We are the Frontier
World

Marvell Skips Chip Nodes to Boost AI Data Efficiency

The company is prioritizing optical interconnects and 5nm production to optimize AI-heavy scale-up network power demands.

Updated on Oct. 26, 2026 in Data Centers

Isometric editorial illustration of a circular silicon semiconductor wafer, representing advanced chip infrastructure for data center power efficiency.
Marvell Technology Inc. is transitioning to 5nm chip production, prioritizing optical interconnects to reduce power consumption in AI-driven data center networks. AI Illustration. Upload story photo >

Live Poll

Is the rapid growth of AI infrastructure and data center capacity good for the country?

Marvell Technology Inc. has shifted its data center infrastructure strategy by skipping semiconductor process nodes, moving directly from 7nm to 5nm. This transition aims to improve energy efficiency for AI-driven data transmission networks.

Why it matters

As 85% of processor-to-processor data traffic occurs within internal scale-up networks, optimizing power efficiency through advanced process nodes is critical for AI infrastructure sustainability. This strategy targets the rising demand for high-bandwidth connectivity in massive data centers.

The Photonic Fabric Memory Module integrates DDR5 memory, high-bandwidth memory caching, and optical connectivity. Marvell uses SerDes (Serializer/Deserializer) technology to handle high-speed data transmission while minimizing power consumption.

The players

Marvell Technology Inc.

A semiconductor provider specializing in data infrastructure, optical interconnects, and custom silicon for AI-driven cloud data centers.

Nvidia Corp.

A leading AI hardware company that invested $2 billion in Marvell in 2026 to secure next-generation data infrastructure.

Google LLC.

A major cloud service provider that entered an agreement to utilize custom AI chips designed by Marvell.

The details

Marvell utilizes SerDes technology, which converts parallel data streams into high-speed serial streams to reduce physical cabling and signal loss. The Photonic Fabric Memory Module acts as a bridge, combining traditional DDR5 memory with HBM (High-Bandwidth Memory) caching to accelerate data movement. By skipping process nodes, the firm aims to optimize the physical chip architecture for these specialized interconnects.

Timeline

  1. 2018: Marvell initiated its strategy to skip semiconductor process nodes.

  2. August 2026: The company announced its Photonic Fabric Memory Module.

  3. September 2026: Marvell unveiled 2-nanometer optical interconnect technologies.

The Tech Race

Marvell is positioning itself at the center of the AI compute stack by bypassing standard node progression to rapidly iterate on optical interconnects. This development directly competes with other providers attempting to solve the bandwidth bottlenecks caused by AI scale-up requirements.

Cloud infrastructure providers adopting these modules will see immediate benefits in energy efficiency for high-density AI clusters. Operators can expect reduced power overhead in scale-up networks once these components are integrated into standard data center racks.

The takeaway

Marvell is betting that moving to smaller process nodes faster than industry standards will grant it a lead in the AI infrastructure market. Watch for the performance metrics of its custom AI chips developed under the Google LLC agreement to confirm if this strategy yields a competitive speed advantage.

Further reading

Explore deeper trends in infrastructure design within our Data Centers section.

Source note: This article includes information reported by SiliconANGLE.

Live Poll

Is the rapid growth of AI infrastructure and data center capacity good for the country?