Hyperscalers raise capex to $725B

Data centers are the story here — not software, not chatbots, not some vague “AI boom.” The biggest cloud companies just told investors they’re going to(bloomberg.com)int to roughly $725 billion of 2026 capital spending combined. That is the real news from this week’s earnings cycle. And it matters because once spending gets this big, the constraint stops being just Nvidia chips and starts becoming everything around them. (bloomberg.com) ### What actually changed t(finance.yahoo.com) April 29, and Meta raised its 2026 range to $125 billion to $145 billion the same day. Microsoft, in investor discussion summarized across coverage of this earnings round, put calendar 2026 capex at about $190 billion, while Amazon kept its figure at $200 billion. Add those together and you get the new headline number — about $725 billion. (abc.xyz) ### Why are they spending this much? Because demand is still there. Google Cloud just crossed $(bloomberg.com)ff enough cash to fund a giant buildout, with revenue up 33% in the March quarter. These companies are basically saying the AI infrastructure they already built is filling up fast enough that they need the next wave now, not later. (abc.xyz) ### So is this mostly GPUs? Not anymore. GPUs are still central, but the cost stack has widened. Microsoft’s spending outlook included about $25 billi(abc.xyz)us tougher competition for land, power, and skilled labor to build data centers. That is the important turn in the story. The shortage is no longer one chipmaker’s problem. It is a full supply-chain problem. (finance.yahoo.com) ### Why does memory suddenly matter so much? AI servers are memory-hungry in a way normal cloud servers were not. High-bandwi(abc.xyz) has been pointing to a memory supercycle, with DRAM revenue surging and suppliers prioritizing AI-linked products because margins are better there. Basically, every extra AI cluster pulls on the same memory pool, and that pushes costs higher across the stack. (spglobal.com) ### Why are land and power in the same sentence as(finance.yahoo.com)eed substations, transformers, cooling, backup generation, transmission access, and permits — then you need the actual building. Meta explicitly flagged competition for land and power. That tells you the limiting factor is becoming how fast you can stand up a site with enough electricity, not just how fast TSMC can package chips. (investor.atmeta.com) ### Does Wall Street l(spglobal.com)w fell to $1.2 billion, driven primarily by a $59.3 billion year-over-year increase in property and equipment purchases tied mainly to AI investment. The market is rewarding proof of demand, but it is getting less patient with blank checks. (ir.aboutamazon.com) ### What does this mean downstream? It means longer lead times, fatter budgets, and more bargaining power for suppliers of(investor.atmeta.com)-and-industrial-base story too. (spglobal.com) ### Bottom line The $725 billion figure matters less as a bragging-rights number than as a signal about where AI has moved. The first phase was a scramble for GPUs. The next phase is a scramble for everything needed to make those GPUs usable at scale — memory, power, buildings, and time. (bloomberg.com)