The Marvell talks surfaced days after Broadcom locked in a through-2031 agreement to design and supply Google's TPUs. What Google is reportedly discussing with Marvell are two inference-side chips: a memory processing unit to ease data movement alongside existing TPUs, and an inference-optimized TPU. Neither chip touches training. Broadcom holds the training relationship through the end of the decade. What looks like Google diversifying away from merchant silicon is actually Google adding a third ASIC vendor to the inference layer while Broadcom's grip on the higher-value training work stays intact.
NVIDIA invested $2 billion in Marvell last month through NVLink Fusion, partnering to build custom XPUs and NVLink-compatible networking. Marvell is simultaneously the ASIC alternative to NVIDIA and a contracted extension of NVIDIA's own ecosystem. The company is playing both sides, which is a fine business to be in.
The more interesting read on Marvell isn't "NVIDIA threat vector." It's toll collector across competing architectures, collecting rent regardless of which silicon wins. Neither company has confirmed the talks are even finalized.