Google and Intel deepen AI CPU ties

Google just locked itself into more Intel chips at the same moment it is also pushing its own custom silicon, which tells you the fight inside data centers is no longer one chip versus another but how many different chips can be stitched together without wasting money or power. On April 9, Intel and Google said Google Cloud will keep deploying Intel Xeon processors and expand joint work on custom infrastructure processing units for artificial intelligence systems. (intel.com) A central processing unit is the general manager of a server, the chip that handles operating systems, memory movement, storage requests, and all the smaller jobs that keep a machine alive while an artificial intelligence model runs. Intel’s point in this deal is that even when a graphics processor does the heavy math, somebody still has to coordinate the warehouse around it. (reuters.com) Google’s cloud already uses Intel in real products customers rent today, including C4 and N4 virtual machines. Google said its C4 machines based on Intel Xeon 6 were generally available in 2025, with up to 30% better general compute performance and up to 60% better machine learning recommendation performance than the prior generation. (cloud.google.com) The new piece is the shared work on infrastructure processing units, which are helper chips that take networking, security, and data-movement chores off the main processor. Intel said these units will be custom application-specific integrated circuits, meaning chips built for a narrow job instead of broad, all-purpose computing. (intel.com) Google has been building more of its own chips for years, so this is not a retreat from custom hardware. Google sells Arm-based Axion central processing units for standard cloud workloads and said its sixth-generation Trillium Tensor Processing Unit is its most powerful artificial intelligence chip, with 4 times the performance and 67% better energy efficiency than the prior version. (cloud.google.com) (blog.google) That means Google is now running a mixed kitchen: its own Tensor Processing Units for artificial intelligence math, its own Axion chips for some general computing, Google Titanium offload hardware for system tasks, and Intel Xeons where x86 software compatibility and broad enterprise support still matter. Google described Titanium as a set of purpose-built microcontrollers and offloads underneath C4 and N4 machines. (cloud.google.com 1) (cloud.google.com 2) Intel needs this kind of deal for a different reason. Chief executive Lip-Bu Tan is trying to prove Intel can still win a place in artificial intelligence infrastructure even after Nvidia became the default name for training chips and even after cloud companies started designing more processors in-house. (cnbc.com) (reuters.com) Google needs it because cloud customers do not move their software stacks overnight. Many enterprise workloads still expect x86 servers, and Intel said Google will keep using multiple generations of Xeon across artificial intelligence, inference, and general-purpose workloads rather than flipping everything to one homegrown design. (intel.com) (cnbc.com) So the real story is not that Google picked Intel over its own chips. The real story is that hyperscale cloud companies are rebuilding the plumbing under artificial intelligence services piece by piece, and the winners will be the vendors that can make central processing units, helper chips, and accelerators work together as one rentable machine. (techcrunch.com) (intel.com)