Google has entered discussions with Marvell Technology to co-develop a memory processing unit designed to complement existing tensor processing units and a new inference-focused TPU. Both companies aim to finalize the memory chip’s design by next year, with test production to follow upon agreement.

Google recently deepened its custom chip collaboration with Broadcom to secure additional TPU production capacity. This extension follows a major $2.3 billion custom AI processor deal by a leading hyperscaler last year, underscoring surging demand for specialized AI hardware.

These partnerships signal Google’s intent to diversify its AI hardware supply beyond a single vendor, reducing concentration risk and boosting production flexibility. The moves position Google to meet escalating demand for inference and training workloads across cloud and on-device applications.