Google in talks with Marvell to build next-gen AI chips, including a memory processing unit and new TPU, aiming to boost efficiency and rival Nvidia’s dominance in GPUs

Adding Marvell alongside Broadcom is Google hedging TPU supply, not a Nvidia killer — H100/B200 still do 80%+ of hyperscaler training runs. MPUs matter more: HBM3e from Hynix/Samsung is the chokepoint, and compute-in-memory sidesteps it entirely. DePIN GPU networks (Akash, Aethir, io.net) priced on permanent Nvidia scarcity should be nervous — hyperscalers building vertically-integrated memory-compute stacks kills the arbitrage.

Top comment by @Benthic

More coverage

• Google is in talks with Marvell to build custom AI inference chips as it ... (thenextweb.com)
• BREAKING: Google is in talks with Marvell to build 2 new AI chips for ... (x.com)
• Google in Talks With Marvell to Build New AI Chips for Inference (theinformation.com)
• Accelerated Infrastructure for the AI Era - Marvell Technology (marvell.com)
• Four Technologies Are Making Marvell the Next Chip Giant (globalsemiresearch.substack.com)

Explore the topic

• Memory
• Google
• Nvidia
• AI