Goldman pegs AI infra at $7.6T

AI infrastructure is turning into a physical build-out story — not just a software story. That is the point Goldman Sachs tried to hammer home in a new May 1 framework that maps the AI boom onto chips, data centers, cooling systems, cabling, and power generation. The headline number is huge: about $7.6 trillion of cumulative AI capital spending from 2026 through 2031 in Goldman’s baseline scenario. But the more important point is the catch — Goldman is not saying this is a clean forecast so much as a map of what the bill looks like if the industry keeps scaling. ### What is Goldman actually counting? Not just Nvidia chips. The model covers the whole physical stack needed to run frontier AI at scale — new silicon, next-generation data centers, industrial cooling, networking gear, and the power assets needed to keep those facilities running. Goldman’s framing is that one AI query feels weightless, but the system behind it is extremely heavy — millions of processors, vast cabling, and electricity demand that starts to look like that of countries, not apps. (goldmansachs.com) ### Why is the number so big? Because Goldman’s baseline ramps fast. The firm says annual AI capex would start around $765 billion in 2026 and rise to roughly $1.6 trillion by 2031. Add those years together and you get the $7.6 trillion figure that is now making the rounds. But Goldman explicitly describes this as a scenario-based framework, not a single-point prediction of what will definitely happen. That distinction matters — people are repeating the total like it is settled fact when the whole exercise is really about sensitivity to assumptions. (goldmansachs.com) ### What assumptions move the total most? Goldman says four things do most of the work. First is how long AI chips stay economically useful before they need replacement. Second is how expensive the next wave of data centers gets as power density rises. Third is the chip and architecture mix — basically, whether efficiency gains lower the bill or just unlock more demand. Fourth is bottlenecks in power, labor, and equipment, which can stretch timelines and change the spending curve. (goldmansachs.com) Small changes in those inputs move total capital requirements by hundreds of billions. ### Why does chip lifespan matter so much? Because this is not like building a bridge once and walking away. AI hardware can go obsolete fast. If accelerators need replacing every few years, the industry is not just funding expansion — it is funding constant renewal. That is why Goldman treats replacement cadence as the most powerful variable in the model. A shorter useful life means the same AI demand produces a much bigger capex bill. (goldmansachs.com) ### Why are power and real estate suddenly central? Because AI data centers are becoming power projects in disguise. Goldman has already argued elsewhere that AI facilities could drive a 160% increase in data-center power demand by 2030, and that much of the growth will require new power capacity rather than just plugging into existing grids. Vacancy in major data-center markets is already tight, and the limiting factor is increasingly access to electricity, transmission, and cooling — not just land. (goldmansachs.com) ### So is this bullish or bearish? Basically both. Bullish for the companies selling the picks and shovels — chips, networking, power equipment, cooling, construction, and data-center real estate. But it is also a warning that AI economics are getting more capital-intensive than many investors assumed. Goldman has separately noted that hyperscaler capex expectations for 2026 have been revised sharply higher, with the largest cloud players expected to spend more than half a trillion dollars that year alone. (pwm.gs.com) ### What’s the bottom line? The real story is not that Goldman picked a giant number. It is that the AI race is now constrained by physical infrastructure. If the models keep improving, the winners will not just be the labs with the best software. They will be the firms that can secure chips, power, cooling, land, and financing fast enough to keep building. (goldmansachs.com 1) (goldmansachs.com 2)