A recent report suggests Google is partnering with chip designer Marvell to develop two new custom chips.
This collaboration is focused on creating a Memory Processing Unit (MPU) and a new Tensor Processing Unit (TPU) specifically optimized for AI inference. This isn't just about making faster chips; it's a strategic move to solve one of the biggest challenges in modern AI: the "memory wall." As AI models get larger, the bottleneck is no longer raw computing power but how quickly data can be moved to and from memory. An MPU is designed to sit next to the main processor and handle this data traffic much more efficiently, reducing costs and energy use for every task the AI performs.
This news is particularly interesting because of its timing. First, Google recently solidified a long-term agreement with its primary chip partner, Broadcom, to continue developing future TPUs until 2031. So, why bring in Marvell? This signals a clever diversification strategy. Google appears to be splitting its chip development: Broadcom will focus on the powerful, general-purpose TPUs used for training massive AI models, while Marvell will specialize in cost-efficient hardware for inference—the process of actually using those trained models to generate answers or images.
Second, this strategy is made possible by Google's enormous financial commitment to AI. The company announced a major increase in its capital expenditures (CapEx) for 2026, creating the budget to fund multiple, parallel custom chip projects. This allows them to reduce reliance on a single supplier and tailor hardware for specific needs.
The groundwork for this partnership has been laid over the past year. Events like the massive TPU demand from AI company Anthropic highlighted the urgent need for cheaper inference. Furthermore, Nvidia's recent $2 billion investment in Marvell validated Marvell's expertise and made it easier for their custom chips to integrate with existing systems. Marvell, for its part, has been building the necessary memory-focused technology and proving it can handle large-scale projects.
In essence, Google's reported partnership with Marvell is a targeted strike against the rising costs of AI. By locking in Broadcom for training and bringing in Marvell for specialized inference and memory chips, Google is building a more resilient and economically sustainable "AI factory" for the future.
- TPU (Tensor Processing Unit): A custom-designed microchip (ASIC) developed by Google specifically for accelerating AI and machine learning workloads.
- Inference vs. Training: Training is the process of teaching an AI model by feeding it vast amounts of data. Inference is the process of using that trained model to make predictions or generate new content. Inference is typically less computationally intensive but happens far more frequently.
- Memory Wall: A term describing the growing gap between the speed of processors and the speed of memory. Processors can often compute faster than data can be supplied to them, creating a bottleneck.