Google Deepens TPU Supply Chain With Performance-Linked Marvell Deal

Google has deepened its custom AI-chip strategy with Marvell Technology in a deal that ties semiconductor purchases to a potentially large equity stake. In a regulatory filing, Marvell disclosed that it entered a commercial agreement with Google on July 29 covering a broad set of custom semiconductor products connected to Google’s TPU ecosystem, then issued Google a warrant on August 18 to buy as many as 58,970,907 Marvell shares at $206.58 each. At the exercise price, the full warrant would be worth about $12.2 billion, but most of the shares vest only if Google buys large volumes of Marvell custom products over the coming years.

The deal is more consequential than a conventional supplier contract because it links procurement, product development and ownership incentives. Marvell says the expanded relationship covers AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute. Those components sit around the core accelerator and determine how efficiently data moves through an AI system, how memory is fed, how storage is accessed and how racks connect. For hyperscale AI infrastructure, those surrounding chips can become as important to throughput and cost as the accelerator itself.

The warrant structure makes the commercial ambition unusually explicit. About 1.36 million shares vest in equal quarterly installments during the first year. The remaining warrant shares are divided into 240 tranches, with one tranche vesting for every $500 million of qualifying custom-product revenue generated from Google and its affiliates from Marvell’s fiscal third quarter of 2027 through the end of fiscal 2033. Multiplying those milestones produces a theoretical $120 billion revenue ceiling, but that number is not a forecast and should not be read as committed spending.

That distinction matters because several early reports framed the agreement around the headline $120 billion opportunity. The underlying filing supports the arithmetic, but not a promise that Google will spend the full amount. Full vesting would require sustained purchases on a scale far above Marvell’s current annual revenue base. The safer interpretation is that Google has created a long-term incentive mechanism that rewards Marvell if it becomes a major supplier across multiple layers of the TPU infrastructure stack.

Independent reporting from Reuters and the Financial Times points to supplier diversification as a central part of the story. Broadcom has been a major partner in Google’s custom AI silicon program, and the Marvell agreement gives Google another large supplier with capabilities spanning custom compute, networking, storage and memory interfaces. Market reaction reflected that interpretation: Marvell shares rose sharply when the deal became public, while Broadcom fell as investors reconsidered how future Google custom-silicon spending might be distributed.

For AI architects, the most important point is that the competitive unit is no longer just the accelerator. Modern AI systems depend on tightly coupled stacks of compute, high-speed networking, memory hierarchy, storage, optical interconnect and software. A company such as Google can reduce dependence on a single vendor by designing more of that stack internally and using multiple semiconductor partners for different components. That can improve negotiating leverage and supply resilience, but it also raises integration complexity and verification requirements across generations of hardware.

The agreement also illustrates how hyperscalers are reshaping semiconductor economics. Instead of simply buying catalog products, large cloud and AI operators increasingly co-design silicon with suppliers, commit to multi-year development programs and use equity or warrants to align incentives. This gives the buyer more control over roadmap, performance and supply, while giving the chip designer a clearer path to large-volume revenue. The tradeoff is tighter customer concentration: a supplier that wins a hyperscaler design can grow quickly, but becomes more exposed to that customer’s architecture decisions and purchasing cadence.

From a business-governance perspective, the warrant creates a form of performance-linked strategic dependence. Google does not receive most of the potential equity simply for signing the agreement. The majority vests only as purchase milestones are reached. That means the ownership relationship grows alongside commercial reliance. Boards and procurement leaders evaluating similar structures should distinguish between nominal deal size, committed spend, contingent incentives and maximum theoretical value because those categories have very different risk implications.

The deal also matters for Nvidia even though it is not a direct replacement story. Google’s TPU strategy is one of the clearest examples of a hyperscaler building an alternative accelerator ecosystem rather than relying exclusively on merchant GPUs. Marvell’s expanded role strengthens the surrounding infrastructure needed to scale that ecosystem. Nvidia remains deeply embedded across AI infrastructure, including through its own partnership and investment in Marvell, which shows how interconnected the supplier landscape has become rather than how neatly it divides into competing camps.

There are still important uncertainties. Marvell’s filing does not specify which current or future TPU generations will use particular Marvell components, what share of Google’s custom silicon workload will move from existing suppliers, or how quickly purchase-linked vesting could accumulate. The $120 billion ceiling should therefore be treated as a contract mechanic, not revenue guidance. It is also possible that Google uses the agreement to diversify sourcing without materially displacing Broadcom in core TPU development.

For practitioners, the practical lesson is to watch architecture boundaries rather than vendor headlines. The most valuable layer in an AI system can shift as bottlenecks move from raw compute to memory bandwidth, network fabric, storage throughput or power efficiency. Google’s agreement with Marvell suggests that custom silicon strategy is expanding outward from accelerator design into the full data path around the accelerator. That creates more opportunities for specialization, but also more places where interoperability, validation and lifecycle management can fail.

The development is therefore significant not because Google has simply chosen a new chip vendor, and not because Marvell is guaranteed $120 billion of revenue. The material change is a multi-year, performance-linked expansion of Google’s custom AI infrastructure supply chain across several semiconductor layers attached to the TPU ecosystem. If Google follows through with large-scale purchases, the deal could materially rebalance the custom-AI-chip market. The next evidence to watch will be Marvell’s revenue disclosures, future TPU platform announcements and any signs of how Google divides work among Marvell, Broadcom and other silicon partners.