Google is discussing plans with Marvell Technology to jointly develop two new chips that could run AI models more efficiently, The Information reported on April 19, citing two sources.

One of the two chips under discussion is a memory processing unit designed to work with Google's tensor processing units (TPUs). The other is a new TPU specialised for running AI models. The new memory processing unit works alongside a TPU and splits AI workloads based on compute and memory demands.

Google's move reflects surging demand for inference chips to run AI products such as autonomous agents, The Information said.

Earlier, Nvidia unveiled a language processing unit (LPU) at its GTC conference in March that improves the efficiency of inference workloads. Nvidia's LPU is based on Grok technology under a licensing deal worth $20 billion. Google had previously planned to develop a new inference chip, but has accelerated related work since Nvidia released the LPU, The Information said. Marvell was a design partner for the first-generation Grok LPU and has experience in designing inference chips.

Google has bought products from Marvell, but the talks aim for a custom design for Google. The effort is an extension of a move to reduce reliance on Broadcom, with which Google has long cooperated on TPU design, The Information said.

Google and Marvell aim to finalise the design as early as next year and then begin trial production.