Marvell Technology's Strategic Alliance with Lumentum Targets AI Infrastructure Evolution

Marvell and Lumentum showcase integrated optical switching and DSPs to cut AI data center latency and power, alongside new PCIe 6.0/CXL switches from Marvell's XConn tech.

A pivotal demonstration at OFC 2026 in Los Angeles is highlighting a collaborative effort between Marvell Technology and Lumentum Holdings. The focus is on defining the connectivity framework for next-generation artificial intelligence data centers, with a particular emphasis on advanced optical high-speed links—a domain becoming increasingly vital for scaling large AI models.

Streamlining Data Center Architecture

The core of the joint showcase involves the integration of Marvell's Aquila and Ara digital signal processors (DSPs) with Lumentum's R300 platform for optical circuit switching. This combination is engineered to establish direct optical pathways between endpoints within a server rack. The approach aims to eliminate conventional intermediary packet processing steps and the need for repeated optical-to-electrical conversions.

This architectural shift is designed to significantly cut both latency and power consumption, two critical constraints in large language model (LLM) training workloads. The setup is further augmented by COLORZ 800 ZR/ZR+ modules, which facilitate high-bandwidth connections across hyperscale infrastructure.

Adding a layer of intelligent management, the RELIANT platform provides real-time telemetry and automated optimization across the entire connectivity portfolio. This system is intended to assist cloud providers in managing the escalating operational complexity of large-scale AI clusters.

Advancing Hardware Integration with XConn

Concurrent with the Los Angeles demonstrations, Marvell is beginning to realize benefits from its integration of acquired XConn technology. This has yielded the new Structera-S series of switches.

The S 60260 model is positioned as the industry's first PCIe 6.0 switch offering 260 lanes, promising maximum design flexibility for AI data centers. Its counterpart, the S 30260, enables CXL-based memory pooling at the rack level. This capability allows multiple servers to share pooled memory resources, a feature deemed crucial for supporting the expanding context windows in generative AI applications.

Both switches are slated to become available for initial customer sampling in the third calendar quarter of 2026. This timeline aligns with the broader industry's transition from 800G to 1.6-terabit connections, a shift that Marvell anticipates will accelerate markedly through 2027. The extent to which major hyperscalers adopt these integrated solutions will be a key indicator of how effectively this technical prowess translates into tangible revenue growth for the company.