Broadcom expands hyperscaler chip deals
Broadcom expanded partnerships to supply custom next‑generation AI silicon to hyperscalers including Google and Anthropic, signalling a move toward hyperscaler‑specific accelerators instead of one‑size‑fits‑all GPUs. The deals reflect hyperscalers' appetite for bespoke silicon that matches their workload and infrastructure requirements. That shift tightens the link between datacenter systems engineering and custom chip development. (x.com)
Broadcom just locked itself deeper into the wiring of artificial intelligence data centers, not by selling one standard chip to everyone, but by helping Google build future versions of its own chips through 2031 and by expanding a separate capacity deal tied to Anthropic. Google’s chips are called Tensor Processing Units, which are custom-built artificial intelligence accelerators made for training and running models instead of drawing video game graphics. Google says these chips power Gemini and services like Search, Photos, and Maps, and it has been building them for more than a decade. The new Broadcom deal means Google is not just buying parts off a shelf. Reuters reported on April 6, 2026 that Broadcom will develop and supply future generations of Google’s custom artificial intelligence chips and other components for Google’s next-generation artificial intelligence racks through 2031. Anthropic is connected to that same hardware pipeline in a different way. Broadcom said Anthropic will get access starting in 2027 to about 3.5 gigawatts of computing capacity that draws on Google’s artificial intelligence processors. A gigawatt is power-plant scale, so 3.5 gigawatts is not a small cloud contract or a few extra servers in a warehouse. Anthropic called it its “most significant compute commitment to date” and said most of the new capacity will be located in the United States. This is the business shift underneath the headline: the biggest cloud companies increasingly want chips shaped around their own workloads, their own software, and even their own buildings. Broadcom has been pitching that model by arguing that custom chips can improve performance and lower total cost when the memory bandwidth, input-output bandwidth, and package are tuned for one customer’s jobs. That tuning now goes beyond the chip itself and into the box around it. Google’s Trillium Tensor Processing Unit is part of what Google calls its Artificial Intelligence Hypercomputer, which combines chips, software, networking, and scheduling into one system and can deploy more than 100,000 Trillium chips on a single Jupiter network fabric. Broadcom has been building toward exactly that kind of system-level sale. In December 2024 it introduced a packaging platform that can combine more than 6,000 square millimeters of silicon with up to 12 stacks of high-bandwidth memory in one device, aimed at next-generation custom accelerators for huge clusters. Google gets a long runway for its in-house chip program, and Anthropic gets a giant reserved lane on that road without having to own the whole highway itself. Anthropic said it still uses Amazon Web Services Trainium, Google Tensor Processing Units, and Nvidia graphics processing units, with Amazon remaining its primary cloud provider and training partner. The pressure behind all of this is demand. Reuters reported that Anthropic’s run-rate revenue has passed $30 billion in 2026, up from about $9 billion at the end of 2025, and Anthropic said the number of business customers spending more than $1 million annually doubled from over 500 in February to more than 1,000 by April. For years, the default answer for artificial intelligence computing was “buy more Nvidia graphics processing units.” This week’s Broadcom deals point to a different future: Google keeps designing its own engine, Broadcom helps build the machine around it, and model companies like Anthropic reserve industrial-scale access to that stack years in advance.
