611 billion. In 2027, Amazon’s data center capacity will be twice that of today, while Google added an additional $92 billion in capex to its 2025 budget and Meta plans to spend almost $100 billion in 2026. These outlays are driving capex intensity to roughly 30% of sales-triple historic norms-and are triggering a second‑order boom in memory, chipmaking tools, and networking equipment. Hammond warns, though, that AI investment as a share of total capital expenditure could be nearing a peak, raising the risk of disappointment should earnings growth lag.
Further complicating matters, enterprise adoption patterns remain fragmented: many corporations remain in pilot phases and are committing limited budgets. On the cloud and semiconductor side, orders associated with AI workloads have been booked; however, this dynamic is not unlike that of the early internet era, where “clicks” and “eyeballs” were voluminous but monetization lagged. Today, few chatbot users pay for premium access, and questions remain over whether end‑user willingness to pay will justify the scale of infrastructure being built.
Third, valuation extremes are sharp. Nvidia trades at about 47 times earnings, Palantir at a P/E above 500, while some AI‑branded firms with negligible revenue have rallied purely on thematic exposure. Palantir’s price‑to‑sales ratio of 117 dwarfs the 30-40 range that marked bubble territory before the dot‑com crash. Even Tesla, which Hammond flags as an AI‑linked name, commands a forward P/E near 200 despite slowing sales growth.
Goldman stops short of calling this a bubble, noting that implied long‑term S&P 500 earnings growth and large‑cap tech valuations are only modestly above historical averages and well below the peaks of 2000 and 2021. But the concentration risk is clear: the largest AI‑exposed stocks now make up nearly 29% of broad U.S. equity benchmarks, which means that most investors are already heavily allocated to the theme whether they intend to be or not.
Execution risks, however, are considerable: data centres are voracious consumers of energy and water, putting a strain on local grids and resources. GPUs account for about 35% of build costs. Replacement cycles could eat into returns if their useful life proves shorter than the assumed five to six years. Much of the hardware is sourced abroad, meaning that the domestic economic multiplier is weaker than the headline capex figures suggest.
For investors, the bottom line of Goldman’s analysis isn’t to shun AI exposure but to recalibrate expectations. Earnings evidence will be the arbiter in what Hammond calls “Phase 3” of the AI trade, as the market starts separating durable winners from overcapitalized hopefuls. Until then, a combination of elevated multiples, slowing earnings momentum and a maturing investment cycle argues for discipline over exuberance.
