Key facts
- Marvell Technology issued Alphabet's Google a warrant for a stake worth approximately $12.18 billion.
- The deal focuses on developing custom chips for Google's AI inference workloads.
- Marvell will create AI inference accelerators, storage, networking, and memory interface controllers for Google.
- Google previously reduced its AI compute costs by migrating workloads to its Tensor Processing Units (TPUs).
- Marvell's custom ASIC business currently generates about $1.5 billion in annualized revenue from cloud providers.
Marvell Technology announced on Wednesday that it has granted Alphabet's Google a warrant to acquire a stake valued at approximately $12.18 billion. This agreement is part of a broader deal aimed at developing custom chips tailored for Google's artificial intelligence workloads.
The collaboration will see Marvell develop AI inference accelerators, storage, networking, and memory interface controllers, along with near-memory computing technologies for Google. This move comes amid a surge in demand for custom AI chips, such as Google's Tensor Processing Units (TPUs), as businesses seek alternatives to Nvidia's graphics processors for AI tasks.
Marvell's shares rose over 11% in premarket trading following the announcement, while competitor Broadcom saw a decline of more than 3%. Broadcom itself had previously signed a long-term agreement with Google in April to supply future generations of custom AI chips and components through 2031.
Reports suggest that Google is actively negotiating with Marvell to engineer a dedicated chip specifically for Large Language Model (LLM) inference. This indicates that even Google's existing TPU infrastructure may not fully meet the demands of commercial-scale LLM inference, highlighting ongoing challenges in LLM inference economics. The talks also involve Marvell acting as a design services provider for Google's TPU customization project, running parallel to the Broadcom agreement to ensure redundancy and architectural differentiation.
Marvell's strategic position is further highlighted by Nvidia's recent $2 billion investment in the company, which integrates Marvell's custom accelerators and networking chips into Nvidia's NVLink Fusion platform. Marvell's custom ASIC business already generates around $1.5 billion in annual revenue from existing cloud-provider design wins.