Google seeks to collaborate with Marvell to develop AI inference chips, accelerating its move away from dependence on Broadcom.

```

Google is seeking to partner with chip designer Marvell Technology to develop two new chips customized specifically for AI inference workloads. This move represents Google's latest effort to systematically reduce its reliance on long-term partner Broadcom and reflects the surging demand for inference chips in the AI industry.

According to a report from The Information on the 19th, sources revealed that Google's negotiations with Marvell involve two chips: one is a memory processing unit (MPU) designed to work with Google’s Tensor Processing Unit (TPU), and the other is a new TPU designed for running AI models. Unlike Google’s previous purchases of off-the-shelf chips from Marvell, this cooperation aims to create custom semiconductor products tailored for Google.

These negotiations have direct implications for the chip market landscape. Broadcom’s stock price faces potential pressure—even though Broadcom this month signed a new agreement with Google extending to 2031, Google’s strategy to diversify its suppliers is becoming increasingly clear. At the same time, Marvell is expected to further expand its custom chip business, which is currently its fastest-growing segment.

Two new chips: clear division of labor, aimed directly at inference efficiency

Sources say the memory processing unit developed jointly by Google and Marvell will work together with existing TPUs, dynamically allocating AI workloads between the two types of chips according to the differences in computing and memory requirements. This design approach stems from the inherent heterogeneity of inference tasks—some steps of generating responses require immense computing power, while others are restricted by the speed at which the chip can read and write memory, making it difficult for a single processor to handle both aspects optimally.

Google plans to produce nearly 2 million memory processing units, but sources note this number may change as negotiations are still at an early stage. For reference, Morgan Stanley estimates Google’s TPU output in 2027 to be around 6 million units. Both sides aim to finalize the design of the memory processing unit as early as next year, with pilot production to follow.

The design timeline and planned output for the second chip—a new TPU built specifically for inference scenarios—are still unclear. Google currently produces chips at TSMC, but whether the new chip will also be manufactured by TSMC remains to be determined.

Inference chip arms race accelerates, Nvidia’s LPU as a catalyst

Google is accelerating this partnership in part due to competitive pressure from Nvidia.

According to a Google employee, Google had already planned to develop inference chips, but after Nvidia released its Language Processing Unit (LPU) at its GTC conference in March this year, Google quickly sped up related work. Nvidia’s LPU is based on technology licensed from startup Groq for $2 billion, and Marvell was Groq’s first-generation LPU chip design partner—meaning Marvell already has hands-on experience designing inference chips.

The boom in demand for inference chips is fundamentally driven by the evolution of AI product forms. As more complex AI applications such as autonomous agents come to market, their computing requirements far exceed those of traditional chatbots. Recently, OpenAI signed an inference chip procurement contract worth more than $20 billion with Cerebras, and is also developing its own inference chip jointly with Broadcom. The entire industry is accelerating its layout in this area.

Diversifying away from Broadcom: clear strategy but constrained progress

Google’s partnership discussions with Marvell are part of its ongoing efforts since 2023 to diversify its suppliers. Broadcom had long been Google’s sole design partner for TPUs and charged licensing fees based on each TPU produced. With a surge in TPU demand, Google’s payments to Broadcom have also soared, which is a key motivation for seeking alternatives.

Last year, Google brought MediaTek into the design and production of TPU chips. The negotiations with Marvell further expand Google’s network of partners. Google had previously purchased CXL controller chips from Marvell, used to manage memory sharing among data center servers. This cooperation established a foundation of mutual trust between the two sides.

However, Broadcom’s position remains difficult to shake in the short term. This month Broadcom signed a new agreement with Google to provide customized TPUs and network components for Google’s next-generation AI data center racks, extending the partnership to 2031. This indicates that Google’s diversification strategy is more about adding new options to the existing structure, rather than making a complete switch.

For Marvell, deeper cooperation with Google means its custom chip business could secure another major client as an endorsement. Marvell’s main business covers standard networking, storage, and optical interconnect chips for data centers, but in recent years, its custom chip business—helping clients design exclusive chips—has become its fastest-growing segment.

The commercialization of Google’s TPUs also expands the market potential for this partnership. Last year, Google began leasing TPUs to customers outside its own data centers, directly challenging Nvidia’s dominance in the AI chip market, with Anthropic, Meta, and Apple all becoming TPU customers. If the new inference chip development succeeds, its potential market won’t be limited to Google’s internal needs.

Risk Warning and DisclaimerThe market carries risks, and investments must be made cautiously. This article does not constitute personal investment advice and does not consider the special investment objectives, financial conditions, or needs of individual users. Users should consider whether the opinions, viewpoints, or conclusions in this article are suitable for their specific situation. Investment is at your own risk. ```