Google and Marvell: Two Million Custom AI Chips

Google has engaged in discussions with Marvell Technology to design two new specialized chips intended for its data centers, according to information reported by The Information. One of these chips is a memory processing unit, designed to work in tandem with Google's internal TPU (Tensor Processing Units) to efficiently distribute AI tasks based on specific computational and memory needs. The second chip is a new TPU, specifically developed for inference, which refers to the execution of already trained AI models.

Google plans to produce nearly two million of these memory processing units, with a design expected to be finalized by next year. Marvell appears to be a strategic partner for Google, having already designed the first inference chip for the startup Groq. In December 2025, Nvidia acquired Groq's LPU (Language Processing Unit) technology for $20 billion, using it to launch the Groq 3 LPU and the Groq 3 LPX rack system at the GTC 2026. Jonathan Ross, founder of Groq and former Google engineer, played a key role in the initial development of the TPUs.

This collaboration with Marvell is part of a broader strategy by Google to reduce its dependence on Broadcom, its current chip design partner, which imposes high costs per unit for each TPU produced. Despite this, Broadcom remains a key player, having recently signed a contract with Google through 2031.