Summary: Google is engaging with Marvell Technology to produce two innovative AI chips—a memory processing unit and an inference-specific TPU—marking a significant expansion in its custom silicon supply chain alongside Broadcom and MediaTek. The ongoing discussions, which have yet to lead to a formal contract, follow Broadcom's recent long-term agreement for TPU development and highlight Google's strategic pivot toward inference as a key computing cost factor, with the custom ASIC market projected to grow by 45% by 2026, reaching an estimated $118 billion by 2033.
According to reports, Google is negotiating with Marvell Technology to create two types of chips tailored for AI applications. The first is a memory processing unit that is expected to complement Google’s existing Tensor Processing Units (TPUs). The second chip is a new TPU, specifically optimized for inference tasks, which are crucial for delivering AI services to users instead of just training the models. Marvell's role would be similar to that of MediaTek, which recently collaborated with Google on the latest Ironwood TPU, providing design services. However, no formal agreement has been established yet.
This development comes shortly after Broadcom, Google’s primary custom chip supplier, announced a long-term agreement to continue supplying TPUs and networking components through 2031. This timing suggests that Google is not eliminating Broadcom from its supply chain but rather adding Marvell as a third design partner. This multi-supplier strategy allows Google to diversify its partnerships while maintaining Broadcom for high-performance chip variants and MediaTek for more cost-effective options, thereby enhancing its overall supply chain resilience.
Importance of Inference Technology
Google has recently launched its seventh-generation TPU, dubbed Ironwood, which the company describes as “the first Google TPU for the age of inference.” Ironwood boasts ten times the peak performance of the previous TPU v5p and is capable of scaling to 9,216 liquid-cooled chips in a superpod configuration that consumes about 10 megawatts of power, delivering an impressive 42.5 FP8 exaflops. Google aims to produce millions of Ironwood units this year, and the chips designed by Marvell would serve as a supplement, potentially targeting different workloads or cost structures as the demand for inference-driven AI computation rises.
The shift from model training to inference as the primary driver of demand is significantly altering the landscape of the chip market. Training a frontier AI model is a massive, resource-intensive task that can take weeks or months, whereas inference is a continuous process that serves real-time queries from users, scaling costs dynamically with demand. As the reach of AI products expands to hundreds of millions of users, the cost associated with inference becomes the dominant expense, emphasizing the need for specialized inference silicon that can outperform general-purpose GPUs in terms of cost and efficiency.
Historical Context of Google and Marvell Collaboration
The collaboration between Google and Marvell is not a new development. Reports indicate that Google has been exploring a chip codenamed “Granite Redux” since 2022, intending to leverage Marvell’s capabilities instead of solely depending on Broadcom, which could potentially save the company billions annually. At that time, Google praised Broadcom as an excellent partner, indicating ongoing collaborations with multiple suppliers for the long haul.
However, the current landscape reflects a shift in strategy, where Google appears to have solidified its partnership with Broadcom through the newly locked-in agreement, opting instead for a diversified supplier model. This strategic framework allows Google to distribute its TPU program across several suppliers, similar to how automotive companies manage their component suppliers, mitigating the risk of any single vendor gaining excessive leverage.
Marvell’s Strategic Position
Marvell has seen significant growth, with its data center revenue reaching a record $6.1 billion in its fiscal year ending February 2026, reflecting a 42% year-over-year increase. The company has established a robust custom silicon business with a $1.5 billion annual run rate, securing design wins with major cloud providers like Amazon, Microsoft, and Meta, alongside its previous work with Google on the Axion ARM CPU. Additionally, Nvidia’s recent investment of $2 billion in Marvell and their partnership to integrate custom chips with Nvidia’s interconnect fabric positions Marvell favorably within both the GPU and ASIC ecosystems.
Furthermore, Marvell’s acquisition of Celestial AI for up to $5.5 billion in December 2025 enhances its photonic interconnect technology, which aims to create a comprehensive connectivity platform for AI and cloud customers. The CEO of Marvell, Matt Murphy, targets a 20% market share in custom AI chips and anticipates around 30% revenue growth in fiscal 2027.
Broadcom’s Competitive Stance
Despite these developments, Broadcom maintains a strong position in the market, commanding over 70% of the custom AI accelerator market. The company reported AI revenue of $8.4 billion in its most recent quarter, marking a 106% increase year-over-year, with projections of $10.7 billion for the upcoming quarter. Broadcom’s ambitions include reaching $100 billion in AI chip revenue by 2027. Following the announcement of the Google extension, Broadcom’s stock rose by more than 6%, with analysts forecasting significant revenue growth attributed to its partnerships with Google and Anthropic.
As the custom ASIC market continues to expand, with projections indicating a 45% increase in 2026 compared to only 16% growth in GPU shipments, Broadcom is expected to maintain a leading market share, while Marvell aims to capture approximately 25% by 2027. The overall market is anticipated to reach $118 billion by 2033.
Implications for Google
Google’s chip strategy now involves collaboration with four partners—Broadcom, MediaTek, Marvell, and TSMC—as well as its in-house design team, encompassing a wide range of products that address training, inference, and general cloud compute needs. This complexity is intentional, as it shields Google from the risks associated with relying on a single chip supplier, such as pricing fluctuations and supply chain vulnerabilities.
The focus on inference in the discussions with Marvell indicates a strategic shift in funding allocation. While Nvidia’s latest chips dominate in training tasks, inference represents the bulk of operational volume, where the cost advantages of custom silicon can significantly increase. Google serves billions of AI-driven search queries, Gemini conversations, and Cloud AI API requests daily. Even marginal cost reductions in inference can translate into substantial savings annually, aligning with the goals outlined in the previous “Granite Redux” discussions.
While the negotiations with Marvell are still ongoing, and any new chip development will likely take years to materialize, the trajectory is unmistakable. Google is actively constructing a chip supply chain capable of meeting the most demanding AI inference workloads globally, ensuring that it has multiple partners to develop the necessary silicon. Securing a contract with Marvell for an inference TPU would solidify Marvell’s position as a leading custom AI chip designer, while also providing Google with additional leverage in a competitive market where dependence on a single supplier is no longer viable.