Google‑Intel AI tie‑up

Google and Intel just made a deal about the least glamorous part of artificial intelligence: the chips that keep giant data centers fed, cooled, and moving data fast enough that expensive processors do not sit idle. Intel said on April 9 that Google will keep using Intel Xeon central processing units across Google Cloud and will expand joint work on custom infrastructure processing units. (intel.com) A central processing unit is the traffic cop of a server. It does not draw every pixel or train every model itself, but it schedules work, moves data, and keeps the rest of the machine from turning into a traffic jam. (intel.com) Google’s own cloud service already uses Intel chips in several virtual machine families, including C4 and N4 instances. Google said last year that C4 machines moved to Intel Xeon 6 for workloads like databases, analytics, gaming, and artificial intelligence. (cloud.google.com) The new piece is the infrastructure processing unit. Intel described it as a custom application-specific integrated circuit, which is a chip built for one narrow job, and here that job is handling the plumbing around storage, networking, and data movement so the main processors can spend more time on actual computing. (intel.com) That sounds small until you remember what an artificial intelligence data center wastes money on. If a graphics processor waits for data for even a few milliseconds, a company that spent billions on accelerators is paying sports-car prices for a machine stuck at a red light. (networkworld.com) Intel is leaning hard into that argument because it is no longer the company setting the pace in the most visible artificial intelligence chips. Its pitch is that modern systems are “heterogeneous,” meaning they mix different kinds of chips, and the central processing unit still matters because somebody has to coordinate training, inference, security, and memory across the whole rack. (intel.com) Google is an unusually important customer because it is not just renting out chips to others. Network World reported this week that hyperscalers control more than 60 percent of global artificial intelligence compute, and Google is the largest single owner by cumulative capacity because it built much of that stack its own way. (networkworld.com) (epoch.ai) Epoch AI estimated that Google held about one quarter of global cumulative artificial intelligence compute capacity by the fourth quarter of 2025. The same analysis said Google stands out because much of that capacity comes from its own tensor processing units instead of Nvidia graphics processors. (epoch.ai) That is why this Intel deal is not a simple comeback story. Google is still investing in its own tensor processing units for model training and inference, while using Intel’s central processing units and new infrastructure chips for the surrounding work that keeps those systems busy and rentable inside Google Cloud. (cloud.google.com) (intel.com) The power question sits underneath all of it. When a few cloud companies own most of the scarce compute, they do not just decide which chip wins inside their buildings; they also shape who gets access, what developers pay per hour, and which hardware suppliers survive long enough to build the next generation. (networkworld.com) So the April 9 announcement is really two stories at once. Intel gets a public vote of confidence from one of the world’s biggest buyers of compute, and Google gets another way to squeeze more useful work out of every server in a market where control of the plumbing is starting to look as valuable as control of the models. (intel.com) (networkworld.com)