Summary: Google is in discussions with Marvell Technology to create two innovative AI chips—a memory processing unit and an inference-optimized TPU—marking a significant addition to its custom silicon supply chain alongside Broadcom and MediaTek. These talks, which have yet to result in a signed agreement, follow closely after Broadcom secured a TPU agreement through 2031, underscoring Google's strategic pivot towards inference as a key computing cost driver. The custom ASIC market is anticipated to expand by 45% by 2026, potentially reaching $118 billion by 2033.
According to reports, Google is engaging with Marvell Technology to design two new chips aimed at enhancing AI model performance. One of these is a memory processing unit, intended to complement Google's existing Tensor Processing Units (TPUs). The second chip is a new TPU, optimized specifically for inference tasks—where models serve end-users rather than engage in data learning. Marvell's role would primarily focus on design services, similar to its collaboration with MediaTek on Google's recent Ironwood TPU. As of now, these discussions have not led to a formal contract.
The timing of these negotiations coincides with Broadcom's announcement of a long-term agreement to supply TPUs and networking components until 2031. This suggests that Google is not looking to replace Broadcom but rather to diversify its chip design partnerships, adding Marvell as a third player in a lineup that already includes Broadcom for high-performance versions and MediaTek for cost-effective alternatives. This diversification strategy aims to strengthen Google's supply chain rather than eliminate existing partnerships.
Importance of Inference in AI
Recently launched, Google's seventh-generation TPU, Ironwood, has been labeled as 'the first Google TPU for the age of inference.' It boasts a peak performance that is ten times greater than the TPU v5p, with capabilities to scale up to 9,216 liquid-cooled chips in a superpod, consuming around 10 megawatts and achieving 42.5 FP8 exaflops. Google plans to manufacture millions of Ironwood units this year. The chips designed in collaboration with Marvell would not serve as replacements but rather as enhancements, potentially catering to different workloads or cost structures, given the increasing proportion of Google's computational resources dedicated to serving AI models as opposed to training them.
The transition from training-focused models to inference-driven demands is reshaping the semiconductor landscape. Training a cutting-edge model can be a resource-intensive endeavor, requiring substantial computational power over extended periods. In contrast, inference operates continuously, catering to user queries and scaling costs with demand. As AI technologies gain traction among hundreds of millions of users, inference becomes a critical expense, positioning purpose-built inference chips as a competitive edge over general-purpose GPUs in terms of cost efficiency and performance.
Background of the Google-Marvell Collaboration
The partnership between Google and Marvell extends beyond the current discussions, with earlier reports indicating that Google had been exploring a chip codenamed 'Granite Redux' leveraging Marvell's capabilities instead of Broadcom's, aiming for significant cost savings. At that time, Google characterized Broadcom as 'an excellent partner,' emphasizing its commitment to long-term relationships with multiple suppliers.
Since then, it appears Google has pivoted from the idea of entirely replacing Broadcom. The recent agreement securing Broadcom's involvement through 2031 solidifies this relationship. Instead, Google is constructing a multi-supplier architecture where Broadcom, MediaTek, and potentially Marvell each contribute unique strengths to the TPU initiative. This collaborative approach mirrors strategies used in the automotive industry, where no single vendor holds enough leverage to dictate terms.
Marvell's Role and Growth
Marvell has reported impressive growth, with data center revenue reaching a record $6.1 billion for the fiscal year ending February 2026, and total revenue climbing to $8.2 billion—a 42% increase year-over-year. The company maintains a robust custom silicon business with an annual run rate of $1.5 billion, comprising 18 cloud-provider design wins, including chips for major players like Amazon, Microsoft, and Meta, alongside its existing collaborations with Google on the Axion ARM CPU.
Furthermore, Nvidia's $2 billion investment in Marvell at the end of March has strengthened their partnership through NVLink Fusion, integrating Marvell's custom chips with Nvidia's interconnect fabric. This positions Marvell at a pivotal intersection in both GPU and ASIC markets. Marvell's acquisition of Celestial AI for up to $5.5 billion in December 2025 further enhances its capabilities, adding photonic interconnect technology aimed at delivering a comprehensive connectivity platform for AI and cloud clients. Marvell's CEO anticipates capturing 20% of the custom AI chip market, with a projected 30% revenue growth in fiscal 2027.
Broadcom's Competitive Edge
Despite the discussions with Marvell, Broadcom's dominant position in the custom AI accelerator market remains unshaken. The company holds over 70% market share, with AI revenue soaring to $8.4 billion in its latest quarter—an increase of 106% year-over-year. Broadcom forecasts $10.7 billion in AI revenue for the upcoming quarter and is targeting $100 billion by 2027. Following the announcement of Google's agreement, Broadcom's stock surged more than 6%, and analysts predict significant future AI revenue growth from its partnerships with Google and Anthropic.
The broader ASIC market is anticipated to outpace the GPU market, with projections indicating a 45% increase in custom chip sales by 2026, compared to a 16% growth in GPU shipments. Research suggests that by 2027, Broadcom will command approximately 60% of the custom AI accelerator market, with Marvell capturing around 25%. The overall market is expected to reach $118 billion by 2033.
Implications for Google
With its chip strategy now involving four partners—Broadcom, MediaTek, Marvell, and TSMC—alongside its in-house design team, Google is poised to create a diverse product line that spans training, inference, and general-purpose cloud computing. This complexity is intentional, as it mitigates risks associated with relying solely on a single chip supplier, a vulnerability that could impact pricing and supply.
The emphasis on inference in the discussions with Marvell reflects a significant shift in financial priorities. Although Nvidia’s latest chips dominate training workloads, inference represents the bulk of demand, where custom silicon can offer substantial cost advantages. With billions of AI-enhanced search queries and Cloud AI API calls to serve daily, even marginal cost reductions in inference can translate to billions saved annually, aligning with the goals of the earlier 'Granite Redux' discussions.
While the negotiations with Marvell have not yet culminated in a deal, the timeline for chip development indicates that any resulting products could take years to materialize. However, the trajectory is unmistakable: Google is constructing a robust chip supply chain designed to handle the most demanding AI inference workloads globally, with a clear intention to engage multiple partners capable of delivering the necessary silicon. For Marvell, securing a Google inference TPU contract would solidify its status as a key player in the custom AI chip landscape, while for Google, it represents an essential step towards ensuring resilience in a competitive market.