Marvell put AI-scale connectivity at the center of its Industry Analyst Day, outlining a full-stack interconnect roadmap designed for the era of hyperscale GPU clusters and 1.6T-class networking. Speaking at the event, Xi Wang, senior vice president and general manager of the Connectivity Business Unit, said AI infrastructure is shifting from single-rack, general-purpose compute to warehouse-scale systems with hundreds of thousands of GPUs and XPUs acting as one. This shift forces operators to abandon “one-size-fits-all” connectivity and instead adopt domain-optimized solutions across servers, racks, data center fabrics, and regional links — spanning PCIe retimers, custom scale-up fabrics, active electrical cables (AECs), active optical cables (AOCs), and high-volume pluggable optics.
Within the rack, Wang described an industry-wide move away from passive copper cables toward active solutions as lane speeds rise and thermal envelopes tighten. Marvell now offers active copper and optical options tuned for different deployment envelopes: low-latency, short-reach active copper; higher-performance electrical and optical cables for mid-range distances; and pluggable optics starting at 800G and moving to 1.6T and beyond for leaf–spine and fabric connectivity. Beyond the rack, Marvell is developing “compound optics” that combine coherent and high-volume data center optics for distances from roughly 10 km to 100 km, allowing multiple AI data center buildings to operate as a single virtual compute domain.
To handle reliability at AI scale, Marvell is introducing Reliant, a real-time telemetry and analytics suite that aggregates link-health and optical performance data from cables and modules and integrates with customers’ network operating systems. The company also announced its new Golden Cable initiative for AECs — a program providing fully validated cable architectures, firmware, calibration data, and reference designs to cable vendors and hyperscalers. Instead of selling Golden Cables, Marvell will distribute these reference designs to partners to accelerate interoperability, enable customization of parameters such as wire gauge and bend radius, and speed time-to-market for 1.6T-capable short-reach AECs used in scale-up and scale-out AI racks.
• AI infrastructure is moving from single-rack servers to massive clusters with hundreds of thousands of GPUs/XPUs acting as a unified system.
• Connectivity priorities now differ sharply by domain: servers require low power and latency; racks and fabrics need higher bandwidth and reach; campuses and regional links prioritize distance and reliability.
• Marvell provides electrical and optical building blocks including PCIe technologies, scale-up fabrics like UEC/UAD, and pluggable optics from 800G to 1.6T and multi-Tbps modules.
• Optical connectivity is penetrating deeper into systems as copper alone can no longer meet power-density and bandwidth requirements.
• Inside the rack, operators are shifting from passive DACs to AECs and AOCs as speeds scale toward 1.6T and thermal constraints intensify.
• Marvell’s AEC portfolio targets short- to mid-reach links, with options optimized for latency, power, reach, and cable construction.
• For data center fabrics and regional interconnects, the company’s optics strategy spans a few meters to about 100 km, supporting multi-building AI campuses.
• The Reliant software suite provides real-time link-health visibility at both near and far ends and integrates into operators’ network OS environments.
• The Golden Cable initiative offers validated AEC designs — mechanical, electrical, firmware, and diagnostics — to speed partner development and manufacturing.
• Marvell positions Golden Cable as an open ecosystem program to expand supply diversity and accelerate hyperscaler adoption.
“As AI infrastructure scales at an unprecedented pace, the need for open, high-performance AEC interconnect solutions has never been more critical,” said Xi Wang, senior vice president and general manager of the Connectivity Business Unit at Marvell. “The Golden Cable initiative empowers ecosystem partners to quickly innovate, iterate and differentiate, delivering validated architectures, advanced firmware and design flexibility that will shape the next generation of hyperscaler AI deployments.”
🌐 Analysis: Marvell’s Golden Cable and Reliant announcements highlight how interconnect intelligence and open-ecosystem enablement are becoming core requirements in AI-scale data centers. As AEC demand rises with 1.6T switching and denser GPU racks, Marvell’s approach — combining silicon, optics, firmware, and real-time link analytics — aligns with hyperscalers’ need for multi-vendor supply chains and complements the company’s broader connectivity and optical silicon roadmap amid increasing competition across the AI infrastructure stack.
