Google + Marvell talks

Google is in talks with Marvell Technology to develop two custom artificial intelligence chips, adding a new partner to its in-house silicon push. (reuters.com) Reuters, citing a report from The Information on April 20, said the discussions cover two chips meant to run AI models more efficiently. One is described as a memory processing unit that would work alongside Google’s Tensor Processing Units, and the other as a Tensor Processing Unit tailored for inference, the step where a trained model answers prompts. (reuters.com) Google has spent more than a decade building Tensor Processing Units, or TPUs, for its own services and for Google Cloud customers. In April 2025, Google introduced Ironwood, its seventh-generation TPU, and called it its first chip designed specifically for inference. (blog.google) Inference is the expensive part of generative artificial intelligence that happens after training, when a model has to respond to millions of user requests. Google’s TPU7x documentation says Ironwood is built for large-scale training and inference, with pods of up to 9,216 chips and 192 gigabytes of high-bandwidth memory per chip. (docs.cloud.google.com) The reported Marvell talks point to a narrower problem inside that workload: moving data fast enough to keep AI chips busy. Google’s own TPU7x documentation says high-bandwidth memory can still become a bottleneck on memory-bound operations, which helps explain interest in a separate memory-focused chip. (docs.cloud.google.com) For Marvell, the attraction is its custom-silicon business with cloud companies that want chips built for one buyer instead of sold off the shelf to everyone. Marvell told investors at a custom AI event that it sees a growing opportunity in custom silicon for next-generation AI infrastructure. (marvell.com) Investors treated the report as a meaningful design opportunity. CNBC reported on April 20 that Marvell shares rose after the news, while Broadcom shares fell, reflecting expectations that Google could spread more of its AI chip design work across multiple suppliers. (cnbc.com) Google already uses more than one chip partner. Reuters said a Marvell deal would make it a third design partner in Google’s custom silicon supply chain alongside Broadcom and MediaTek, a sign that the company is trying to expand capacity as demand for AI computing keeps climbing. (reuters.com) Nothing has been announced by Google or Marvell, and Reuters said it could not immediately verify the report independently. For now, the clearest read is that Google is still spending heavily on custom chips, but is getting more specific about where training ends and inference begins. (reuters.com)