Key facts
- Marvell Technology has granted Alphabet's Google a warrant to acquire a stake valued at approximately $12.18 billion.
- The deal involves Marvell developing custom AI inference accelerators, storage, networking, and memory interface controllers for Google.
- This collaboration aims to address the growing demand for custom AI chips and optimize LLM inference economics.
- Marvell's shares surged over 11% in premarket trading, while rival Broadcom's shares fell more than 3%.
- Google is also in parallel discussions with Marvell for a dedicated chip for Large Language Model (LLM) inference.
Marvell Technology announced a significant deal with Alphabet's Google, granting the tech giant a warrant to acquire a stake potentially valued at $12.2 billion. This strategic partnership focuses on the development of custom artificial intelligence chips, specifically AI inference accelerators, storage, networking, and memory interface controllers for Google's advanced workloads.
The agreement comes as demand for specialized AI hardware intensifies, with companies like Google seeking alternatives to Nvidia's GPUs for tasks such as Large Language Model (LLM) inference. Google's pursuit of custom silicon, even alongside its existing Tensor Processing Units (TPUs), highlights the ongoing challenges and costs associated with commercial-scale LLM inference.
Marvell's shares experienced a notable surge of over 11% in premarket trading following the announcement. In contrast, rival chipmaker Broadcom saw its shares dip by more than 3%. Broadcom itself has a separate, long-term agreement with Google, underscoring Google's strategy of pursuing multiple development paths for its AI chip needs.
Marvell's custom ASIC business is already a significant contributor, generating approximately $1.5 billion in annualized revenue from existing cloud provider engagements. The potential revenue from this new deal, estimated at $120 billion through fiscal 2033, hinges on Google achieving specific performance targets related to the custom chips.