Intel and Google expand AI chip work
Intel and Google announced a multiyear collaboration to extend custom IPUs and cloud AI infrastructure, reflecting continued hardware-level specialization for large AI workloads. That kind of vendor collaboration speaks to the shifting economics of inference and how providers are trying to squeeze more performance and efficiency out of AI deployments. (x.com)
An artificial intelligence system does not run on one magic chip. A Google Cloud data center splits the job between general-purpose central processing units that manage traffic and specialized chips that do the heavy math, and Intel and Google said on April 9, 2026 that they are extending that division of labor in a new multiyear deal. (intel.com) The central processing unit is the traffic cop. Intel said its Xeon server processors will keep powering Google Cloud across artificial intelligence, inference, and general-purpose workloads, which means the old workhorse chip is still in the loop even as custom silicon gets more attention. (intel.com) Inference is the moment a model answers a question instead of learning from new data. Google has been pushing chips built specifically for that stage, including its seventh-generation Tensor Processing Unit called Ironwood, which it introduced in April 2025 as its first design made specifically for inference. (blog.google) The new Intel-Google piece is not that Google is replacing its Tensor Processing Units. It is that Google and Intel are expanding work on custom infrastructure processing units, which are chips that move data around the network and keep expensive artificial intelligence accelerators from sitting idle. (intel.com) Intel said these infrastructure processing units are application-specific integrated circuits, which means they are built for one narrow job instead of many. In a cloud system, that narrow job can be packet handling, storage movement, and security offload, like hiring warehouse staff so the star machinery never stops. (intel.com) Google has used that approach before. In 2021, Google Cloud introduced its own infrastructure processing unit to offload networking and security tasks from the main server processor, saying the design gave customers near line-rate performance while freeing central processing unit cores for applications. (cloud.google.com) That is why this deal is about cost as much as speed. Intel said the two companies will align across multiple future Xeon generations and co-develop more custom infrastructure processing units to improve efficiency, utilization, and performance at scale, which is cloud-company language for getting more work out of the same rack of hardware. (intel.com) The backdrop is that artificial intelligence spending has moved from training giant models once to serving answers constantly. Google’s Ironwood launch called inference the new center of gravity for generative artificial intelligence, and inference burns money every time a user types a prompt, so shaving watts and milliseconds off each request adds up fast. (blog.google) Intel also needs this kind of win. The company has been trying to prove that Xeon server chips still matter in an artificial intelligence stack dominated by graphics processing units and custom accelerators, and Google publicly said Intel has been a trusted partner for nearly two decades and that the Xeon roadmap still fits Google’s performance and efficiency needs. (intel.com) So the short version is not “Intel builds Google’s artificial intelligence chips.” The real story is that Google is stitching together a layered machine room where Tensor Processing Units handle model math, Xeon chips coordinate the system, and custom infrastructure processing units keep data moving cheaply enough for cloud artificial intelligence to scale. (intel.com)
