Google Shifts TPU Partnership to MediaTek for Next-Gen AI Chip Development

In a strategic move to enhance its AI capabilities, Google is reportedly transitioning from Broadcom to MediaTek as its design partner for the development of its seventh-generation Tensor Processing Units (TPUs). As Google continues to innovate in the realm of artificial intelligence, this partnership signifies a shift in its approach to hardware development, aiming for cost efficiency and improved performance in its AI systems.

Google has engaged with Broadcom for the design of its AI accelerator chips, known as Tensor Processing Units (TPUs). It is essential to clarify that these TPUs are distinct from the Tensor Gx application processors found in Pixel devices. Recent reports suggest that Google may replace Broadcom with MediaTek for the development of new TPUs, marking a significant change in Google’s design strategy for its AI chips.

According to the reports, Google is not completely severing ties with Broadcom; however, there are compelling reasons for this potential partnership shift to MediaTek. Primarily, MediaTek maintains a strong relationship with TSMC, the world’s largest semiconductor foundry, which may enable MediaTek to provide chips to Google at a lower cost compared to Broadcom. Last year, Google is estimated to have spent between $6 billion and $9 billion on TPUs, as reported by Omdia.

Google designed the TPU AI accelerators to reduce its reliance on Nvidia’s GPUs, which are currently the predominant chips used for training AI models. The TPU chips, tailored specifically for AI tasks, not only serve Google’s internal workloads but also cater to Google Cloud customers. This strategic move places Google in a distinct position, as it is less dependent on Nvidia compared to other key players like OpenAI and Meta Platforms, who still heavily rely on Nvidia chips, exposing them to risks during chip shortages.

For instance, OpenAI CEO Sam Altman recently revealed that the company faced a shortage of Nvidia GPUs, which compelled them to delay the launch of the new GPT-4.5 model. While the reasons for choosing GPU chips over CPUs for AI modeling are often unclear, GPUs are preferred for their ability to handle extensive data processing simultaneously, aligning well with AI's matrix-style data processing needs. In contrast, CPUs generally operate in a sequential manner, making them less effective for AI applications.