Marvell Secures $120 Billion Custom AI Chip Agreement With Google, Grants $12.2 Billion Stake Option

The speed, cost, and capability of the artificial intelligence tools you use daily are currently bottlenecked by a physical constraint: how fast data can move inside massive server farms. For years, the technology industry has relied on general-purpose processors to handle this traffic, but the sheer scale of modern AI models is forcing a rethink. To eliminate these internal traffic jams, the companies building frontier AI are increasingly designing their own hardware from the ground up.[1][2]

That shift accelerated sharply this week when Alphabet’s Google finalized a sweeping agreement with Marvell Technology to co-design custom silicon for its data centers. The partnership, disclosed in regulatory filings, covers a broad portfolio of specialized chips designed to attach directly to Google’s existing AI infrastructure. It marks a significant escalation in the race among cloud providers to control their own hardware destinies, ensuring they have the exact components needed to run next-generation models efficiently.[1][4]

To cement the long-term relationship, Marvell issued Google a stock warrant that allows the search giant to purchase up to 58.97 million shares at a fixed price of $206.58. If fully exercised, that position would be worth approximately $12.2 billion and give Google a roughly 7 percent stake in the semiconductor company. However, the financial structure of the agreement is entirely performance-based. Google does not receive the equity simply for signing the contract.[3][6]

Instead, the vast majority of the shares will only vest if Google actually purchases billions of dollars worth of custom chips from Marvell between now and 2033. According to the regulatory filings, only about 1.36 million shares will vest automatically over the first year. The remaining 57.6 million shares are divided into 240 equal tranches. One tranche of shares will unlock for every $500 million in custom product revenue that Marvell recognizes from Google.[3][7]

To unlock the entire $12.2 billion stake before the warrant expires in August 2033, Google would need to purchase roughly $120 billion worth of Marvell products over the next seven years. That ceiling represents a staggering potential shift in Marvell’s business scale. For context, Marvell generated a record $2.42 billion in total revenue during its most recent fiscal quarter. Even annualized, the company’s entire global business currently operates at less than $10 billion a year.[5][7]

At the engineering level, the collaboration is focused on solving the data movement problem through a concept known as near-memory compute. When a large language model generates text, it requires heavy, repeated access to massive parameter matrices stored in memory banks. Moving that data back and forth to the primary processor consumes enormous amounts of energy and introduces significant latency.[2]

By designing custom storage controllers, memory interface controllers, and specialized network interface controllers, Google and Marvell aim to move the actual computation closer to where the data lives. Cutting down the physical distance that information must travel is critical for reducing the energy consumption that currently dominates AI inference workloads. The less glamorous chips that manage data flow are becoming just as important as the processors doing the math.[2][7]

These new custom chips are not designed to replace Google’s primary AI engines. Google has spent years developing its own Tensor Processing Units (TPUs), which act as the heavy lifters for training and running its AI models. Instead, the Marvell-designed silicon will sit alongside and attach to the TPU ecosystem. The scope of the agreement includes AI inference accelerators that can offload specific workloads from the primary TPUs.[1][2]

This architecture allows Google to use specialized silicon for the parts of the computing stack where a general-purpose processor—or even a standard TPU—might be the wrong tool for the job. By optimizing the entire server rack for maximum efficiency, Google can lower the cost of serving AI to end users while managing the massive power demands of its expanding data centers.[1][4]

The partnership also signals a broader shift in the semiconductor industry as cloud providers reduce their reliance on off-the-shelf components. Google joins Amazon and Microsoft in deepening its bespoke chip efforts, placing indirect pressure on market leader Nvidia to maintain its dominance through its proprietary software and networking edge. Furthermore, the Marvell deal expands Google’s supplier base beyond Broadcom, which has historically been the search giant’s primary custom-chip development partner.[1][2]

While Broadcom recently signed its own long-term agreement to supply Google through 2031, the addition of Marvell indicates that the sheer volume of AI infrastructure required in the coming decade is large enough to support multiple custom silicon giants simultaneously. Financial analysts note that the current environment is not necessarily a zero-sum game; companies with custom silicon design capabilities are finding abundant opportunities as the world's largest tech firms race to build the physical foundations of the AI era.[4][5]