Google picks Intel pragmatically
Google Cloud is expanding beyond custom silicon and will deploy Intel Xeon platforms alongside its own IPUs, signalling a multi‑architecture, optionality-first approach to AI infrastructure. This move balances performance with supply resilience and bargaining power rather than an ideological bet on one chip vendor. For platform leaders, it’s an example of making infrastructure choices that preserve future flexibility while supporting near-term capacity needs. (Tom's Hardware)
Google just made a very un-Google move in artificial intelligence infrastructure: after years of pushing its own chips, Google Cloud said on April 9 it will keep deploying Intel Xeon processors and expand joint work with Intel on custom infrastructure chips. Reuters, CNBC, Intel, and Google all describe it as a multi-year deal, not a one-off purchase. (reuters.com) (cnbc.com) (newsroom.intel.com) A cloud data center does not run on one kind of chip. It uses central processing units, which are the general-purpose brains; accelerators, which do the heavy artificial intelligence math; and networking chips, which move data around like traffic lights on a freeway. (cloud.google.com) (blog.google) Google already has its own accelerator family, called Tensor Processing Units. In April 2025, Google introduced Ironwood, its seventh-generation Tensor Processing Unit, and said it was built specifically for inference, which is the step where a trained model answers questions and generates outputs for users. (blog.google) (techcrunch.com) Google also has its own server processor now. In 2025, Google Cloud rolled out Arm-based Axion virtual machines alongside Ironwood, which made it look like Google was steadily replacing outside chip vendors with more in-house silicon. (cloud.google.com) (blog.google) This week’s Intel deal shows that was only half the story. Intel said Google Cloud is already using Xeon 6 processors in its C4 and N4 instances, and the new agreement commits Google to multiple future generations of Xeon for artificial intelligence, inference, and general cloud workloads. (newsroom.intel.com) (siliconangle.com) The second piece is the custom infrastructure chip. Intel and Google started working together on this category in 2021 with a project called Mount Evans, and these chips are designed to take housekeeping work away from the main processor so the central processing unit can spend more time on customer workloads. (datacenterdynamics.com) (techcrunch.com) That housekeeping work includes networking, security, and storage tasks that every cloud server has to do before any artificial intelligence model can answer a prompt. Intel and Google said the next phase will expand co-development of application-specific integrated circuit infrastructure processing units, which means chips built for one narrow job instead of everything at once. (newsroom.intel.com) (intc.com) The practical reason is capacity. Reuters reported that rising artificial intelligence use is reviving demand for traditional central processing units, because even clusters packed with accelerators still need lots of host processors to coordinate training, feed data, and run the rest of the cloud stack. (reuters.com) (newsroom.intel.com) The business reason is leverage. If Google can mix Tensor Processing Units, Arm-based Axion chips, Nvidia graphics processors, and Intel Xeons inside one cloud, it is less exposed to any single supplier’s delays, prices, or product roadmap. That is an inference from Google’s public multi-chip portfolio and the new Intel commitment, not a quoted line from either company. (cloud.google.com 1) (cloud.google.com 2) (cnbc.com) For Intel, the deal is a foothold inside one of the world’s biggest cloud operators at a moment when Nvidia dominates the artificial intelligence conversation. For Google, it is a reminder that owning some chips does not mean you stop buying others when you need more capacity, more flexibility, or both. (cnbc.com) (reuters.com)
