Google is reportedly in talks with Marvell Technology to co-develop a new set of custom AI chips focused on inference. The discussions center around building specialized silicon that can handle the massive, always-on demand of running AI models across products like search and cloud services, rather than just training them, reports The Information. If the deal moves forward, it would expand the tech titan’s existing chip strategy beyond its in-house TPUs and long-standing partners, while also positioning the company to rely less on external GPU suppliers like Nvidia.
At the core of the reported discussions are two distinct chip designs. One is believed to be a memory-centric processor intended to complement Google’s existing Tensor Processing Units (TPUs), improving how efficiently data is moved and accessed during AI workloads. Notably, memory bandwidth has become a major limitation in modern AI systems, especially for large language models that need to quickly access huge amounts of stored data.
Meanwhile, the second design is described as a next-generation inference accelerator – a custom TPU variant optimized for serving AI models with lower latency and reduced energy consumption. Such specialization is increasingly important as AI models scale. General-purpose GPUs, while highly flexible, are not always the most efficient solution for repetitive inference tasks. Custom ASICs can be engineered to perform a narrower set of operations far more efficiently, reducing both power usage and cost per query. And for a company operating at Google’s scale, even marginal gains in efficiency can translate into substantial financial impact.
However, the move does not signal a departure from existing associations. Google has worked closely with Broadcom for years on TPU development under a long-term agreement that extends into the next decade. Instead, bringing in an additional partner appears to be part of a broader strategy to reduce single-vendor dependency, improve supply chain resilience, and maintain leverage in pricing and innovation. In an environment where demand for AI hardware is surging, such diversification is increasingly seen as a strategic necessity.
The potential partnership also shows the growing role of custom silicon providers like Marvell in the AI supply chain. Unlike vertically integrated chipmakers, Marvell focuses on designing application-specific integrated circuits customized to the needs of large clients. Its expertise in data center infrastructure, high-speed interconnects, and custom compute solutions makes it a natural partner for hyperscalers aiming to diversify their hardware stack without building every component internally. If the discussions with Marvell lead to a formal agreement, the development process would likely unfold over several phases, including design finalization, prototyping, and testing before full-scale deployment.
It is worth noting that this potential move could impact Nvidia over time. Nvidia still leads the AI chip market, with its GPUs widely used for both training and running AI models, supported by its strong CUDA software ecosystem. However, as companies like Google build their own chips for specific tasks like inference, they may gradually reduce their dependence on Nvidia’s hardware.
The Tech Portal is published by Blue Box Media Private Limited. Our investors have no influence over our reporting. Read our full Ownership and Funding Disclosure →
Ashutosh is a Senior Writer at The Tech Portal, largely reporting on new tech, and intersection of technology and business. Ashutosh’s career spans across nearly a decade of technology writing across multiple platforms and languages.
