Google–Intel infra deal

Google just made a point that gets lost in the artificial intelligence chip race: the expensive graphics processor is not the whole machine. On April 9, Google and Intel said they were expanding a multiyear deal built around Intel Xeon central processing units and jointly developed infrastructure chips inside Google’s cloud. (intel.com) That sounds less flashy than a new model launch, but it sits underneath every answer a chatbot gives. A data center has to move data, schedule jobs, isolate customers, and keep thousands of servers busy before the graphics processors do the math people notice. (intel.com) Google said Intel Xeon chips will keep powering Google Cloud infrastructure across artificial intelligence, inference, and general-purpose workloads. Intel also said the two companies will expand co-development of application-specific integrated circuit infrastructure processing units, which are custom chips for data-center plumbing. (intel.com) An infrastructure processing unit is basically a traffic cop on a crowded highway. It takes networking, security, and storage chores away from the main central processing unit so the main chip can spend more time on customer work. (intel.com) Google is not starting from zero here. Verdict reported that Google already uses Intel in C4 and N4 cloud instances, which are virtual server families sold to customers for tasks including artificial intelligence training, inference, and ordinary computing. (verdict.co.uk) The timing matters because the market spent the last two years talking as if one bottleneck explained everything. Reuters reported on April 9 that rising artificial intelligence use is also reviving demand for traditional central processing units, because serving models at scale still needs general-purpose compute around the accelerator. (reuters.com) Google has been building a mixed hardware stack for years rather than betting on one chip type. In April 2025, Google introduced Ironwood, its seventh-generation tensor processing unit, and said the system could scale to 9,216 chips linked by its own inter-chip network for inference-heavy workloads. (blog.google) That is the backdrop for this Intel deal: Google already has its own tensor processing units for the heavy artificial intelligence math, and it still wants Intel central processing units and custom infrastructure chips around them. The architecture looks less like one super-engine and more like a warehouse full of forklifts, loading docks, and conveyor belts that all have to be tuned together. (blog.google, intel.com) For Intel, this is also a credibility win in a market where investors mostly ask about graphics processors and custom accelerators. CNBC reported that Google committed to using multiple generations of Intel chips in its artificial intelligence data centers, extending a supplier relationship that goes back to Google’s early server-rack buildout. (cnbc.com) For cloud customers, the practical point is cost and utilization, not brand names. If Google can offload networking and security work to infrastructure processing units and keep central processing units fed with the right jobs, it can squeeze more useful work out of the same racks, power, and floor space. (intel.com) The deal is a reminder that artificial intelligence is turning data centers into systems-engineering contests. The winners will not just be the companies with the fastest chip, but the ones that make the whole rack act like one machine. (intel.com, reuters.com)