Google–Marvell Talks

Alphabet’s Google is in talks with Marvell Technology to help develop two custom artificial-intelligence chips aimed at running models more efficiently. (theinformation.com) The Information reported Sunday, April 20, that the discussions cover a memory processing unit to work alongside Google’s Tensor Processing Unit, or TPU, and a separate TPU tuned for inference, the step when a trained model answers prompts. Reuters said Marvell shares rose nearly 5% on Monday after the report. (theinformation.com) (reuters.com) Inference is the expensive, nonstop part of artificial intelligence service: every chatbot reply, search summary, or coding suggestion has to be computed in real time. Google has been building chips for that workload for years, and in April 2025 it introduced Ironwood, its first TPU designed specifically for inference. (blog.google) Google already sells earlier TPU systems for both training and serving through Google Cloud. In August 2023, Google said Cloud TPU v5e was built for “large-scale AI inference” at lower cost, and in February 2026 Alphabet said Google Cloud ended 2025 at an annual revenue run rate above $70 billion, driven by demand for AI products. (cloud.google.com) (blog.google) A memory processing unit is meant to handle the constant movement of model data, a bottleneck that can slow down chips even when raw compute is available. Pairing a memory-focused chip with a TPU can cut pressure on bandwidth and power use in the same way a separate loading dock keeps a warehouse floor from clogging up. (theinformation.com) (cloud.google.com) The talks also point to a broader supplier strategy. Reuters said a Marvell project would add another design partner as Google pushes TPUs as an alternative to Nvidia’s graphics processors, while reports said the move would not replace Google’s existing work with Broadcom, MediaTek, and Taiwan Semiconductor Manufacturing Co. (reuters.com) (gadgets360.com) That matters for Marvell because custom silicon has become one of its main pitches to cloud companies building their own artificial-intelligence infrastructure. Marvell says it is building a platform for “custom AI infrastructure,” and investors treated the Google report as a sign it could win a larger role in hyperscale data centers. (marvell.com) (finance.yahoo.com) The companies have not announced a deal, and Reuters said it could not immediately verify The Information’s report. For now, the clearest signal is where Google appears to be spending: not just on bigger models, but on cheaper, faster hardware to serve them at scale. (reuters.com)