Goldman Sachs now expects hyperscalers to spend about $1.1 trillion on AI infrastructure in 2027, well above a Wall Street consensus near $920 billion and with a bullish upside that could push 2027 capex toward $1.4 trillion. That revision redraws which firms are likeliest to benefit early: cloud platforms, chipmakers, networking and power suppliers, and data-centre contractors. Goldman frames the market as a stretched rubber band, hyperscalers pulling on one end and semiconductor and infrastructure suppliers on the other, a divergence that creates both opportunity and risk. The bank says supply and demand for AI infrastructure are unlikely to balance until the second half of 2027, keeping capital spending elevated for the next 18 months.

A bank meeting in Manhattan set the tone: analysts there were not debating whether AI would need more servers. They were debating how many, and how fast the rest of the economy could absorb that demand.

Goldman Sachs raised its central forecast for hyperscaler capital spending to about $1.1 trillion for 2027, well above a Wall Street consensus near $920 billion, and sketched a separate bullish scenario that takes spending toward $1.4 trillion. The revision matters because it reshapes which companies are likely to benefit first. Goldman says cloud providers, chipmakers, networking and power suppliers, and the data centre contractors stand to capture the bulk of near-term revenue from the AI buildout.

Concrete signals and a stretched market

Rich Privorotsky, strategist in Goldman’s global banking and markets division, wrote that the market has largely ignored several negative developments around the AI capex trade. Goldman points to concrete demand indicators, notably a large combined backlog at the major cloud providers as of the first quarter, up sharply from six months earlier. The bank treats that jump as evidence hyperscalers are funding a sustained buildout rather than a short burst of one-off purchases.

Goldman’s forecast rests on the view that token consumption, the compute unit tied to how much AI models are used, remains at an early stage. The firm projects token consumption could rise roughly 24 times through 2030, a scale-up that would require far more compute, data centres, networking gear and power capacity than exists today. The analysts noted that AI-related investment already amounted to a noticeable share of GDP in 2026, and they point to historical precedent for multi-year infrastructure booms as a reason to expect elevated hyperscaler capex and continued earnings support for suppliers.

Mastheads

Goldman warned the rubber band can only stretch so far before it snaps back. The note identifies physical and operational bottlenecks that could constrain how quickly spending converts into installed capacity. Delayed data-centre projects, shortages in memory chips, limited power capacity and labour constraints are all real limits on buildout speed.

Those bottlenecks make the path from orders to revenue choppy and extend the time before end customers see consistent service improvements.

Valuation is another pressure point. Parts of the AI infrastructure trade have become increasingly crowded, and stretched prices amplify downside risk if revenue growth slips. Goldman’s analysts highlight that while semiconductor firms have so far captured the lion’s share of the economic value from AI, other parts of the stack haven't yet shown proportional returns.

Jim Covello, head of global equity research at Goldman, said on the Exchanges podcast that the economic value from AI so far has flowed disproportionately to semiconductor companies, while cloud platforms and many downstream adopters have yet to demonstrate matching returns. He added that companies continue to spend aggressively, in part out of fear of being left behind, even when the direct business case for some investments remains unproven. That dynamic can keep capex high even as the revenue payoff remains uncertain.

For suppliers, the mixed picture is both a sales opportunity and a business risk. Data-centre contractors and component suppliers may pull forward revenues as hyperscalers race to secure capacity and parts. But if hyperscalers slow orders because of cost pressure or weaker end-market demand, those same suppliers could face sudden revenue shortfalls, and valuations priced for uninterrupted growth would suffer.

Goldman doesn't expect supply and demand for AI infrastructure to reach balance until at least the second half of 2027. That timing implies elevated capex through the next year and a half and a market in which the rubber band remains taut. The question Goldman leaves for investors is straightforward: how far can the stretch continue before supply constraints, valuation pressure or disappointing returns pull it back?

Related Articles

Goldman’s calendar marker is the second half of 2027, when the bank expects AI infrastructure supply and demand to align. That deadline will determine whether the current stretch turns into a durable expansion or a rapid retracement. Originally reported by MarketWatch.

This article was created with AI assistance.