Google in Talks with Marvell to Develop New AI Inference Chips

Alphabet's Google is in talks with Marvell Technology, Inc. to design two new chips that aim to run artificial intelligence models more efficiently, according to The Information. The discussions involve developing a memory processing unit that will work alongside Google's tensor processing unit and a new TPU designed specifically for inference tasks.

The companies plan to finalize the design of the memory processing unit as early as next year before moving to test production. Google intends to produce about two million of these units, although the figure may change as talks continue. The chips are expected to divide AI workloads between compute and memory demands to improve efficiency.

Google has previously purchased data center chips from Marvell but is now pursuing exclusive semiconductor designs for its own infrastructure. The move reflects Google's effort to diversify its chip design partnerships beyond Broadcom Inc., which has been its primary TPU design partner. Broadcom recently signed a new agreement with Google to continue supplying custom TPUs and networking components through 2031.

Google currently manufactures its chips through Taiwan Semiconductor Manufacturing Co., though it is not yet clear whether the new chips will also be produced there. The collaboration with Marvell would add to Google's expanding custom chip strategy as demand for efficient inference hardware continues to grow.