Google and Intel deepen AI CPU tie‑up

Google just made a bet that the artificial intelligence boom will not run on graphics chips alone. On April 9, Google and Intel said they were expanding a multiyear deal so Google Cloud keeps using future Intel Xeon central processing units and co-develops more custom infrastructure chips with Intel. (reuters.com) (intel.com) A graphics processing unit is the chip that does the heavy math for training large artificial intelligence models, but a central processing unit is still the traffic cop that feeds data, manages memory, and runs a lot of the surrounding software in a data center. Intel and Google are leaning into that split instead of pretending one chip does everything. (techcrunch.com) (intel.com) The new piece is a deeper push into infrastructure processing units, which are custom chips built to take over chores like networking, security, and moving data around a server cluster. Think of them as warehouse loaders that keep boxes moving so the expensive robots do not sit idle. (intel.com) (techcrunch.com) Intel said the companies started working together on these chips in 2021, and the April 9 announcement expands that work into custom application-specific integrated circuit infrastructure processors for Google Cloud. That means Google is not just buying off-the-shelf parts; it is shaping silicon around its own data centers. (intel.com) (reuters.com) This comes at an awkward moment for the usual artificial intelligence story, because the market spent the last two years talking as if Nvidia’s graphics chips were the whole machine. Reuters reported that demand is now shifting from training giant models toward deploying them, and that change is reviving demand for more general-purpose central processing units that can handle steady, mixed workloads. (reuters.com) Google also has a reason to spread its bets. The company already uses its own Tensor Processing Units for artificial intelligence jobs, but cloud systems still need host processors and networking hardware around those accelerators, and Intel wants to be the supplier for that surrounding layer. (cnbc.com) (intel.com) For Intel, this is about staying inside the biggest cloud fleets while rivals push Arm-based server chips and Nvidia pushes full-system artificial intelligence racks. Bloomberg reported that Google committed to using future Xeon generations, which gives Intel a rare public endorsement from one of the companies that buys chips by the data-center hall, not by the box. (bloomberg.com) (cnbc.com) Intel’s own announcement used the phrase “heterogeneous artificial intelligence systems,” which is industry shorthand for mixing different kinds of chips in one cluster. In plain English, the future server row looks less like one giant engine and more like a pit crew, with each chip doing one job well. (intel.com) That is why this deal is bigger than one supplier contract. If Google, one of the world’s largest cloud operators, is locking in central processing units plus custom infrastructure chips alongside accelerators, other hyperscalers are likely to keep building the same layered stack instead of chasing an all-graphics-chip fantasy. (reuters.com) (techcrunch.com)