Google doubles down on Intel CPUs

Google just made a very un-fashionable bet in the middle of an artificial intelligence chip rush: it signed a new multiyear deal to keep buying Intel central processing units for Google Cloud and to build more custom data-center chips with Intel. Intel announced the deal on April 9, 2026. (intel.com) That stands out because the loudest story in artificial intelligence infrastructure has been graphics processing units, especially Nvidia’s. Google is saying its next wave of cloud and artificial intelligence systems still needs Intel Xeon chips for training coordination, inference, and general-purpose computing. (networkworld.com) Google is not choosing Intel instead of its own chips. Google already sells Arm-based Axion central processing units for some cloud customers, and it also runs Tensor Processing Units, which are Google’s in-house chips built specifically for neural networks. (cloud.google.com 1) (cloud.google.com 2) What Google is buying from Intel is the part of the machine room that acts like the traffic manager. A central processing unit handles operating systems, memory movement, storage, networking, and a long list of jobs that specialized artificial intelligence chips do not handle well on their own. (intel.com) (cloud.google.com) The second part of the deal is about infrastructure processing units, which are custom chips that peel off chores like moving data, handling security, and managing network traffic. Intel said Google and Intel will expand co-development of application-specific integrated circuit-based infrastructure processing units to improve efficiency, utilization, and performance at scale. (intel.com) Google has already been mixing Intel chips into current cloud products. Google Cloud says its C4 virtual machines now run on Intel’s sixth-generation Xeon Granite Rapids processors, and Google documentation lists both C4 and N4 machine families among its general-purpose compute options. (cloud.google.com) (docs.cloud.google.com) Intel needs this kind of win badly. Google launched Axion in April 2024 with claims of up to 50% better performance and up to 60% better energy efficiency than comparable current-generation x86 cloud instances, which showed that even longtime Intel customers were building alternatives inside their own walls. (cloud.google.com) That is why the wording of this deal matters. Intel said Google will align across multiple generations of Xeon processors, which suggests this is not a one-cycle purchase but a roadmap commitment that helps Intel lock in future volume and helps Google plan power, cooling, and software years ahead. (intel.com) (datacenterdynamics.com) The bigger picture is that hyperscale cloud companies are no longer building around one magic chip. Google now has graphics processing units from outside suppliers, Tensor Processing Units for neural networks, Axion chips for some general computing, Titanium offload hardware inside its servers, and now a fresh multiyear Intel Xeon commitment on top. (cloud.google.com 1) (cloud.google.com 2) (intel.com) That mix is really a power bill story as much as a chip story. Every watt saved in a data center turns into less heat, less cooling, and lower cost per artificial intelligence query, which is why Google and Intel framed the partnership around performance and energy efficiency instead of raw chip counts. (intel.com) (networkworld.com) So the surprise here is not that Google likes Intel. The surprise is that in 2026, after launching its own central processing unit and after years of graphics processing unit hype, Google still thinks the safest way to build artificial intelligence infrastructure is to spread the job across central processing units, custom offload chips, and specialized accelerators instead of betting the whole data center on one kind of silicon. (cloud.google.com) (intel.com)