Goldman: data centers hit $602B capex

Data centers are turning into the physical bottleneck of the AI boom. Not the software. Not the models. The buildings, power lines, cooling systems, racks, and chips. Goldman Sachs’ new framing is basically that the market is arguing about AI demand, but the bigger near-term story may be the sheer scale of infrastructure needed to keep the whole thing running. ### What did Goldman actually say? On May 1, Goldman Sachs published a scenario piece that puts the potential AI build-out at about $7.6 trillion in capital spending from 2026 to 2031 across compute, data centers, and power. The point was not “this exact number will happen.” The point was that even small changes in a few infrastructure assumptions can swing the total by huge amounts. (goldmansachs.com) ### Why are data centers the hard part? Because AI infrastructure is physical in a way normal software stories are not. A single model run looks weightless on screen, but underneath it sits a stack of GPUs, cabling, transformers, cooling gear, and utility connections that can take years to secure. Goldman’s piece leans on that gap — AI demand can rise fast, but the supporting hardware and power systems do not. (goldmansachs.com) ### Which assumptions matter most? Goldman highlights four. First, how long AI chips stay economically useful before companies swap them out. Second, how expensive next-generation data centers become as power density rises. Third, what chip mix and system architecture end up winning. Fourth, how much delay comes from shortages in power, labor, and equipment. Those are not side details — they are the levers that can add or subtract hundreds of billions. (goldmansachs.com) ### Why is power suddenly central? Because AI is making data centers much more electricity-hungry. Goldman had already said in late 2025 that data-center power consumption could rise 175% by 2030 from 2023 levels. Then in March 2026 it raised that outlook again, to 220% growth by 2030 versus 2023. That is a huge revision in a short time, and it tells you the constraint is moving from server rooms to the grid. (goldmansachs.com) ### Are companies actually spending at that pace? Yes — and the estimates keep moving up. In December 2025, Goldman said Wall Street’s consensus for 2026 capital spending by AI hyperscalers had already climbed to $527 billion, up from $465 billion at the start of that earnings season. Goldman’s broader trillion-dollar framing goes beyond hyperscaler capex alone, but the direction is the same: the biggest cloud companies are still stepping harder on the gas. (goldmansachs.com) ### Is there enough capacity? Not comfortably. Goldman’s late-2025 data-center work said supply and demand were both rising, but demand was rising a bit faster. In its base case, occupancy stays near peak levels through 2026, with constraints easing only later as more capacity comes online. That means the market is not just expensive — it is tight. (goldmansachs.com) ### Why does this matter beyond tech? Because once capex gets this large, data-center strategy becomes corporate finance strategy. Companies have to decide how hard to reinvest, whether to fund more with debt, where to lock in power, and how much near-term margin pain they can tolerate. Investors are already rewarding the names that can show a clearer path from AI spending to revenue. (goldmansachs.com) ### Bottom line? The AI race is no longer just about who has the best model. It is about who can secure the most real-world infrastructure — fastest, cheapest, and at the right scale. Goldman’s big message is that the next phase of AI will be decided as much by substations and replacement cycles as by algorithms. (goldmansachs.com 1) (goldmansachs.com 2)