On August 19, 2026, Marvell Technology announced that Google had the right to purchase nearly 59 million Marvell shares at $206.58 each. The potential $12.2 billion stake would make Google the company’s fifth-largest investor. Marvell shares rose 9.9%, while Broadcom, Google’s longtime primary custom chip partner, fell more than 5%.
The agreement ties Google’s chip purchases to an opportunity to own part of its supplier. It also extends beyond AI processors into storage and networking chips. For Google, that means a closer financial relationship with a company helping build several parts of its AI infrastructure. For competing suppliers and cloud customers, the consequences could take years to become clear.
How the warrant agreement works
The $12.2 billion figure isn’t an upfront cash payment. The arrangement uses warrants, which give Google the right to buy shares at a specified price. Approximately 1.4 million shares vest in the first year. The rest become available in tranches triggered by every $500 million in chip purchases Google makes from Marvell.
Buying more chips therefore unlocks more rights to purchase equity. It doesn’t automatically give Google ownership of all those shares. If Google reaches the full target, Marvell estimates that the partnership could generate approximately $120 billion in revenue through fiscal 2033. That is potential revenue tied to the arrangement’s full scale, rather than a completed $120 billion purchase.
The expanded agreement covers technologies used with Google’s tensor processing units, or TPUs, its custom AI processors. These include AI inference accelerators, which run trained models, storage controllers, and network interface controllers. The scope gives Marvell a role across the surrounding infrastructure as well as AI compute.
This is a form of vertical integration through contracts and financial incentives. Google isn’t acquiring Marvell outright, but it is tying more of its chip supply chain to its own purchasing plans.
Custom silicon now reaches beyond the processor
Large cloud providers have been developing their own chips for years. Google shipped its first TPU internally in 2016. AWS developed Nitro, Trainium, and Inferentia, while Microsoft built Azure Maia. Those efforts established custom silicon as part of cloud infrastructure, even while providers continued to rely on Nvidia, Broadcom, and other suppliers.
The Marvell agreement broadens that approach. Designing a compute die, the piece of silicon that performs the processing, leaves storage and networking decisions partly in other suppliers’ hands. An agreement spanning those components gives Google more room to coordinate how data moves through an AI system, not just how quickly a processor can execute a model.
The equity structure also gives Marvell a financial reason to align its business with Google’s roadmap over the next seven years. That is a deeper relationship than a series of separate component orders.
Other reported arrangements point in a similar direction. AMD reportedly struck a strategic arrangement with OpenAI. Nvidia assembled a $500 billion infrastructure financing consortium with Apollo, Blackstone, BlackRock, and others to support customers’ data center construction. These arrangements suggest that long investment cycles and large purchase volumes are encouraging relationships lasting seven to ten years, with financing becoming part of the supplier relationship.
Potential effects on cloud costs and supplier bargaining
For hosting companies, enterprise data centers, and engineering organizations that depend heavily on cloud services, the effects may take another eighteen months to become visible. The most direct connection is between chip purchasing economics and the cost of the cloud services built on those chips.
If Google’s purchases reach the scale Marvell projects, negotiated chip prices and the equity arrangement could shape Google Cloud’s AI compute costs through 2033. Those costs could eventually affect customer pricing, whether through lower prices or greater price stability. Providers with favorable custom-silicon agreements may gain a cost advantage over competitors buying merchant silicon, meaning chips sold broadly to customers rather than designed around one cloud provider’s requirements.
That advantage remains an expectation, not something the headline deal value establishes. The arrangement doesn’t by itself show how much of any savings Google would pass on to customers.
Broadcom’s position also deserves a closer reading than its share-price decline alone allows. Analysts have described the agreement as an expansion of the overall market rather than a displacement of Broadcom. That interpretation can coexist with greater negotiating pressure. Google’s willingness to expand Marvell’s role and link purchases to a potential equity stake gives it another supplier to bring into future procurement discussions.
Broadcom needn’t lose revenue immediately for that alternative to matter. The strategic question is how much influence each supplier retains over Google’s next generation of systems.
The agreement also reflects an effort to reduce dependence on a single vendor’s production schedule. GPU availability has constrained AI infrastructure for years. Coordinating a broader custom-chip supply chain offers another way to address that constraint, although the agreement alone doesn’t establish that supply bottlenecks have been resolved.
Google is bringing chip and cloud decisions closer together
Bloomberg reported that Google recently restructured its organization to elevate custom chips and AI infrastructure within Google Cloud operations. That places chip architecture closer to the leadership responsible for the infrastructure running on those chips.
The potential benefit is more coordinated engineering. A TPU’s memory subsystem affects how efficiently it can receive and process data. Storage systems affect whether that data arrives quickly enough to keep the processor busy. Bringing those decisions into a closer organizational relationship could help engineers optimize the whole system rather than separate components.
The restructuring signals that Google sees chip design as part of operating its cloud, rather than as an isolated hardware program. Whether it produces better systems will depend on how those teams work together, but the organizational direction fits the broader supplier agreement.
A longer planning horizon for both companies
The warrant schedule gives Marvell a reason to plan around sustained Google demand. Each additional $500 million in purchases unlocks another tranche, linking the relationship’s expansion to actual spending.
A relationship extending through fiscal 2033 could help Marvell plan research and development for workloads Google expects to run in 2030 and 2031. Better visibility into a major customer’s requirements can reduce some of the uncertainty in semiconductor investment. The warrant terms alone, however, don’t establish that Marvell has complete visibility into Google’s future roadmap.
Google, in turn, gains a supplier with a substantial incentive to accommodate its engineering priorities. The prospect of $120 billion in revenue gives Marvell a strong reason to keep investing in the relationship. It also makes Google’s purchasing decisions consequential for Marvell’s business.
This differs from the end-to-end stacks built by companies such as IBM in the 1990s. Google is pursuing coordination through a separate supplier, using purchasing commitments and equity rights rather than bringing every part of production under one owner. The argument for that approach is that AI infrastructure has reached a scale where closer control over component design and supply may justify the added complexity.
What to watch next
A reasonable forecast is that Amazon will deepen the ecosystem around Trainium and Inferentia, potentially through similar equity-linked partnerships, and that Microsoft will expand its Azure Maia commitments. More such arrangements could reduce the role of off-the-shelf merchant silicon in AI inference as large cloud providers shift spending toward co-designed systems. The speed and extent of that shift remain uncertain.
For infrastructure planning, the useful comparison is between the costs and capabilities these supplier relationships produce. Providers that coordinate custom silicon successfully could develop cost structures that are difficult to match. Google’s Marvell agreement shows how it intends to pursue that advantage. Cloud pricing, available capacity, and the performance of the resulting services will show how much of it reaches customers.
