Marvell‑Google chip talks

Google is in talks with Marvell Technology to develop two custom chips for running artificial intelligence models, according to reports published April 19 and April 20. (theinformation.com) (reuters.com) The discussions cover two designs: a memory processing unit that would work alongside Google’s tensor processing units, and a new tensor processing unit built for inference, the stage when a trained model generates answers. Reuters said Marvell shares rose nearly 5% on Monday after the report, and U.S. News said the stock was up about 7% in premarket trading. (reuters.com) (usnews.com) Inference is the expensive part of artificial intelligence that happens every time a user sends a prompt and the model has to respond in real time. Google said when it introduced Ironwood on April 23, 2025 that the chip was its seventh-generation tensor processing unit and its first designed specifically for inference. (blog.google) A memory chip in this setup acts like a faster pantry next to the kitchen: it keeps data closer to the processor so the model spends less time waiting. Marvell has been pitching that approach to cloud customers, saying its custom high-bandwidth-memory architecture can increase compute area by up to 25%, raise memory capacity by 33%, and cut memory-interface power by 70%. (marvell.com) Google has been building its own artificial intelligence chips for more than a decade, but it still relies on outside partners to help design and manufacture them. Reuters reported on April 6 that Broadcom signed a long-term agreement with Google to develop and supply future generations of custom artificial intelligence chips and other rack components through 2031. (reuters.com) That makes the Marvell talks less a break with Google’s chip strategy than an expansion of it. The Information said a Marvell deal would add another supplier to a Google custom-silicon chain that already includes Broadcom, while reports in late 2025 also tied MediaTek to parts of Google’s tensor processing unit roadmap. (theinformation.com) (trendforce.com) The pressure behind all of this is cost. Nvidia’s general-purpose graphics processors still dominate the market for training and serving large models, but cloud companies have been trying to lower power use and reduce per-query costs with chips tailored to their own software and data centers. (blog.google) (reuters.com) Neither Google nor Marvell had announced a deal as of Monday, April 20. For now, the clearest signal is the market reaction: investors treated the report as another sign that the artificial intelligence chip race is shifting from buying more processors to designing more specialized ones. (reuters.com)