Google inks Intel chip deal
Google and Intel announced a multi‑year agreement under which Google will continue deploying Intel Xeon CPUs alongside custom infrastructure processing units for its next‑generation AI and cloud platform. The deal is framed as part of a broader move away from single‑vendor stacks toward heterogeneous mixes of CPUs, accelerators and custom silicon. Observers see this as another sign hyperscalers are assembling bespoke hardware stacks rather than relying on one accelerator type. (tomshardware.com) (simplywall.st)
Google just signed up to keep buying Intel server chips for its next wave of cloud and artificial intelligence systems, even after spending years building more of its own silicon. Intel said on April 9 that Google Cloud will keep deploying Intel Xeon processors and will expand joint work on custom application-specific integrated circuit infrastructure processing units. (intel.com) That sounds strange only if you picture an artificial intelligence data center as one giant pile of graphics chips. In practice, a modern data center is more like a kitchen line: one machine chops, one cooks, one plates, and the server central processing unit still coordinates a lot of the work. (intel.com) Intel said Google is already using its latest Xeon 6 chips in Google Cloud C4 and N4 instances. Those systems handle jobs ranging from general computing to latency-sensitive inference, which is the step where a trained model answers a prompt in real time. (intel.com) Google has not been betting on Intel alone. Google Cloud also sells Axion processors, which are Google-designed Arm-based central processing units for standard cloud workloads, and Google says C4A instances with Axion deliver up to 10% better performance per virtual central processing unit than rival Arm cloud instances. (cloud.google.com) Google also keeps building its own Tensor Processing Units, which are chips made specifically for artificial intelligence math. At Google Cloud Next in April 2025, Google introduced Ironwood, its seventh-generation Tensor Processing Unit, and said it was designed specifically for inference. (blog.google) So the picture inside Google now is not Intel or Google silicon. It is Intel Xeon for a range of cloud and orchestration jobs, Google Axion for some general-purpose computing, and Google Tensor Processing Units for heavy artificial intelligence serving. (intel.com) (cloud.google.com) (blog.google) The extra piece in this week’s deal is the infrastructure processing unit, which is a chip that takes over networking, storage, and security chores that would otherwise eat up the main server processor. Google and Intel have worked on custom Intel infrastructure processing units before, including the custom Intel infrastructure processing unit used in Google Cloud C3 machines. (cloud.google.com) (intel.com) That is why this is less a comeback story for Intel than a map of how hyperscale cloud companies now build computers. They are mixing central processing units, artificial intelligence accelerators, and custom helper chips in the same rack instead of trying to force every task onto one kind of processor. (intel.com) (tomshardware.com) Intel gets something important out of that mix even without owning the whole stack. Google Cloud is one of the biggest buyers of server hardware in the world, and Intel said this agreement covers multiple generations of Xeon, which means its chips stay embedded in Google’s cloud fleet for years rather than quarters. (intel.com) For Google, the deal buys flexibility more than exclusivity. Some customers still want x86 compatibility, some workloads care about single-threaded central processing unit speed, and some artificial intelligence systems need custom accelerators plus ordinary server chips working side by side. (tech.yahoo.com)
