Hyperscalers to spend $690B on AI infrastructure

The story here is data centers — not chatbots. AI demand is now big enough that the largest cloud companies are rebuilding the physical layer of computing around it. That means more GPUs, yes, but also substations, transformers, liquid cooling, networking gear, and entire campuses of new server halls. The big shift in early 2026 is that five companies — Microsoft, Alphabet, Amazon, Meta, and Oracle — have now sketched out a combined capex bill of roughly $660 billion to $690 billion for the year, with most of that tied to AI infrastructure. (futurumgroup.com) ### Who is actually spending the money? It’s the hyperscalers — the companies that already own the world’s biggest cloud platforms. Futurum’s February 12 estimate puts Amazon at about $200 billion in 2026 capex, Alphabet at $175 billion to $185 billion, Meta at $115 billion to $135 billion, Microsoft at $120 billion or more, and Oracle at $50 billion. Add that up and you get the headline number. (futurumgroup.com) ### Why is the number so huge? Because AI infrastructure is lumpy. You do not add a little capacity and see what happens. You reserve land, power, networking, and cooling years ahead, then fill those sites with accelerators as fast as supply allows. The interesting part is that the big platforms are acting as if demand is s(futurumgroup.com) AI compute if the capacity existed. (futurumgroup.com) ### Is this all just GPUs? No — and that’s the part people miss. A lot of the spending is really about everything around the chips. The bottleneck has shifted from “can NVIDIA ship enough silicon?” to “can anyone get enough electricity, cooling equipment, and construction done on time?” If the grid connection is late, the G(futurumgroup.com) capex increasingly looks like industrial infrastructure spending with semiconductors attached. (futurumgroup.com) ### Where does NVIDIA fit in? NVIDIA is still the pacing supplier because better inference economics change how much infrastructure buyers are willing to deploy. In February, NVIDIA highlighted SemiAnalysis performance data showing GB300 NVL72 systems with Blackwell Ultra delivering up to 50x higher throughput per megawatt (futurumgroup.com)atters because if inference gets dramatically cheaper, customers run more of it — which can justify even more buildout. (blogs.nvidia.com) ### Why does “35x cheaper” not end the spending boom? Because cheaper compute often increases usage. The cloud business learned this years ago. When the unit cost drops, developers do more. They make models larger, keep context windows open longer, and run more agent-style workflows. NVIDIA’s own framing here is about(blogs.nvidia.com)owing quickly. (blogs.nvidia.com) ### What does this mean for everyone around the hyperscalers? It pushes value outward into the supply chain. Chipmakers benefit first, but so do networking vendors, data-center developers, power equipment makers, cooling specialists, and utilities with available capacity. The catch is that revenue can scale faster tha(blogs.nvidia.com)facility that serves it. (futurumgroup.com) ### So what’s the bottom line? The $690 billion figure is really a signal that AI has moved from a software story into a systems story. The winners will not just be the companies with the best models. They’ll be the ones that can secure power, ship hardware, cool it, and keep lowering the cost of useful inference. (futurumgroup.com)