Data Gravity Analysis: AI Interconnects Shift from Copper to Optical, with NVIDIA Leading with CPO-Equipped Spectrum-X Switches

Data Gravity USA

Overview

Data Gravity’s analysis indicates that frontier AI performance constraints are shifting from accelerator supply to interconnects, highlighting challenges in chip-to-chip, GPU-to-GPU, and rack-to-rack connectivity. NVIDIA is prioritizing optical technology through significant investments in optical suppliers and the deployment of its Spectrum-X Photonics SN6800 switch, featuring Co-Packaged Optics (CPO). This strategy aims to replace copper interconnects and address power consumption and resilience issues. Diverse optical alternatives, including silicon photonics, VCSEL-based Near-Packaged Optics, and THz radio-over-wire, are becoming practical in 2026.

In Depth

Key Findings

According to a comprehensive analysis by Data Gravity, the primary bottleneck impeding the performance of cutting-edge AI systems has shifted from traditional accelerator supply to interconnect capabilities. This transition has spawned new challenges in chip-to-chip, GPU-to-GPU, and rack-to-rack connectivity within data centers. NVIDIA is positioning optical technology at the heart of its solutions, demonstrated by substantial investments in optical suppliers and the deployment of its Spectrum-X Photonics SN6800 switch, which incorporates Co-Packaged Optics (CPO). This strategy aims to replace copper interconnects, significantly reduce power consumption, and enhance system resilience.

Technical / Clinical Details

AI workloads demand terabit-scale data transfer and microsecond-level latency, pushing conventional copper interconnects to their limits in terms of both performance and power efficiency. CPO technology addresses this by directly integrating optical transceivers with switch chips or GPU packages, drastically shortening the electrical signal paths. This approach leads to significant reductions in signal loss and power consumption. NVIDIA’s Spectrum-X Photonics SN6800 switch, by implementing CPO, substantially boosts data center bandwidth density and optimizes the performance of AI clusters. Furthermore, a diverse array of optical technologies, including silicon photonics, VCSEL-based Near-Packaged Optics (NPO), and Terahertz (THz) radio-over-wire, are becoming practical and commercially viable in 2026, expanding the options for advanced AI interconnects.

Background & Context

As AI models scale from gigabytes to terabytes and even petabytes, data center power consumption has become a major sustainability concern. NVIDIA CEO Jensen Huang’s ‘Copper-and-Optics Roadmap’ outlines an industry-wide direction for addressing these challenges. NVIDIA’s investment in the optical supply chain, engaging key players like Lumentum, Coherent, and Corning, is driving a fundamental transformation of AI infrastructure. This move is particularly significant given that competitors such as Broadcom and Intel are also pursuing their distinct silicon photonics strategies, indicating a broad industry consensus on the critical role of optics.

Strategic Significance & Outlook

The shift from copper to optical technologies for interconnects in the AI era is an irreversible trend, and Data Gravity’s analysis confirms the acceleration of this transition. NVIDIA’s deployment of CPO-equipped switches is paving the way for optical technology to become the standard in AI data centers. As silicon photonics and CPO/NPO solutions mature, data center energy efficiency and scalability are expected to improve dramatically. This will, in turn, enable the development and deployment of even larger and more complex AI models, further expanding the impact of AI technology across industries and society. Investors and technologists should closely monitor the innovations and market dynamics in this pivotal sector.