Google‑Intel AI Deal

Google just made a very old chip category newly important. On April 9, Google and Intel said Google Cloud will keep deploying multiple generations of Intel Xeon central processing unit chips, including Xeon 6, across artificial intelligence, inference, and general-purpose workloads. (intel.com) The second part of the deal is more unusual: Google and Intel are expanding their work on custom infrastructure processing units, which are chips built to handle the plumbing of a data center instead of the math inside an artificial intelligence model. Intel said these chips are meant to improve efficiency, utilization, and performance at scale. (intel.com) A central processing unit is the general manager inside a server. It handles lots of different jobs, and in an artificial intelligence system that includes coordinating training runs, feeding data to accelerators, and serving responses fast enough that a chatbot does not feel slow. (networkworld.com) An infrastructure processing unit is more like a building superintendent. Google’s 2022 launch of C3 machines said its custom Intel infrastructure processing unit offloads networking and storage work so the main server chip can spend more time on customer workloads. (cloud.google.com) That sounds like a side detail until you look at where artificial intelligence capacity sits. Network World reported this week that more than 60% of global artificial intelligence compute capacity is now controlled by hyperscalers, the giant cloud companies that own the biggest data centers. (networkworld.com) One research estimate goes further and puts Google alone at about one quarter of global cumulative artificial intelligence compute capacity as of the fourth quarter of 2025, driven mainly by Google’s own tensor processing units. That means Google is not just renting out servers; it is shaping the supply of the raw industrial input behind modern artificial intelligence. (epoch.ai) This is why Intel’s role here is not mainly about beating Nvidia in the graphics processor race. Google already buys Nvidia chips and builds its own tensor processing units, but this deal locks Intel deeper into the layers that connect servers, storage, and networks into one working cloud product. (cnbc.com, cloud.google.com) Google is also making a bet about what the next bottleneck looks like. Its own network team wrote in 2025 that artificial intelligence is driving an explosive surge in network capacity demand, with new traffic patterns from large-scale training and inference. (cloud.google.com) If network traffic, storage access, and security checks become the choke points, then the company that owns those layers can squeeze more useful work out of the same building, power feed, and cooling plant. DatacenterDynamics reported that the infrastructure processing unit frees server central processing unit cycles for revenue-generating tasks. (datacenterdynamics.com) Intel first disclosed this infrastructure processing unit collaboration with Google in October 2021 under the name Mount Evans. The new agreement turns that earlier experiment into a multiyear supply and design relationship inside Google Cloud. (datacenterdynamics.com, intel.com) For Google Cloud customers, the practical result is not a shiny new consumer product. It is that the company selling the cloud is trying to control more of the machine underneath it, from the custom chip that runs the model to the chip that moves the data to the central processing unit that keeps the whole service fed. (intel.com, networkworld.com)