Broadcom joins big AI supply deals
Reports say Broadcom entered strategic collaborations with Anthropic and Google Cloud under a 3.5‑gigawatt infrastructure arrangement to support large TPU clusters for model training, highlighting chip and networking firms as central players in AI capacity co‑design. The story underscores that AI infrastructure procurement is now an ecosystem negotiation — compute, networking and power are being contracted together. (markets.financialcontent.com)
Broadcom just turned an artificial intelligence deal into a power deal: in a filing on April 6, it said Anthropic will get about 3.5 gigawatts of next-generation Google Tensor Processing Unit capacity through Broadcom starting in 2027. Broadcom also said it will keep supplying Google’s future Tensor Processing Units and networking parts for Google’s next artificial intelligence racks through 2031. (sec.gov) That 3.5-gigawatt figure is the kind of number utilities use for cities and industrial zones, not the kind software companies used to talk about in public. Broadcom’s filing says the three companies are already talking with operating and financial partners to make the deployment happen. (sec.gov) Anthropic confirmed the agreement the same day and said the new capacity will start coming online in 2027. The company called it its biggest compute commitment yet and said most of the new buildout will be in the United States. (anthropic.com) The reason Anthropic is locking in power this far ahead is simple: its customer growth is running faster than normal cloud buying cycles. Anthropic said its revenue run rate has passed $30 billion in 2026, up from about $9 billion at the end of 2025, and the number of business customers spending more than $1 million a year has risen from more than 500 in February to more than 1,000 now. (anthropic.com) Google’s Tensor Processing Unit is a custom artificial intelligence chip, built by Google for neural-network math the way a race car is built for one track. Google Cloud says these chips are designed specifically for training and running artificial intelligence models, which is why companies like Anthropic want guaranteed access years in advance. (cloud.google.com) Broadcom sits in the middle because it is not just selling one chip off a shelf. In its April 6 filing, Broadcom said it has a long-term agreement to develop and supply Google’s future Tensor Processing Units, plus a separate supply-assurance deal for the networking gear inside Google’s next-generation artificial intelligence racks. (sec.gov) That networking piece is easy to miss, but it decides whether thousands of chips act like one machine or like a traffic jam. Broadcom’s own March results showed why this matters: the company said first-quarter artificial intelligence revenue reached $8.4 billion, up 106 percent from a year earlier, driven by custom accelerators and artificial intelligence networking. (broadcom.com) Anthropic is also making a point of not betting on one supplier. Its April 6 post said Claude runs on Amazon Web Services Trainium, Google Tensor Processing Units, and NVIDIA graphics processing units, while Amazon remains its primary cloud provider and training partner on Project Rainier. (anthropic.com) So this is not one company buying chips from another company anymore. It is an artificial intelligence lab reserving power, a cloud provider reserving data-center space, and a chip-and-networking company reserving the hardware path years before the models are trained. (sec.gov)
