Marvell Shows 2nm Optical Chips for AI Data Centers

Marvell Technology is preparing to demonstrate a series of first-of-their-kind optical interconnects built on a 2-nanometer process node. These components are designed specifically for the demanding connectivity requirements of artificial intelligence data centers, where speed and energy efficiency are critical.

The demonstrations, scheduled for the European Conference on Optical Communication in September 2026, mark a significant step in the company’s roadmap. By moving to the 2nm node, Marvell aims to deliver higher bandwidth density and lower power consumption per bit, addressing the escalating energy demands of large-scale AI deployments.

New optical speeds for AI networks

The centerpiece of the showcase is the industry’s first 2nm 400G/lane optical PAM4 technology. This innovation supports the transition to 3.2T data center connectivity, offering a faster and more energy-efficient way to link AI accelerators and servers. It builds on previous advancements, such as the 3nm 1.6T platform released in 2024, pushing the boundaries of what is possible with optical transmission.

Alongside the speed improvements, Marvell will display the first live demonstration of a 2nm 800G ZR/ZR+ pluggable module with integrated MACsec security. This feature is crucial for hyperscale networks that require robust data protection without sacrificing performance. The module leverages the company’s Libra coherent DSP to ensure secure and efficient optical transmission across global data center networks.

Co-packaged optics for high-density fabrics

A key part of the offering is the Marvell co-packaged optics (CPO) platform, demonstrated at 102.4 Tbps. This technology integrates optical components directly with switch silicon and AI accelerators, bypassing the physical limitations of traditional copper cabling. As AI pods expand to multiple racks, this direct integration is necessary to maintain high bandwidth density while managing power consumption.

Marvell will also showcase 1.6T ZR and coherent-lite O-band technologies. These solutions are optimized for connectivity within and between data center facilities, paving the way for future 3.2T upgrades. The broad portfolio of 38 demos highlights a strategy focused on providing comprehensive solutions for the entire AI infrastructure stack.

Balancing performance with energy costs

The primary trade-off addressed by these new 2nm technologies is the growing power cost of data center operations. While increasing bandwidth is essential for AI workloads, the energy required to transmit that data has become a major operational expense. By improving power efficiency per bit, these optical interconnects help operators manage these rising costs without compromising network speed.

However, adopting these advanced nodes requires significant investment in new infrastructure and supply chain adjustments. Operators must weigh the benefits of higher density and security against the complexity of integrating next-generation optical components into existing data center architectures. The move to 2nm represents a leap in capability, but it also demands a corresponding shift in how data centers are designed and maintained.