bytevyte
bytevyte
Language
EN DE ES FR PT IT
ai-beats —

Marvell and Alphabet Co-Develop Custom AI Inference Silicon

custom AI inference silicon

Marvell Technology and Alphabet are in negotiations to co-develop two specialized semiconductors, according to industry reports. This collaboration focuses on custom AI inference silicon designed to optimize Google’s Tensor Processing Units (TPUs) and reduce reliance on third-party hardware providers.

The project involves the development of a memory processing unit (MPU) and a new generation of TPUs specifically tuned for inference tasks. Alphabet and Marvell expect to finalize these designs as early as next year, with a target for deployment by 2027.

Scaling Custom AI Inference Silicon for the Cloud

This move positions Marvell as a central player in the competitive AI ASIC market. Marvell shares rose 6.3% following reports of the expanded deal. The company’s custom silicon division has reached a $1.5 billion annual run-rate, securing 18 design wins across major cloud providers. These include infrastructure projects for Amazon’s Trainium and Microsoft’s Maia chips.

Alphabet’s decision to deepen its ties with Marvell follows the semiconductor firm's $3.25 billion acquisition of Celestial AI. This acquisition integrated "Photonic Fabric" technology into Marvell’s portfolio. The technology provides high-speed interconnect capabilities essential for modern data center workloads.

Financial performance reflects this growth. Marvell reported data center revenue of $6.1 billion for fiscal year 2026. While the company maintains a $2 billion networking silicon partnership with Nvidia, this custom AI inference silicon venture directly challenges Nvidia's current stronghold on the AI infrastructure market.

By developing custom AI inference silicon in-house with Marvell, Google aims to enhance the thermal efficiency and processing power of its cloud infrastructure. As of April 2026, the shift toward bespoke hardware continues to redefine the balance of power between traditional chipmakers and hyperscale cloud operators.

While we strive for accuracy, bytevyte can make mistakes. Users are advised to verify all information independently. We accept no liability for errors or omissions.

✔Human Verified

Share