Google–Broadcom TPU extension
Google and Broadcom extended their TPU partnership out to 2031, signaling a multi‑year commitment to custom TPU hardware and components. UBS and market commentary say the deal pushed analysts to raise Broadcom TPU shipment forecasts for 2027, suggesting TPU capacity is being planned at industrial scale rather than as a short‑term experiment. Separately, reports say Anthropic has aligned much of its future compute with Google TPU capacity and has secured roughly 3.5GW of TPU capacity starting in 2027, reinforcing the move toward long‑horizon supplier commitments. (au.investing.com (news.futunn.com) (www.indexbox.io)
Google and Broadcom have extended their custom artificial intelligence chip partnership through 2031, locking in Google’s Tensor Processing Unit roadmap for years ahead. (reuters.com) Broadcom said on April 6 that it will develop and supply future generations of Google’s Tensor Processing Units, or TPUs, and provide components for Google’s next-generation artificial intelligence racks through 2031. Reuters reported the agreement alongside a separate Anthropic deal tied to Google’s TPU capacity. (reuters.com) Anthropic said the same day that it signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity expected to come online starting in 2027. The company said the capacity will power future Claude models and support rising customer demand. (anthropic.com) A Tensor Processing Unit is Google’s in-house chip built specifically for artificial intelligence, rather than for the broader mix of computing jobs handled by central processing units or graphics processing units. Google says it has been building TPUs for more than a decade to run training and inference workloads more efficiently. (cloud.google.com) Google’s latest widely available TPU generation is Trillium, its sixth-generation chip. Google Cloud said in December 2024 that Trillium was generally available and delivered four times the performance and 67% better energy efficiency than the prior generation. (blog.google) The new Broadcom agreement points to a supply plan measured in product generations, not quarter-to-quarter orders. Reuters said the deal covers future TPU generations through 2031, and Broadcom’s filing described support for Google’s next-generation artificial intelligence racks as well as the chips themselves. (reuters.com) (convergedigest.com) Analysts moved quickly to raise their forecasts after the announcement. UBS lifted its 2027 TPU shipment estimate to about 7 million units from about 6 million and raised its Broadcom fiscal 2027 revenue forecast to $194.9 billion from $182.2 billion, according to market coverage of the note. (finance.biggo.com) (au.investing.com) UBS also raised its estimate for TPU-related spending to $26 billion in 2026, $61 billion in 2027, and $79 billion in 2028, Yahoo Finance reported. That forecast suggests Google is planning TPU capacity at data-center scale as it sells more artificial intelligence computing through Google Cloud. (finance.yahoo.com) Broadcom’s role is not new, but the public commitment is. Reporting in 2023 said Broadcom had already been helping translate Google’s TPU designs into manufacturable chips and supplying key high-speed interfaces that let the processors communicate with the rest of the system. (theregister.com) The Anthropic piece shows why the timetable matters. Reuters reported Anthropic secured about 3.5 gigawatts of artificial intelligence computing capacity based on Google processors starting in 2027, turning Google’s custom chips into reserved infrastructure for one of the largest model developers. (reuters.com) By 2031, the question will be less whether Google keeps building TPUs than how much of the artificial intelligence market it can supply with them. This month’s agreements show Google, Broadcom, and Anthropic are already planning on that horizon. (reuters.com) (anthropic.com)
