Google is in discussions with Marvell Technology to develop new AI inference chips in conjunction with the Broadcom TPU program.

      Summary: Google is negotiating with Marvell Technology to design two new AI chips: a memory processing unit and an inference-optimized TPU. This move adds Marvell as a third design partner, joining Broadcom and MediaTek in Google's custom silicon supply chain. The discussions haven't yet resulted in a signed agreement, following a recent long-term TPU contract with Broadcom extending to 2031. These talks indicate Google's focus on inference as a primary computing cost, with the custom ASIC market expected to experience a 45% growth in 2026, reaching $118 billion by 2033.

      According to The Information, Google is engaged with Marvell Technology to create two new chips for AI model execution. One chip is a memory processing unit designed to work in tandem with Google’s existing Tensor Processing Units, while the second is a new TPU specifically for inference—where models serve users instead of learning from data. Marvell's role would be similar to that of MediaTek in the recent Ironwood TPU project, providing design services. No formal agreement has yet been established.

      These discussions follow shortly after Broadcom, Google’s main custom chip collaborator, announced a significant agreement to develop and supply TPUs and networking components until 2031. This timing indicates that Google is not replacing Broadcom but rather expanding its supply chain with Marvell, which will contribute alongside Broadcom for high-performance chip variants and MediaTek for more cost-effective "e" variants, while TSMC handles manufacturing. Google's strategy is one of diversification rather than replacement.

      Why inference is becoming more significant now

      This month, Google introduced its seventh-generation TPU, known as Ironwood, which it labels as “the first Google TPU for the age of inference.” Ironwood boasts peak performance ten times that of the TPU v5p and can scale up to 9,216 liquid-cooled chips in a superpod, covering approximately 10 megawatts, delivering 42.5 FP8 exaflops. Google aims to produce millions of Ironwood units this year, and the chips designed by Marvell would complement Ironwood rather than replace it, potentially catering to various workload profiles or cost considerations as the proportion of Google's computing tasks shifts toward serving AI models instead of training them.

      The transition from training to inference as the main demand driver is transforming the chip market. Training a cutting-edge model is a significant one-time task requiring extensive computational power over weeks or months. Conversely, inference runs continuously to handle every user query, with costs scaling with demand instead of capability. As AI applications reach hundreds of millions of users, inference becomes the major cost factor, and custom-designed inference silicon offers an edge that general-purpose GPUs struggle to match in terms of cost and efficiency.

      The background

      The collaboration between Google and Marvell dates back further than this week's discussions. In 2023, reports indicated Google had been working on a chip known as “Granite Redux” since 2022, opting for Marvell over Broadcom, anticipating significant annual savings. Despite calling Broadcom “an excellent partner,” Google has now seemingly moved away from the idea of cutting off ties with Broadcom entirely, given the binding agreement through 2031. Instead, Google is constructing a multi-supplier framework where Broadcom, MediaTek, and potentially Marvell each play distinct roles in the TPU program, competing for specific portions rather than the whole project—a strategy reminiscent of the automotive industry's management of component suppliers to prevent any single vendor from having too much influence.

      What Marvell contributes

      Marvell achieved a record $6.1 billion in revenue from its data center in the fiscal year ending February 2026, with total revenue reaching $8.2 billion, marking a 42% year-over-year growth. The company has a custom silicon division generating a $1.5 billion annual run rate across 18 cloud-provider design victories, producing chips for Amazon (Trainium processors), Microsoft (Maia AI accelerator), and Meta (a new data processing unit), in addition to its ongoing collaboration with Google on the Axion ARM CPU.

      In March, Nvidia invested $2 billion in Marvell, establishing a partnership via NVLink Fusion to integrate Marvell's custom chips and networking capabilities with Nvidia's interconnect infrastructure. This positions Marvell at a pivotal point within both GPU and ASIC ecosystems. In December 2025, Marvell acquired Celestial AI for as much as $5.5 billion, gaining photonic interconnect technology, which CEO Matt Murphy stated would provide “the industry’s most complete connectivity platform for AI and cloud customers.” Murphy is targeting a 20% market share in custom AI chips, anticipating approximately 30% year-over-year revenue growth by fiscal 2027.

      Broadcom's standing

      The discussions with Marvell do not seem to have undermined Broadcom's influence. Broadcom holds over 70% market share in custom AI accelerators, with AI revenue reaching $8.4 billion in its latest quarter, representing a