Google Partners with Marvell to Co-Develop New AI Chips; TPU Expansion Announced
Google announced it is in talks with Marvell Technology to jointly develop new AI chips including a memory processing unit and inference-optimized TPU, as the company pushes to diversify its chip suppliers and challenge Nvidia's dominance in AI hardware.
Partnership Details
Google is set to unveil its next-generation Tensor Processing Units at Google Cloud Next this week, with dedicated inference chips designed to accelerate the execution of trained AI models, and the company is in talks with Marvell Technology to develop a memory processing unit and inference-optimized TPU, with the push announced in reports on April 20 and building on recent deals supplying TPUs to Meta and Anthropic.
Strategic Significance
Google Chief Scientist Jeff Dean noted the shift toward specialized chips for training versus inference as workloads evolve, and by challenging Nvidia's dominance in a market projected to grow rapidly, Google aims to lower costs and boost performance for its cloud customers.
Broader Market Context
Agentic AI's shift toward CPU and memory spending broadens the AI hardware boom, creating new avenues for investment and innovation for chipmakers and startups beyond today's GPU-centric landscape. The move represents Google's effort to control more of its AI infrastructure stack and reduce reliance on external chip suppliers as demand for specialized AI processors explodes.