Google reportedly taps Marvell Technology to co-design custom AI chips, signaling shift away from Broadcom partnership

Google is in negotiations with chip designer Marvell Technology to develop new custom AI processors, including a memory processing unit to support Google's Tensor Processing Units (TPUs — specialized chips that run AI models), according to reporting by The Information. The companies plan to complete the memory chip design next year and move to test production.

Marvell's memory controller chip (called Structera) moves memory storage physically closer to AI accelerators, reducing the distance data must travel and boosting the speed at which AI models can process information. This architecture matters because it reduces a bottleneck: today, retrieving data from memory is often slower than the AI chip itself can compute.

For data center and cloud engineers, this partnership could mean faster, cheaper AI inference (the step where an AI produces answers to queries) if Marvell succeeds — Google may eventually sell these in-house chips to customers who currently buy from competitors like NVIDIA. For Marvell shareholders, the deal could accelerate the company's goal of capturing 20% of the custom AI chip market by 2028, up from less than 5% in 2023.