Google and Marvell Team Up to Develop Next-Generation AI Chips for Cloud Inference
Google has entered into advanced discussions with semiconductor company Marvell Technology to jointly develop two new artificial intelligence chips — a next-generation Tensor Processing Unit (TPU) and a dedicated memory processing unit — in a move designed to diversify Google's AI chip supply chain and improve inference performance across its cloud infrastructure.
Background
The race to develop faster, more efficient AI chips has become one of the defining technology competitions of the decade. While Nvidia has dominated the market for AI training chips, the inference segment — where trained models are deployed to generate responses — has become an increasingly contested battleground, with major cloud providers seeking to reduce their dependence on any single supplier.
Google has long developed its own TPUs for internal use, but the scale of AI workloads has grown so dramatically that the company is now seeking to expand its hardware partnerships. Marvell, known for its custom silicon and data infrastructure chips, has emerged as a key partner in this effort.
Key Developments
According to reports, the partnership would see Marvell contribute its expertise in custom chip design and advanced packaging to the development of a new TPU generation optimised for inference tasks. The memory processing unit, meanwhile, would address one of the key bottlenecks in large language model deployment — the speed at which data can be moved between memory and processing cores.
The collaboration is part of a broader strategy by Google to work with multiple chip partners, reducing its exposure to supply chain disruptions and giving it greater flexibility in optimising hardware for specific AI workloads. The company has also been expanding its partnerships with other semiconductor firms as part of this diversification effort.
Why It Matters
Custom AI chips can deliver significant performance and cost advantages over general-purpose graphics processing units, particularly for inference workloads that run continuously at massive scale. For Google, developing more efficient inference hardware could translate into substantial savings in data centre operating costs and a stronger competitive position against Amazon Web Services and Microsoft Azure.
The partnership also signals Marvell's growing importance in the AI chip ecosystem, potentially boosting its revenue and market position in a rapidly expanding segment.
What's Next
The chips are expected to enter production in the coming years, with Google likely to deploy them first in its own data centres before potentially offering them to external cloud customers. The partnership is subject to finalisation of commercial terms, and both companies are expected to make a formal announcement in the near future.
Sources: TechStartups
