Industry analysis scrutinizes reported discussions between Google and Marvell regarding a potential $120 billion custom silicon procurement agreement.
Another day, another major transaction pairing a technology titan with a chip designer or supplier—sometimes both simultaneously—to fund tens of billions of dollars in generative AI infrastructure. This latest agreement, announced last week and reviewed during my virtual attendance at Hot Chips 2026, involves hyperscaler and cloud provider Google and chip architect and manufacturer Marvell. Marvell aims to leverage its multifaceted portfolio, which spans controllers, switch ASICs, and photonics, alongside its growing custom chip design services that guide customers through CPU and XPU architecture, foundry production, and packaging.
The arrangement was disclosed in an 8-K filing by Marvell, which provided limited detail, while Google issued no public statement. Examining the share warrant agreement signed this week clarifies the structure. Contrary to headlines claiming Google is investing $12.5 billion in Marvell stock in exchange for $120 billion in chip sales and engineering work through 2033, the actual terms are more nuanced. The SEC filing states the companies “entered into a commercial agreement relating to the Company’s development of custom semiconductor products to Google (the “Custom Products”.)” It further notes: “The expanded partnership spans a comprehensive range of custom silicon programs that attach to the TPU ecosystem, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute.”
Trade and business media quickly latched onto reports that the partnership includes collaboration on a specialized AI chip codenamed “Merope,” details of which remain closely held by Google and possibly Marvell. Industry observers characterize Merope as a language processing unit (LPU), optimized for the decode phase of generative AI inference—the stage where the model generates a response to a query, as opposed to the prefill phase, which processes prompts and context. In April, The Information reported that Google was negotiating with Marvell to develop two chips: the aforementioned LPU (referred to as Merope) and a memory processor, potentially leveraging CXL-based memory area networking to allow TPU systems to share DRAM as a caching tier, thereby reducing host memory requirements. Given that DRAM prices have surged 7.1 times over the past year, such architectural efficiency has become essential. Neither company has commented on these rumors.
The $18.6 billion surge in Marvell’s market capitalization on the day of the announcement already exceeds the $12.5 billion warrant value, meaning the deal has effectively paid for itself before any components ship. The warrants cover 58.97 million Marvell shares, representing a 6.3 percent stake upon full conversion. Under the vesting schedule, 1.36 million shares vest across the first four quarters, with the remainder distributed across 240 tranches, each triggered by $500 million in cumulative revenues from Marvell’s sales to Google.
This arrangement functions as a strategic feedback loop within the current generative AI investment cycle. Google commits to purchasing $500 million worth of Marvell stock quarterly over six years to secure the partnership. The resulting commercial activity subsequently drives Marvell’s valuation higher, creating a cyclical benefit that helps offset the costs of acquiring TPUs, LPUs, custom silicon, and off-the-shelf components for Google’s servers and networks. Much of this hardware powers Google Cloud and will likely supply TPU systems sold to external clients like OpenAI and Anthropic. Should Marvell’s market capitalization grow as projected, Google’s increasing equity stake would act as a partial rebate on its chip design and manufacturing fees. In effect, Google funds the hardware buildout and receives equity appreciation that strengthens its balance sheet, improving its capacity to raise private debt for further AI infrastructure investments—a cycle that ultimately benefits Marvell.
A critical detail lies in the agreement’s addendum: Google is not obligated to purchase the shares. It may exercise the vested warrants at its own discretion, potentially delaying execution until August 18, 2033. If Google chooses not to exercise them—for instance, if Marvell’s stock falls below the established floor of $206.58 per share—the warrants expire worthless. While unlikely given Google’s interest in board influence and cost recovery for its AI expansion, the optionality remains structurally significant.
Another notable detail involves product nomenclature. While “Merope” does not appear in the filing, a product codenamed “Kestrel” is explicitly referenced. This likely denotes the memory server or associated controller suite, with a qualification deadline set for no later than November 10, 2027. Compounding the confusion, Google also operates “Project Kestrel,” a separate initiative involving a series of hyperscale datacenters in the Kansas City, Missouri region. Spanning approximately 430 acres, the site will accommodate up to five facilities valued at roughly $100 billion, offering around 2.5 million square feet of aggregate capacity (though earlier reports suggested six buildings totaling 1.8 million square feet).
The “Kestrel” product launch outlined in the addendum may therefore represent a broader portfolio of storage, memory, and network controllers, memory expanders, and a custom AI inference chip designed to operate alongside Google’s TPUs. While TPUs excel at training and prefill workloads, they face latency constraints during decode phases—a limitation shared by GPUs and Trainium chips. Alternatively, the Kestrel designation could refer to a CXL-based memory area network. My assessment is that Kestrel represents the custom silicon deployed within the Kansas City datacenters, effectively transforming them into a massive clustered AI inference accelerator. Qualifying these accelerators for deployment in late 2027 aligns with a 2028 rollout timeline, as Marvell guides the designs through Taiwan Semiconductor Manufacturing Co. foundries and advanced packaging. That assessment, however, remains a human inference—a hunch.