Google in talks with Marvell on new AI chips

Google is in talks with US chipmaker Marvell Technology to develop two new chips for AI workloads, as it looks to run AI models more efficiently and build an alternative to Nvidia’s GPUs.

One chip would be a memory processing unit designed to work with Google’s tensor processing unit, while the other would be a new TPU built specifically to run AI models.

The companies aim to finish the memory processing unit’s design as early as 2027 before sending it for test production. Google has been pushing TPUs more aggressively as they become a bigger part of cloud revenue.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

A Google partnership would fit Marvell’s focus on big cloud customers

This potential deal fits Marvell’s focus on hyperscalers, very large cloud companies such as Google, Amazon and Microsoft 1. It also matches its AI data center connectivity work, including the Golden Cable active electrical cable (AEC) program 1.
Marvell says it spans custom AI accelerators, chips built for specific AI tasks, plus related infrastructure silicon for hyperscalers 2. It has also cited more than $100 billion in added cloud capital spending by hyperscalers this year alone in a Marvell LinkedIn post 2.
A Google design win could broaden Marvell’s customer mix beyond large programs with Amazon and Microsoft 1. That could ease investor concern about leaning too hard on a small group of hyperscaler customers 1.

Custom silicon has become a battleground for cloud AI infrastructure

Hyperscalers are building bespoke chips to run AI models more efficiently and cut reliance on Nvidia’s GPUs 3.
Google already works with Broadcom and may also be exploring a chip design relationship with Marvell 3.
The shift reaches beyond compute chips as companies work together on more of the data center technology stack.
Marvell’s planned acquisition of Celestial AI signals a push into optical interconnects and optical scale-up fabrics, high-speed optical links that connect many AI chips inside next-generation AI data centers 4. The deal was announced and had not closed at the time of the release 4.