Summary: Google is currently negotiating with Marvell Technology to create two innovative AI chips: a memory processing unit and an inference-optimized TPU. This partnership adds a third design collaborator to Google's existing relationships with Broadcom and MediaTek in its custom silicon supply chain. These talks follow Broadcom's recent agreement with Google, extending through 2031, and underscore Google's strategic pivot towards inference as a primary compute cost driver, as the custom ASIC market is anticipated to grow by 45% in 2026, reaching an estimated $118 billion by 2033.
According to reports, Google is engaging with Marvell Technology to design two new chips aimed at enhancing AI model performance. The first chip is a memory processing unit, intended to complement Google’s existing Tensor Processing Units (TPUs). The second is a new TPU specifically tailored for inference, which is the phase where AI models respond to user queries rather than learn from new data. Marvell is expected to take on a design-services role, akin to MediaTek’s contributions on Google’s latest Ironwood TPU. However, these discussions have yet to culminate in a formal agreement.
The timing of these negotiations is noteworthy, as they come shortly after Broadcom, Google’s primary custom chip collaborator, announced a long-term arrangement for designing and supplying TPUs and networking components until 2031. This suggests that Google is not seeking to replace Broadcom but rather to diversify its design partnerships, adding Marvell as a third player in a supply chain that already includes Broadcom for high-performance variants and MediaTek for cost-optimized versions, along with TSMC for manufacturing. This approach signifies a strategy of diversification rather than outright substitution.
Understanding the Importance of Inference
Recently, Google introduced its seventh-generation TPU, dubbed Ironwood, which the company labels as “the first Google TPU for the age of inference.” This new TPU boasts ten times the peak performance compared to its predecessor, the TPU v5p, and can scale up to 9,216 liquid-cooled chips in a superpod, drawing roughly 10 megawatts of power and achieving 42.5 FP8 exaflops. Google plans to manufacture millions of Ironwood units in the current year. The chips designed by Marvell are expected to complement, rather than replace, Ironwood, targeting distinct workload profiles or cost points as an increasing portion of Google’s compute resources shift towards serving AI models instead of training them.
The transition from training to inference as the leading demand driver is transforming the chip market landscape. While training a cutting-edge model is a substantial, one-time task requiring significant computational resources over weeks or months, inference is a continuous operation that processes every user query, leading to costs that scale with demand rather than capability. As AI applications grow to serve hundreds of millions of users, inference becomes the primary expense, making purpose-built inference silicon a competitive edge that general-purpose GPUs cannot match in terms of cost efficiency.
The History of Google and Marvell
The relationship between Google and Marvell extends beyond the recent discussions. Reports from 2023 indicated that Google had been developing a chip codenamed “Granite Redux” with Marvell since 2022, aiming to save billions of dollars annually by opting for Marvell over Broadcom. At that time, Google’s representatives acknowledged Broadcom as a valued partner while noting their productive engagement with multiple suppliers for the long term.
However, it appears that Google has shifted its strategy since 2023, opting not to completely sever ties with Broadcom. The agreement extending to 2031 solidifies that partnership. Instead, Google is constructing a multi-supplier ecosystem where Broadcom, MediaTek, and potentially Marvell will each participate in different segments of the TPU program, competing on specific areas rather than against each other for the entire contract. This methodology resembles the supply chain management utilized by automotive manufacturers, preventing any single vendor from gaining excessive leverage.
Marvell's Contributions and Market Position
Marvell reported a record $6.1 billion in data center revenue for its fiscal year ending February 2026, with total revenue reaching $8.2 billion, reflecting a 42% increase year-over-year. The company has established a custom silicon division with an annual run rate of $1.5 billion, having secured 18 cloud-provider design wins, including collaborations with Amazon (for Trainium processors), Microsoft (for the Maia AI accelerator), and Meta (for a new data processing unit). In addition, Marvell continues its partnership with Google on the Axion ARM CPU.
In March, Nvidia made a significant investment of $2 billion in Marvell, facilitating collaboration through NVLink Fusion to integrate Marvell’s custom chips and networking with Nvidia’s interconnect fabric. This positioning places Marvell at a pivotal juncture within both the GPU and ASIC landscapes. Additionally, Marvell's acquisition of Celestial AI for up to $5.5 billion in December 2025 further enhances its capabilities in photonic interconnect technology, which Marvell's CEO claims will provide “the industry’s most complete connectivity platform for AI and cloud customers.” Marvell aims to capture a 20% market share in custom AI chips and anticipates around 30% annual revenue growth in fiscal 2027.
Broadcom's Dominance and Future Projections
The ongoing discussions with Marvell do not seem to undermine Broadcom’s strong market position. Broadcom boasts over 70% of the market share in custom AI accelerators, with AI revenue reaching $8.4 billion in its latest quarter—a 106% increase year-over-year—and forecasts suggest it could hit $10.7 billion in the upcoming quarter. The company aims for $100 billion in AI chip revenue by 2027. Following the announcement of the Google extension, Broadcom’s share prices surged by more than 6%, and analysts predict that the company could generate around $21 billion in AI revenue from its relationships with Google and Anthropic by 2026, climbing to $42 billion in 2027.
The broader custom ASIC market is expanding at a faster pace than the GPU sector. TrendForce projects custom chip sales to rise by 45% in 2026, in contrast to a 16% growth rate for GPU shipments. Counterpoint Research estimates that Broadcom will maintain approximately 60% of the custom AI accelerator market share by 2027, with Marvell capturing around 25%. The overall market is expected to reach $118 billion by 2033.
Implications for Google
Google’s chip strategy now encompasses four partners (Broadcom, MediaTek, Marvell, and TSMC), in conjunction with its in-house design team, leading to a diverse product line that covers training, inference, and general-purpose cloud computing. This complexity is intentional, as hyperscalers relying on a single chip supplier—be it Nvidia or another entity—face risks related to pricing, supply, and the strategic vulnerability of depending on external silicon.
The focus on inference in the ongoing discussions with Marvell highlights a significant shift in where financial resources are allocated. While Nvidia’s latest chips remain dominant for training workloads, inference represents a larger volume of demand, allowing custom silicon to leverage cost advantages. Google processes billions of AI-enhanced search queries, Gemini interactions, and Cloud AI API requests daily. Even marginal reductions in inference costs across such a vast scale can translate into substantial savings, which aligns with the objectives discussed during the earlier “Granite Redux” initiatives.
Though the talks with Marvell have yet to result in a formal deal, the trajectory is evident. Google is establishing a chip supply chain capable of supporting the most demanding AI inference workloads globally, ensuring it has multiple partners equipped to produce the necessary silicon. For Marvell, securing a contract for Google’s inference TPU would affirm its position as a leading custom AI chip designer, while for Google, it would provide an essential additional supplier in a market where reliance on a single provider is increasingly untenable.