Compute choices are reshaping AI stacks

The fight over artificial intelligence is no longer just about whose model is smartest. It is also about which chips run the boring but essential work around those models, and where companies are willing to build the buildings to house them. (intel.com) (bloomberg.com) A modern artificial intelligence system does not run on one kind of chip. Graphics processors handle the heavy math for training and serving models, while central processing units manage memory, storage, networking, scheduling, and a long list of background jobs that keep the whole system from stalling. (intel.com) Google and Intel said on April 9, 2026 that they are extending a multiyear partnership built around Intel Xeon server chips and custom infrastructure processing units for Google Cloud. Intel said those chips will keep powering Google Cloud infrastructure across artificial intelligence, inference, and general-purpose workloads. (intel.com) An infrastructure processing unit is the traffic cop of a data center. Instead of asking the main chip to handle every packet, storage request, and security check, the infrastructure processing unit takes over that plumbing so the expensive compute chips can spend more time on customer work. (intel.com) (cloud.google.com) Google already uses Intel-based designs in parts of its cloud fleet, including C3 virtual machines that pair Intel Xeon processors with Google’s custom Intel infrastructure processing unit. That makes this week’s announcement less like a surprise switch and more like a decision to keep scaling a design Google already knows how to operate. (cloud.google.com) (siliconangle.com) That matters because the market story around artificial intelligence chips has tilted hard toward graphics processors and custom accelerators. Intel and Google are making the opposite point: even in a heterogeneous system packed with specialized silicon, the central processing unit and the network-and-storage offload chip still decide how efficiently the whole machine runs. (intel.com) (tech.yahoo.com) At the same time, OpenAI has paused its Stargate United Kingdom data-center plan. Bloomberg reported on April 9, 2026 that OpenAI cited high energy costs, and Reuters reported the pause also came amid concerns about regulation around artificial intelligence and copyright. (bloomberg.com) (reuters.com) That pause shows the other half of the compute story. A data center is not just chips in racks; it is power prices, grid access, permits, and legal rules, and one bad line item can stop a multibillion-pound build even when demand for artificial intelligence keeps rising. (reuters.com) (engadget.com) While the hyperscalers are deciding where to place giant workloads, companies like Nutanix are selling a different answer to enterprises that do not want to hand everything to a public cloud. Nutanix said this week that its Agentic AI platform will bundle compute, storage, networking, Kubernetes software, governance, and management into one stack, with broader availability planned for the second half of 2026. (nutanix.com) So the stack is starting to split three ways at once. Google is betting that carefully tuned central processing units plus custom infrastructure chips still belong under huge cloud artificial intelligence systems, OpenAI is being reminded that geography can kill expansion plans, and Nutanix is pitching companies on owning more of the machinery themselves. (intel.com) (bloomberg.com) (nutanix.com) The result is that “where will artificial intelligence run” is becoming a more useful question than “who has the best model.” In 2026, the answer depends on which company can secure chips, cheap electricity, and enough control over the software stack to keep costs from exploding. (intel.com) (reuters.com) (nutanix.com)