Google in talks with Marvell to co-develop AI inference chips

Alphabet’s Google is reportedly working with Marvell Technology on new custom chips aimed at improving AI inference efficiency, as it seeks to strengthen its in-house silicon strategy and challenge Nvidia’s dominance in AI hardware.

Google is exploring two AI chip designs with Marvell focused on inference and memory optimisation.
The plans include a memory processing unit and a next-generation TPU for inference workloads

What happened

Google is in discussions with Marvell Technology to jointly develop two new AI chips, according to reports cited by Reuters and The Information.

The chips target improved efficiency in running AI models, especially during inference tasks. One design is a memory processing unit that would support Google’s existing tensor processing units (TPUs). The second is a new TPU tailored specifically for inference workloads.

The firms are still in early-stage design talks. They aim to finalise the memory chip design next year. After that, they would move it into test production.

Google has been steadily expanding its custom silicon strategy. Its TPUs already support large-scale AI services in its cloud business. The company also competes with Nvidia in AI infrastructure markets.

Why it’s important

The deal reflects a wider shift in AI computing towards inference efficiency. Training AI models once dominated chip demand. Now, inference is becoming the main workload in real-world applications.

This shift increases demand for specialised chips that reduce latency and energy use. Google aims to reduce reliance on general-purpose GPUs from Nvidia. Instead, it is building a vertically integrated stack of hardware and software.

The collaboration also highlights intensifying competition in AI semiconductors. Companies like Broadcom, AMD, and Marvell are all positioning themselves as custom-chip partners for hyperscalers.

For Google, stronger in-house chips could improve margins in its cloud business. It could also help lock in enterprise customers needing large-scale AI deployment.

At the same time, the move deepens strategic risks. Designing custom silicon is costly and complex. It also increases dependency on supply-chain partners for manufacturing and packaging.

The broader industry trend suggests AI infrastructure spending is still accelerating. Cloud providers continue to invest heavily in specialised compute, even amid concerns about oversupply or slowing demand cycles.

Also read: Nvidia to supply up to one million AI chips to Amazon in cloud deal

Also read: Anthropic considers developing its own AI chips