Google and Marvell enter talks to co-develop custom AI chips

8 articles · Updated · The Motley Fool · Apr 29

The discussions cover a memory processing unit and a next-generation TPU for inference, sending Marvell shares up more than 13% in one session.
The move underscores how major cloud groups are building in-house AI silicon to reduce reliance on Nvidia, which still controls about 85% to 92% of the accelerator market.
Marvell says its custom ASIC business reached $1.5bn in annual revenue and has 18 hyperscaler design wins, as it expands further through photonic interconnect technology from Celestial AI.

Are tech giants trading their Nvidia dependency for a new one on Marvell's essential connectivity technology?

As data movement becomes AI's bottleneck, will interconnects soon be more valuable than the chips themselves?