Broadcom inks long AI deal with Google

Google just locked one chip supplier into its artificial intelligence buildout for as long as 5 years, and it is not Nvidia. Broadcom said on April 6 that it will develop future generations of Google’s Tensor Processing Units and supply networking parts for Google’s next-generation artificial intelligence racks through up to 2031. (sec.gov) The important word here is “racks.” A rack is the metal cabinet inside a data center that holds rows of chips, power gear, and network switches, and Google’s filing says Broadcom is tied not just to the chips but also to the parts that connect those racks together. (sec.gov) Google has been building its own Tensor Processing Units for years instead of buying only off-the-shelf graphics processors, and its current cloud lineup includes the sixth-generation Tensor Processing Unit called Trillium, which Google says is used to train and serve large artificial intelligence models. (cloud.google.com) (docs.cloud.google.com) Broadcom’s job in that setup is not to run Google’s cloud. Broadcom designs and supplies custom silicon and the networking chips that move data between thousands of processors, which is the plumbing that keeps giant artificial intelligence clusters from choking on their own traffic. (sec.gov) (sdxcentral.com) That is why investors reacted fast. Reuters reported on April 6 that Broadcom shares rose after the company disclosed the Google agreement, and the filing gave Broadcom something Wall Street loves in chip cycles: visibility that runs all the way to 2031. (economictimes.indiatimes.com) (sec.gov) This also says something about how the artificial intelligence hardware market is splitting in two. Nvidia still dominates the market for general-purpose artificial intelligence systems, while Google is doubling down on custom chips built for its own software and data centers. (cnbc.com) (cloud.google.com) Broadcom has been telling investors that custom artificial intelligence chips are becoming a huge business on their own. In its September 2025 earnings release, Chief Executive Hock Tan said Broadcom’s third-quarter artificial intelligence revenue reached $5.2 billion, driven by custom accelerators and networking. (sec.gov) The Google deal arrived with a second clue about scale. In the same April 6 filing, Broadcom said Anthropic will get access beginning in 2027 to about 3.5 gigawatts of next-generation Tensor Processing Unit-based computing capacity through the companies’ expanded collaboration. (sec.gov) A gigawatt is power-plant language, not software language. When artificial intelligence deals start getting described in gigawatts, the constraint is no longer just smarter models; it is electricity, buildings, cooling systems, and who has already reserved the hardware years in advance. (sec.gov) (theregister.com) So this is not just a supplier win for Broadcom. It is Google deciding that the race for artificial intelligence will be fought with long contracts, custom chips, and guaranteed network capacity, and Broadcom now has a signed seat inside that machine through 2031. (sec.gov)