What Google’s Reported $120 Billion Marvell Chip Deal Actually Says

A closer look at the Google-Marvell agreement

Google and Marvell have expanded their partnership around custom silicon for generative AI infrastructure. The arrangement was announced through an 8-K filing that Marvell submitted to the U.S. Securities and Exchange Commission. Google did not issue a public statement about the agreement.

News reports described the arrangement as a $12.5 billion investment by Google in Marvell stock, paired with as much as $120 billion in chip-related business and services through 2033. The underlying documents, however, describe a more conditional structure.

The filing says the companies entered into a commercial agreement concerning Marvell’s development of custom semiconductor products for Google. It specifies that the expanded partnership covers several types of silicon connected to Google’s TPU ecosystem, including:

• AI inference accelerators
• Storage controllers
• Network interface controllers
• Memory interface controllers
• Near-memory compute products

Marvell’s business includes controllers, switch ASICs, photonics, and custom-chip development services. Its custom semiconductor operation helps customers design CPUs and XPUs, move those designs through foundries, and complete packaging.

How the stock warrants work

Under the share warrant agreement, Google receives warrants that can be converted into Marvell shares as revenue targets are met. The warrants cover 58.97 million shares, which would represent approximately 6.3 percent of Marvell after completion.

A total of 1.36 million shares vest during the first four quarters. The remaining warrants vest in 240 tranches, with each tranche tied to $500 million in revenue from products and services Marvell provides to Google.

Google is not required to exercise the warrants immediately. The agreement allows it to choose when to convert vested warrants, potentially waiting until August 18, 2033. If Google does not exercise its rights by the following day, the warrants can expire. The documents specify a floor price of $206.58 per share.

Marvell’s market value rose by approximately $18.6 billion on the day the agreement was announced. That increase exceeded the reported $12.5 billion value of the warrants, although no components had yet shipped under the new arrangement.

Reports about the Merope project

Trade and business publications connected the agreement with an alleged Google chip code-named “Merope.” Neither Google nor Marvell has confirmed that Merope is part of the deal.

Merope has been described in reports as a language processing unit, or LPU. Such a processor would focus on the decode phase of generative AI inference, when a system generates a response to a query. This differs from prefill, which processes the prompt and its context before response generation begins.

Earlier reports said Google was discussing two chip projects with Marvell: the rumored LPU and a memory-focused processor. The latter could involve a CXL-based memory-area network that allows TPU systems to share DRAM as a cache tier, potentially reducing the amount of local DRAM required in TPU host systems.

The agreement itself does not name Merope. It does, however, mention a product code-named “Kestrel,” which Google reportedly wants qualified by November 10, 2027.

The separate Kestrel data center project

Google also uses the name Project Kestrel for a planned data-center development in its Kansas City, Missouri, region. The project covers approximately 430 acres and is expected to include as many as five hyperscale data centers. Reported estimates put the development at around $100 billion and approximately 2.5 million square feet of combined capacity. Earlier reports described six data centers and 1.8 million square feet.

The Kestrel product mentioned in the warrant agreement could therefore refer to a group of components rather than a single chip. Possibilities include storage, memory, and network controllers, memory expanders, a custom inference accelerator, or a CXL-based memory-area network.

Google’s TPUs are designed for AI training and prefill workloads. A separate accelerator could address low-latency decode workloads, an area where other AI accelerators, including GPUs and Trainiums, also face different performance considerations.

A conditional arrangement tied to AI infrastructure

The agreement creates a connection between Google’s purchases from Marvell and Marvell’s stock performance. Google can receive additional shares as Marvell reaches revenue milestones, while the commercial relationship itself may contribute to Marvell’s future revenue and market valuation.

Google could use custom TPUs, inference accelerators, memory systems, controllers, and networking components across its own infrastructure. Some of those systems may also support TPU deployments made available to Google Cloud customers, including large AI customers such as OpenAI and Anthropic.

If Kestrel refers to a custom accelerator for Google’s planned Kestrel data centers, qualification in late 2027 would allow deployment as those facilities are completed in 2028. Marvell’s role would include guiding the designs through Taiwan Semiconductor Manufacturing Co. foundries and packaging operations.

The public documents establish the commercial agreement, the revenue-linked warrants, and the Kestrel qualification date. They do not confirm whether Merope is included, what Kestrel specifically contains, or whether either name identifies a single product or a broader program.