Google is in discussions with Marvell Technology to develop new AI inference chips as part of its collaboration with the Broadcom TPU program.

      **Summary:** Google is negotiating with Marvell Technology to create two new AI chips—a memory processing unit and an inference-focused TPU—adding Marvell as a third design partner alongside Broadcom and MediaTek in its custom silicon supply chain. These discussions, which have not yet resulted in a signed contract, occurred shortly after Broadcom secured a TPU agreement lasting through 2031. This indicates a strategic shift at Google towards inference as the primary compute cost, with the custom ASIC market anticipated to grow by 45% in 2026, reaching $118 billion by 2033.

      According to The Information, Google is collaborating with Marvell Technology to develop two new chips aimed at executing AI models. One of these is a memory processing unit meant to complement Google's existing Tensor Processing Units. The other is a new TPU specifically designed for inference, which is the phase of AI where models interact with users instead of learning from data. Marvell will provide design services, similar to MediaTek's role with Google's latest Ironwood TPU. As of now, no formal contract has been signed.

      These discussions followed Broadcom's announcement of a long-term partnership with Google to design and supply TPUs and networking components through 2031. This timing suggests that Google is not eliminating Broadcom as a partner but is rather expanding its supply chain by adding another design partner. Broadcom will provide high-performance chip variants, MediaTek will supply more cost-effective "e" variants at 20 to 30% lower costs, and TSMC will handle fabrication. This strategy emphasizes diversification rather than replacement.

      **Why inference is critical now:**

      Google has just launched its seventh-generation TPU, Ironwood, which it describes as "the first Google TPU for the age of inference," offering ten times the peak performance of the TPU v5p. It can scale to 9,216 liquid-cooled chips in a superpod that consumes about 10 megawatts, achieving 42.5 FP8 exaflops. Google plans to manufacture millions of Ironwood units this year. The chips designed by Marvell are expected to enhance rather than replace Ironwood, targeting different workload profiles or cost points, catering to the growing percentage of compute dedicated to serving AI models rather than training them.

      The transition from training to inference as the main demand driver is transforming the chip market. Training a top-tier model is a one-time process that needs significant compute power for weeks or months, while inference occurs continuously to handle user queries, with costs increasing alongside demand rather than capability. As AI products reach hundreds of millions of users, inference becomes the largest expense, and specialized inference silicon offers a competitive edge that general-purpose GPUs cannot match in terms of cost or efficiency.

      **The background:**

      The partnership between Google and Marvell dates back further than the recent reports indicate. In 2023, it was reported that Google had been developing a chip codenamed "Granite Redux" using Marvell instead of Broadcom, purportedly aiming to save billions annually. At that time, a Google spokesperson praised Broadcom as "an excellent partner" and noted their productive long-term engagements.

      However, what's changed since 2023 is Google's apparent decision to maintain its collaboration with Broadcom. The long-term agreement solidified that relationship. Rather than completely replacing Broadcom, Google aims to create a multi-supplier framework in which Broadcom, MediaTek, and potentially Marvell manage different aspects of the TPU program, competing in specific areas rather than for the whole contract. This approach is akin to how automotive firms handle their component suppliers, ensuring that no one vendor gains enough power to set terms.

      **What Marvell offers:**

      Marvell's data center revenue hit a record $6.1 billion for the fiscal year ending February 2026, with total revenue of $8.2 billion—a 42% increase compared to the previous year. The company has a custom silicon division with a $1.5 billion annual run rate and has secured design wins with 18 cloud providers, producing chips for firms like 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, collaborating through NVLink Fusion to connect Marvell’s custom chips and networking with Nvidia’s interconnect fabric. This partnership places Marvell at the crossroads of both the GPU and ASIC markets. In December 2025, Marvell acquired Celestial AI for up to $5.5 billion, gaining photonic interconnect technology that CEO Matt Murphy stated would provide "the industry’s most complete connectivity platform for AI and cloud customers." Murphy aims for a 20% share in the custom AI chip sector and anticipates roughly 30% year-over-year revenue growth in fiscal 2027.

      Marvell’s stock has surged about 50% this year, with a 30% increase in