Google doubles down on Intel

Google just made a bet that the most important chip in an artificial intelligence data center is not only the graphics processor that trains the model, but also the older-looking central processor that keeps the whole machine fed, scheduled, and online. On April 9, Google and Intel said Google will keep deploying multiple generations of Intel Xeon chips across its cloud and artificial intelligence infrastructure. (cnbc.com) (intel.com) That sounds less flashy than a new chatbot, but it reaches into the plumbing of Google Cloud. Intel said the deal covers artificial intelligence training coordination, inference, and general-purpose computing, which means the same supplier is being locked in for both the heavy lifting and the traffic control around it. (intel.com) (reuters.com) Artificial intelligence data centers do not run on one kind of chip. Graphics processors are the race cars, but central processors are the dispatchers, warehouse managers, and air-traffic controllers that move data, assign jobs, and keep thousands of machines from sitting idle. (intel.com) (cnbc.com) That division of labor is getting more important because the market is shifting from training giant models once to serving them constantly. Reuters reported that companies are moving toward deployment and inference, which raises demand for general-purpose central processors that can handle steady, mixed workloads around the clock. (reuters.com) Google said Intel’s newest Xeon 6 chips will run both training and inference jobs, not just ordinary cloud tasks. Intel also said Xeon 6 already powers Google Cloud’s C4 and N4 instances, which are the rented server types customers use to run software inside Google’s cloud. (cnbc.com) (intel.com) The two companies are also expanding work on something called an infrastructure processing unit. That chip is a specialist helper that takes networking, storage, and security chores away from the main central processor, the way a restaurant adds a dedicated expediter so the head chef can keep cooking. (intel.com) (techcrunch.com) Google and Intel have been working on those infrastructure processing units since at least 2021 or 2022, depending on how the companies date the effort, and the new announcement says that work will continue with custom application-specific integrated circuit designs. In plain English, Google is not only buying more Intel parts off the shelf; it is also helping shape chips built for its own data centers. (techcrunch.com) (cnbc.com) This lands in the middle of a broader scramble to rebalance artificial intelligence hardware. CNBC reported that Nvidia’s own head of artificial intelligence infrastructure said in March that central processors are “becoming the bottleneck” as newer agent-style systems demand more coordination outside the graphics processor. (cnbc.com) For Intel, that is exactly the opening it needed. The company missed much of the first wave of the artificial intelligence boom while Nvidia became the standard bearer for graphics processors, so a multiyear Google commitment gives Intel a way back into the center of how hyperscale computing gets built. (cnbc.com) (bloomberg.com) Google’s buying choices also tend to ripple outward because cloud customers test software on whatever hardware the big platforms make easy to rent. When Google standardizes around Xeon generations and Intel-built infrastructure chips, that can quietly shape which combinations of software, drivers, and networking features get tuned first across the wider artificial intelligence stack. (intel.com) (siliconangle.com) The part both companies did not disclose is the money. Intel and Google gave no financial terms and no public timeline, which means the clearest signal here is not a sales number but the fact that Google wants hardware options from Intel across several chip generations instead of treating central processors as a commodity it can swap at the last minute. (cnbc.com) (techcrunch.com)