Google discusses AI chips for inference with Marvell, reducing Broadcom dependence

Google is taking the next step in its chip strategy: the company is discussing with Marvell the creation of two specialized AI accelerators to expand its chip development pipeline beyond Broadcom. If negotiations conclude with a deal, Google will gain another custom silicon partner—and will strengthen its bet on cheaper and more mass-market inference, not just the training of large models. According to available information, the subject of negotiations involves two new chips. The first is a memory processing unit, a component designed specifically for memory operations and data transfer within AI systems. The second is a TPU optimized precisely for inference—when a model is already trained and needs to respond quickly to user requests.

For Google, this is a sensitive area: the more requests that flow through Gemini, search, advertising, and cloud services, the more heavily the cost of each response and data center power consumption weigh on the company. Currently, Google is not building its custom silicon ecosystem alone. Broadcom has long been a key partner, and MediaTek also figures in the development chain. The emergence of Marvell as a potential third partner means more than simply expanding the list of contractors.

It is a way to distribute risks, not depend on a single supplier of critical components, and simultaneously test different engineering approaches for different classes of workloads—from training to inference in production services. Importantly, these are still just negotiations: there is no signed contract yet, based on available information. But the very fact of such discussions shows where the market is heading.

Against the backdrop of the generative AI boom, companies are increasingly designing their own chips because universal GPUs remain expensive, scarce, and not always optimal for a specific task. Inference is a particularly attractive target for customization: it is where user-facing products ultimately scale, where latency, request cost, and predictable performance matter.

A separate interest is drawn to the memory processing unit. In AI infrastructure, the bottleneck often becomes not raw computational power, but the speed at which data reaches the accelerator. The more efficiently a system works with memory, the less the compute blocks sit idle and the better one can utilize already installed hardware. Such solutions can be just as important as TPUs themselves: they improve the overall balance of the system and reduce data movement overhead.

For Google, this is a particularly logical step. The company has been developing its TPU line for many years and knows how to tightly integrate hardware with its software, data centers, and cloud infrastructure. New specialized components can help it fine-tune the stack for different scenarios: separately for training, separately for serving a huge number of requests, separately for bottlenecks related to memory. This approach typically yields not only performance gains but also tighter control over cost, delivery schedules, and dependence on external ecosystems.

Another layer of this story is Google's negotiating position. When a company has multiple design partners, it is easier for it to distribute projects by specialization and not tie its entire roadmap to one contractor's pace. In an industry where rolling out the next generation of AI hardware directly impacts cloud margins and product launch speed, such flexibility becomes a strategic asset.

Even without a final contract, the very dialogue with Marvell shows that Google is preparing alternatives in advance and expanding its room for maneuver. In practical terms, this news is important not because Google found yet another chip maker, but because the largest AI companies are moving toward a more granular and specialized hardware architecture. The winner will not be the one with simply more accelerators, but the one who best distributes roles among them. If the Marvell deal goes through, it will be yet another signal: the next phase of the AI race is not only about models, but about the economics of inference, where every watt and every millisecond of improvement becomes a competitive advantage.