Apr. 20 at 12:04 AM
$GOOGL in talks w/
$MRVL to co-develop 2 custom chips: a memory processing unit (MPU) designed to pair w/ existing TPUs, & a new TPU optimized for inference
The move would add a 3rd design partner alongside
$AVGO & MediaTek, reflecting a strategy of supply-chain diversification & cost-targeted silicon for inference workloads
Inference is now Google’s #1 cost center
Training is episodic & inference is continuous - Serving Gemini, Search, Ads, Workspace, YouTube, Android, etc. is a perpetual cost firehose
$GOOGL is building a 3-vendor TPU stack
Broadcom is elite at compute‑dense accelerators, Mediatek is a cost‑optimized TPU “e” variants & Marvell is elite at memory‑centric companion chips
Google is adding Marvell b/c the future of AI hardware is heterogeneous: compute dies + memory dies + networking dies. Marvell is the memory‑system specialist Google needs to make inference cheap enough to scale Gemini everywhere