Google is currently negotiating with Marvell to develop new AI chips focused on inference, in a step aimed at reducing dependence on Nvidia and strengthening control over performance and profit margins. This news led to a rise in Marvell's stock in financial markets.
Development plans include two chips: the first will act as a memory processing unit alongside current tensor processors, and the second will be a new optimized TPU for running language models more efficiently. The data center CPU market is currently estimated at about $100 billion, with expectations of adding an additional $32.5 to $60 billion by 2030 thanks to Agentic AI.
Inference is the new battlefield in AI infrastructure, as focus shifts from training massive models to serving them globally. This alliance indicates that cloud service providers are still trying to build their own solutions to control performance and costs.
