Goldman Sachs projects $1.2T in AI infrastructure capex by 2027 as energy becomes key bottleneck

Goldman Sachs strategists are projecting that the five biggest US hyperscalers will pour roughly $1.2 trillion into AI infrastructure capital expenditures in 2027. That figure represents a 50-54% jump from an estimated $800 billion in 2026, and it blows past Wall Street’s consensus forecast of around $1.1 trillion.

The research note, led by strategist Ryan Hammond, goes further: an upside scenario puts the number closer to $1.4 trillion. And looking at the longer horizon, Goldman estimates cumulative AI infrastructure spending could hit $7.6 trillion from 2026 through 2031.

The hyperscaler arms race

The five companies driving this spending tsunami are Amazon, Alphabet, Microsoft, Oracle, and Meta. Each is racing to build out the compute and data center capacity needed to train increasingly powerful AI models and run inference at scale.

Goldman’s team has been revising its projections upward throughout 2026, citing strong quarterly capex results from hyperscalers in Q2 and Q3 as evidence that previous consensus figures were too conservative.

AI infrastructure spending grew nearly 100% in 2026. Goldman expects that rate to slow to 54% in 2027 and then to 12% in 2028, when total spending is projected to reach $2 trillion.

The $300 billion revenue question

There’s a catch embedded in all this optimism. Goldman’s analysis indicates that these five hyperscalers will need approximately $300 billion in annual AI-related revenue just to break even on their infrastructure investments.

That financing shift is worth watching. When companies start borrowing heavily to fund capex, the calculus changes. Debt service costs add pressure to generate returns faster, and any slowdown in AI revenue growth could turn what looks like visionary investment into a balance sheet problem.

Energy, labor, and chips: the triple constraint

Goldman’s strategists flagged three key bottlenecks that could slow or reshape this spending wave: energy supply, labor availability, and memory chip shortages.

Energy is arguably the most fundamental constraint. Training and running large AI models requires enormous amounts of electricity, and data centers are already straining power grids in key markets. The firms building these facilities are exploring everything from nuclear power to natural gas to renewable energy partnerships, but new power generation takes years to bring online.

Memory chip shortages present a third bottleneck. High-bandwidth memory, essential for AI workloads, has been in tight supply as demand has outstripped manufacturers’ ability to scale production. Companies like SK Hynix and Samsung have been ramping capacity, but the gap between supply and demand hasn’t closed.

What this means for markets

For the hyperscalers themselves, the stakes are existential in a competitive sense. Any company that under-invests risks falling behind in AI capabilities, losing cloud customers, and watching rivals capture the market. That dynamic creates a spending floor that’s hard to walk back, even if returns take longer than expected to materialize.

But the risk side of the ledger is real. If AI-related revenue growth doesn’t meet the $300 billion break-even threshold, the market will reprice these companies quickly. And the shift toward debt financing means that interest rate sensitivity is creeping into what has historically been a cash-rich corner of the market.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.