Ewerton posts $930B data capex

Data centers are becoming the physical shape of the AI boom. Not apps, not chatbots, not model demos — concrete, transformers, cooling loops, and racks of accelerators. That is why Ewerton Costa’s $930 billion figure landed: it puts a giant number on something investors already feel but haven’t fully sized. The point is not just that spending is huge. It is that the industry may be building ahead of proven productivity gains. ### What is the $930 billion actually measuring? Costa’s post appears to refer to hyperscaler-style data-center capital expenditure over roughly six years — basically the money the biggest cloud and AI platforms have poured into new facilities, servers, networking, and power infrastructure. The cleanest outside cross-check is not his post itself but the broader market work around it: Federal Reserve researchers nowcast aggregate U.S. data-center investment rising to a $370 billion annualized pace by 2026:Q2, while Dell’Oro sees worldwide data-center capex reaching $1.7 trillion by 2030. (frbsf.org) ### Why are people suddenly treating this like heavy industry? Because it is heavy industry now. Cushman & Wakefield says operational capacity across the 97 global markets it tracks now exceeds 40GW, with AI demand accelerating the build-out and power availability becoming the top constraint. JLL goes even bigger and (frbsf.org)lize the internet.” (cushmanwakefield.com) ### Is Costa’s warning basically “this could be a bubble”? More or less — but not in the simple dot-com sense. The concern is not that AI demand is fake. The concern is that infrastructure is being financed on the assumption that model usage, enterprise adoption, and monetization will catch up fast enough to justify all the steel in the ground. Dell’Oro itself says AI spe(cushmanwakefield.com) looks intact. That is the tension. (delloro.com) ### What makes this build-out different from older cloud cycles? Power. Older cloud waves could expand inside an existing utility and real-estate envelope. AI pushes much harder on both. CBRE says global power shortages are already constraining growth and pushing operators toward secondary markets with better grid access. Knight Frank puts 2024 data-center capex near $228 billion and (delloro.com)r, much of it tied to AI. (cbre.com) ### So is the bottleneck chips or electricity? Increasingly electricity. GPUs are still scarce and expensive, but more markets can buy hardware than can secure enough power, substations, and transmission on the right timeline. McKinsey says global data-center spend could reach $1.7 trillion through 2030, and a big part of the challenge is simply building larger sites faster and cheaper. In other (cbre.com)ctually run them?” (mckinsey.com) ### Does that make the $930 billion number too high? Not obviously. If anything, it may end up looking conservative depending on what is included. A six-year cumulative figure of about $930 billion sits comfortably beside a $370 billion annualized run rate in 2026 and forecasts of $1.7 trillion by 2030. The real debate is not arithmetic. It is utilization — whether enough high-value AI work shows up to keep all this capacity busy. (frbsf.org) ### What should investors watch next? Watch three things: utility interconnection timelines, signs that inference demand is broadening beyond a handful of labs, and whether returns improve before another spending leg begins. If productivity gains lag, this starts to look speculative. If usage and pricing hold up, the build-out starts to resemble a new industrial base for computing. (cushmanwakefield.com) ### Bottom line Costa’s post works because it translates AI from abstraction into capex reality. The catch is that the industry is no longer just betting on better models. It is betting that trillions in power-hungry physical infrastructure will earn their keep. (frbsf.org)