Marvell and Lumentum have successfully demonstrated a next-generation optical network for AI, signaling a major step toward commercialization.
Today's AI models require enormous amounts of data to be moved between tens of thousands of GPUs, creating a problem known as the 'networking wall'. This refers to the physical limits of current networks, which are struggling with data traffic jams (latency), high power consumption, and bandwidth constraints. As AI clusters grow, these problems become critical bottlenecks, increasing costs and limiting performance. Hyperscalers are now urgently seeking a new architecture to overcome this challenge.
This is where Optical Circuit Switching (OCS) comes in. Think of it as creating a dedicated, private highway for data made of light. Instead of converting light signals to electrical signals to be routed through a traditional switch and then back to light (OEO conversion), OCS directly connects endpoints with an uninterrupted path of light. This simple yet powerful change brings significant benefits. First, it dramatically reduces latency—Lumentum claims a 98% reduction in switch latency. Second, by eliminating power-hungry electrical conversions, it can cut network power consumption by over 65% in a large-scale 100,000 GPU cluster.
This demonstration is not an isolated event but the culmination of several key industry trends. The timing is significant, coming just a week after major tech companies like NVIDIA, Microsoft, and Meta launched the 'Optical Compute Interconnect (OCI) MSA', an alliance to create open standards for optical connections. This move provides the ecosystem foundation. Furthermore, there is clear market demand; Lumentum has already secured over $400 million in OCS backorders, proving that major customers are ready to invest.
In this partnership, Marvell provides the advanced 'brains'—the Digital Signal Processors (DSPs) that manage the light signals—while Lumentum builds the physical 'highways' with its high-port OCS. Their combined solution provides a credible, tangible path from today's strained networks to tomorrow's efficient optical fabrics. While challenges like managing the complex scheduling of these light paths remain, the industry's direction is now clearer than ever. The race to build the optical infrastructure for AI is officially on.
- Optical Circuit Switching (OCS): A networking technology that creates a dedicated physical path of light between two points, bypassing traditional electronic switches to reduce latency and power consumption.
- OEO (Optical-to-Electrical-to-Optical) Conversion: The process in conventional networks where an optical signal is converted to an electrical one for processing (e.g., switching) and then converted back to an optical signal for transmission.
- East-West Traffic: Data traffic that moves laterally between servers within a data center, which is especially heavy in distributed AI training workloads.