The Copper Wall: Why Nvidia and Industry Giants are Shifting to Optical Scale-Up
For years, the backbone of AI cluster scaling has relied on copper. It is cost-effective, reliable, and, until recently, sufficient. Still, as large language models (LLMs) push toward “super intelligence,” the physical limitations of copper are becoming a bottleneck. Nvidia, once a staunch defender of copper for short-reach connections, is now embracing optical scale-up to ensure AI clusters can continue to grow in size and speed.
The industry is hitting a “copper wall” where the physical reach and power requirements of copper cables can no longer keep pace with the bandwidth demands of next-generation GPUs. To solve this, a massive industry pivot is underway, shifting the connectivity paradigm from module-centric copper to silicon-centric optics.
The NVL72 Experiment: The Peak of Copper Connectivity
To understand why the shift to optics is necessary, one must look at the Nvidia DGX GB200 NVL72. This system represents the absolute limit of copper-based scale-up. The NVL72 integrates 72 GPUs into a single rack, utilizing a staggering 5,184 direct-drive copper cables—totaling roughly two miles of cabling.
Nvidia initially pushed this copper-heavy design for a specific reason: power efficiency. According to CEO Jensen Huang, replacing those copper cables with optical transceivers and retimers would have consumed an additional 20 kilowatts of power just to drive the NVLink spine. By using copper, Nvidia saved that power for actual computation.
However, as clusters expand beyond a single rack and bandwidth requirements climb, the power and distance penalties of copper become unsustainable. This is where optical scale-up enters the frame.
The Rise of the Optical Compute Interconnect (OCI)
Recognizing that no single company can define the future of AI interconnects alone, a powerhouse consortium was formed on March 12, 2026. The Optical Compute Interconnect (OCI) Multi-Source Agreement (MSA) brings together Nvidia, AMD, Broadcom, Meta, Microsoft and OpenAI.
The OCI MSA’s goal is to create an open, protocol-agnostic specification for optical scale-up interconnects. This allows hyperscalers to use optical cables instead of copper to connect accelerators within large AI systems and racks, ensuring predictable power and higher speeds.
The Technical Shift: NRZ, WDM, and Silicon-Centric Design
The OCI specification isn’t just about replacing a wire with a fiber; it’s a fundamental architectural change. The new standard focuses on several key technologies:
- NRZ Signaling and WDM: The OCI uses Non-Return to Zero (NRZ) modulation and Wavelength-Division Multiplexing (WDM). This allows multiple wavelengths of light to carry data simultaneously.
- Scaling Bandwidth: The technology starts at four wavelengths × 50 Gb/s (200 Gb/s per direction) and is designed to scale up to 800 Gb/s per fiber.
- Silicon-Centric Model: Instead of relying on bulky external modules, OCI moves toward tighter integration of optics with compute and networking silicon. This approach, often associated with Co-Packaged Optics (CPO), reduces latency and increases bandwidth density.
- Protocol Flexibility: The OCI physical layer (PHY) is designed to support various protocols, including Nvidia’s NVLink and the UALink standard used by AMD and Broadcom.
Strategic Impact on Hyperscalers
The move toward optical scale-up is already influencing the roadmaps of the world’s largest cloud providers. AWS is a primary example, acting as one of Nvidia’s largest NVLink Fusion customers with plans to integrate this technology into its next-generation Trainium4 compute clusters.
For companies like Meta and Microsoft, the transition to OCI means they can build larger, more efficient AI clusters without being limited by the physical distance that copper can travel before signal degradation occurs. This is critical for training the next generation of frontier models that require thousands of GPUs to act as a single, massive compute engine.
- Copper (e.g., NVL72): Low cost, extremely low power for very short distances, but limited by physical reach and cable bulk.
- Optical (OCI MSA): Higher scalability, supports much longer reaches within the rack, and uses WDM to dramatically increase bandwidth per fiber.
- Industry Alignment: The formation of the OCI MSA signals that the industry is moving toward an open standard to avoid vendor lock-in for optical physical layers.
Conclusion: The Future of AI Hardware
The transition from copper to optical scale-up marks a pivotal moment in AI infrastructure. Although copper served as the foundation for the first wave of LLMs, the path to super intelligence requires a silicon-centric optical approach. By embracing CPO and the OCI standard, Nvidia and its partners are removing the physical barriers to scale, ensuring that the only limit to AI growth is the imagination of the developers—not the length of the cable.