Goldman projects $7.6T AI buildout
- Goldman Sachs published a May 1 note arguing AI infrastructure is a supply-side story too, with baseline capital needs reaching $7.6 trillion through 2031. - The swing factors are not just model demand but chip replacement cycles, pricier next-generation data centers, architecture choices, and power bottlenecks delaying buildouts. - That matters because AI’s winners may be utilities, builders, and landlords too — not just model labs and chip designers.
AI looks weightless on a screen. Type a prompt, get an answer, move on. But the thing underneath is brutally physical — chips, cooling, substations, transformers, land, concrete, and a lot of electricity. That is why Goldman Sachs’ new May 1 explainer matters. The firm is arguing that the AI boom is not just a software story or even just a model-training story. It is a giant infrastructure cycle, and the baseline price tag it sketches is $7.6 trillion from 2026 through 2031. ### What did Goldman actually say? The new Goldman Sachs Global Institute note is called “Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out.” The key move in it is simple: stop treating AI capex as one clean forecast number. Goldman says the total is highly sensitive to a few underlying assumptions about how the hardware gets built, refreshed, and powered, which is why the range around any headline estimate matters almost as much as the estimate itself. (goldmansachs.com) ### Where does the $7.6 trillion come from? That $7.6 trillion is Goldman’s baseline aggregate estimate for AI capital spending between 2026 and 2031 across three buckets — compute, data centers, and power. So this is not just a bet on buying more GPUs. It is also a bet on constructing the buildings that hold them and the energy systems that keep them running. Goldman frames it as a scenario baseline, not a promise that spending will land on that exact number. (goldmansachs.com) ### Why is this more than a chip story? Because AI data centers are getting harder and more expensive to build. Goldman says next-generation facilities are rising in cost and complexity as workloads push power density higher and system integration deeper. In plain English — the box around the chip is becoming a huge part of the story. You need denser racks, more cooling, more cabling, and better grid access, and all of that drags in construction firms, equipment suppliers, utilities, and private capital. (goldmansachs.com) ### What are the big swing factors? Goldman highlights four. First, how long AI chips stay economically useful before they need replacing. Second, how expensive the next wave of data centers becomes. Third, the mix of chips and system architectures, which can change margins or total spending depending on whether demand stretches to absorb efficiency gains. Fourth, bottlenecks in power, labor, and equipment, which can slow the buildout enough to shake confidence on the demand side too. (goldmansachs.com) ### Why does power keep showing up? Because power is the hard ceiling. Goldman has been warning for months that data-center electricity demand is set to surge, with one estimate calling for a 165% rise in global data-center power demand by 2030 from 2023 levels, and another 2026 outlook lifting that to 175%. It also expects occupancy in third-party data centers to stay extremely tight into late 2026 before easing as more capacity arrives. Basically, the AI race is now partly a race for megawatts. (goldmansachs.com) ### Why are investors rethinking who wins? Because the profit pool may spread out. If the constraint is not only model quality but also land, grid access, cooling, and financing, then more of the upside can leak away from pure-play AI software into infrastructure owners and enablers. Goldman has already been making that broader case — from digital infrastructure and power to “AI factories” financed by different pools of capital. (goldmansachs.com) ### Is Goldman saying the boom is safe? Not exactly. The catch is that big spending alone does not guarantee big returns. Goldman made that point back in 2024, when it argued that the industry was pouring enormous sums into AI with limited visible payoff so far. The new note does not reverse that warning. It sharpens it. The scale may be enormous, but the exact size — and who captures the economics — depends on assumptions that can move by hundreds of billions. (pwm.gs.com) ### So what is the real takeaway? The cleanest way to read this is that AI is becoming an infrastructure asset class, not just a technology theme. If Goldman is even roughly right, the next phase of the boom will be decided as much by replacement cycles, construction timelines, and grid constraints as by model benchmarks. That changes the map of who matters — and who gets paid. (goldmansachs.com 1) (goldmansachs.com 2)
