Hyperscalers Own AI Compute

A few cloud companies now own most of the machines that train and run large artificial intelligence models, and Network World reports that hyperscalers control more than 60% of global artificial intelligence compute capacity, with Google in the lead. (networkworld.com) That means the market is shifting from “who has the best model” to “who can get a slot on the biggest cluster,” because training a frontier model takes thousands of chips wired together like one giant factory line. (networkworld.com) Google got there differently from Microsoft and Amazon because it built its own Tensor Processing Units, which are custom chips designed for artificial intelligence math instead of buying only standard graphics processors from Nvidia. Epoch AI estimates Google held about one quarter of global cumulative artificial intelligence compute capacity by the fourth quarter of 2025, mostly on those in-house chips. (epoch.ai) Owning the chip changes the rest of the stack too, because the same company can decide the silicon, the networking, the software tools, and the cloud pricing in one package. Data Center Frontier says Google is not really selling chips the way Nvidia does; it is selling access to accelerator-heavy computing through Google Cloud. (datacenterfrontier.com) Google is still adding more of that capacity. This week Google Cloud said its new Ironwood Tensor Processing Unit improves carbon efficiency by 3.7 times versus the prior TPU v5p generation, based on measurements from January 2026. (cloud.google.com) The new Intel deal shows how sticky this gets once a cloud platform is built around custom hardware. Intel said on April 9, 2026 that Google and Intel are deepening a multiyear partnership that combines Intel Xeon central processors with custom infrastructure processing units for Google’s artificial intelligence systems. (intel.com) An infrastructure processing unit is the traffic cop of a data center, handling networking and data movement so the expensive artificial intelligence chips spend more time doing model work. Intel said the Google design uses Intel’s Mount Evans and Granite Rapids-D platforms to offload that infrastructure work. (newsroom.intel.com) Amazon Web Services is seeing the other side of the same squeeze: too many customers chasing too little capacity. Network World reported on April 10, 2026 that two large Amazon Web Services customers asked to buy all available 2026 instance capacity for Graviton, Amazon’s custom central processor family, and Amazon said no because other customers also need access. (networkworld.com) Amazon is trying to build its way out of that bottleneck with spending. GeekWire reported in February that Amazon plans about $200 billion in capital expenditures in 2026, with Chief Executive Andy Jassy saying most of it is aimed at Amazon Web Services infrastructure and that the company is “monetizing capacity as fast as we can install it.” (geekwire.com) Once compute is scarce, price and speed stop being purely technical questions and start looking like landlord power. The cloud provider that owns the chips and the data centers can decide who gets low-latency access, which workloads are worth reserving capacity for, and how much customers pay to avoid waiting in line. (networkworld.com) That leaves startups and even large enterprises renting from the same handful of gatekeepers they are trying to compete with. Google’s lead in owned artificial intelligence compute, Intel’s deeper tie-in to Google’s infrastructure, and Amazon Web Services customers bidding for future capacity all point to the same reality: the bottleneck in artificial intelligence is no longer just the model, but the landlord. (epoch.ai, intel.com, networkworld.com)