Goldman sees $7.6T AI buildout
- Goldman Sachs published a May 1 scenario analysis arguing AI infrastructure could require $765 billion in 2026 capex and $1.6 trillion by 2031. - The report says the total hinges on a few assumptions — especially chip replacement cycles, rising data-center complexity, and bottlenecks in power, labor, and equipment. - That matters because the chokepoint may shift from GPUs to optics, cabling, and grid-connected megasites.
AI suddenly looks less like a software story and more like a construction story. That’s the real point of Goldman Sachs’ new May 1 note on the AI buildout. The headline number is huge — annual AI capital spending of $765 billion in 2026, rising to $1.6 trillion by 2031 — but the more interesting part is that Goldman is not treating that path as destiny. It’s treating it as a physical supply-chain problem with a lot of moving parts. ### What did Goldman actually say? Goldman’s piece is a scenario framework, not a clean forecast. The bank lays out how a few assumptions can swing the total capital needed for AI infrastructure by hundreds of billions of dollars — even if demand for AI itself stays strong. The focus is the hardware stack behind AI: chips, data centers, power, cooling, cabling, and the replacement cycle that keeps all of it current. (goldmansachs.com) ### Why is the number so big? Because AI at scale is brutally physical. Goldman’s framing is simple: one chat query feels weightless, but the system answering it depends on millions of processors, industrial cooling, vast cable runs, and power demand that can rival midsize countries. Once you think in those terms, trillion-dollar totals stop sounding like hype and start sounding like a buildout model for a new utility. (goldmansachs.com) ### Which assumptions matter most? Goldman highlights four. First, how long AI chips stay economically useful before operators rip them out and replace them. Second, how expensive next-generation data centers become as rack density and system integration rise. Third, what mix of chips and architectures win(goldmansachs.com)ically, the size of the boom depends on refresh speed, building complexity, and how clogged the supply chain gets. (goldmansachs.com) ### Why does replacement cadence matter so much? Because AI silicon is not like a warehouse roof or a transmission line that you expect to keep for decades. If operators decide top-end accelerators need replacing faster to stay competitive on training or inference economics, cumulative spending jumps hard. (goldmansachs.com)goldmansachs.com) ### What does a real site look like? Crusoe’s Abilene, Texas campus is a good example of the scale. In March, Crusoe said it was adding a new 900 MW AI factory campus for Microsoft next to its existing site, bringing the full Abilene buildout to about 2.1 GW. Land clearing is already underway, and the firs(goldmansachs.com)ritory. (crusoe.ai) ### So are GPUs still the bottleneck? Maybe not for long. Wall Street has started talking about optical interconnects as the next choke point — the fiber, lasers, and high-speed links that let giant GPU clusters act like one machine. Copper runs out of room as clusters scale. If data ca(crusoe.ai)king that risk. (morningstar.com) ### Why do power and labor keep coming up? Because you can finance a GPU order faster than you can permit power, train specialized workers, or build dense liquid-cooled facilities at gigawatt scale. The catch is that AI demand can look explosive on paper while deployment lags in the real world. Goldman’s point is that these frictions do not just slow projects — they can reshape the whole spending curve. (goldmansachs.com) ### What’s the bottom line? The interesting part of Goldman’s note is not just the trillions. It’s the reminder that AI’s next phase will be decided by replacement schedules, substations, optics, and construction crews — not just better models. (goldmansachs.com)
