Google and Marvell in talks to co-develop custom AI inference chips

Google is reportedly in discussions with semiconductor firm Marvell Technology to develop two new types of custom artificial intelligence chips, a move that signals a significant expansion of Google's internal silicon strategy .

 According to reports from The Information on April 20, 2026, the potential partnership aims to enhance the efficiency of "inference"—the process where an AI model actually runs and responds to user queries, rather than the initial training phase, News.Az reports, citing Economic Times.

This shift reflects a broader industry trend where tech giants are moving beyond general-purpose hardware to create specialized components that can handle the massive computational demands of modern AI at a lower cost.

The collaboration is said to focus on two distinct pieces of hardware. The first is a dedicated Memory Processing Unit (MPU) designed to work in tandem with Google’s existing Tensor Processing Units (TPUs). This MPU would specifically tackle the memory bottlenecks that often slow down AI performance, allowing for faster and more efficient data handling. The second chip under discussion is a new version of the TPU itself, specifically optimized for inference workloads. By splitting these functions, Google hopes to create a more modular and high-performance ecosystem for its Google Cloud services and internal AI products like Gemini.

This move positions Marvell as a key third partner in Google’s custom chip supply chain, joining established players like Broadcom and MediaTek. While Broadcom has historically been Google's primary partner for high-performance TPU development, bringing Marvell into the fold allows Google to diversify its vendors and potentially reduce its reliance on any single manufacturer. The news has already had a significant impact on the market, with Marvell’s stock surging as investors recognize the company’s growing footprint in the highly lucrative custom ASIC (Application-Specific Integrated Circuit) market, which is projected to grow substantially by the end of the decade.

As Google continues its quest to provide a viable alternative to Nvidia's dominant GPUs, these custom chips are becoming a cornerstone of its competitive strategy. If the talks lead to a finalized contract, the design of the new memory processing unit could be completed as early as next year, with test production to follow. For the broader tech industry, this partnership underscores a new era of "bespoke silicon," where the biggest cloud providers are no longer just customers of chipmakers, but active co-architects of the hardware that powers the next generation of artificial intelligence.

News.Az 

By Leyla Şirinova