Article contents:
- Introduction: the bottleneck in the AI era
- Why co-packaged optics? The problem with copper cables
- The optical solution: how CPO technology works
- Nvidia’s plan: Quantum-X and Spectrum-X
- TSMC’s pivotal role: the COUPE roadmap
- Not an option but a necessity: the future of data centers
- Conclusion: accelerating AI at the speed of light
Introduction: the bottleneck in the AI era
Modern artificial intelligence clusters require thousands of graphics processing units (GPUs) working as a single integrated system. But as these clusters grow, communication between these units becomes the biggest bottleneck. The massive data transfer demands between GPUs are driving a radical shift: using light instead of electricity for communication. Nvidia has revealed its ambitious plans to use silicon photonics and co-packaged optics (CPO) technology by 2026, in a move that may become mandatory for the next generation of AI data centers.
Why co-packaged optics? The problem with copper cables
In large AI clusters, the distance between servers and network switches must be very short to maintain low latency. But this design limits cluster size. The solution is to move switches to the end of the row, allowing a larger number of servers to be connected. However, this significantly increases the distance, making copper cables impractical at high speeds such as 800 Gbps. Here, optical connections become essential.
The optical solution: how CPO technology works
Co-packaged optics avoid the penalties associated with traditional pluggable optical modules. Instead of the electrical signal traveling across the board and connectors before being converted to light, CPO embeds the optical conversion engine directly next to the switch ASIC. The signal is converted to light almost instantly. As a result, electrical loss is drastically reduced, and power consumption per port drops from 30 watts to just 9 watts. This design eliminates many points of failure and greatly simplifies the implementation of optical interconnects.
Nvidia’s plan: Quantum-X and Spectrum-X
Nvidia will introduce CPO-based optical connectivity platforms for both Ethernet and InfiniBand technologies.
- Quantum-X InfiniBand: Scheduled for launch in early 2026. Each switch will deliver 115 Tbps of throughput, supporting 144 ports operating at 800 Gbps each.
- Spectrum-X Photonics: Scheduled for launch in the second half of 2026. It will rely on the Spectrum-6 chip powering two devices, one delivering 102.4 Tbps and the other reaching 409.6 Tbps.
Nvidia expects these switches to eliminate thousands of discrete components in AI clusters, offering faster installation, easier maintenance, and lower power consumption.
TSMC’s pivotal role: the COUPE roadmap
The development of Nvidia’s initiative is closely tied to the evolution of TSMC’s COUPE (Compact Universal Photonic Engine) platform. TSMC’s roadmap unfolds in three phases:
- First generation: An optical engine for OSFP connectors, delivering 1.6 Tbps.
- Second generation: Transitions to CoWoS packaging with onboard optics, enabling 6.4 Tbps at the motherboard level.
- Third generation: Aims for 12.8 Tbps within processor packages, with further reductions in power and latency.
This means the performance of Nvidia’s optical platforms will continuously improve as TSMC’s technology advances.
Not an option but a necessity: the future of data centers
Nvidia stresses that co-packaged optics are not an optional upgrade, but a “structural requirement” for future AI data centers. This means the company will position its optical interconnects as a major advantage over competing AI solutions, such as those offered by AMD. This explains why AMD recently acquired Enosemi, a company specializing in this field.
Conclusion: accelerating AI at the speed of light
Nvidia’s plans to use light for communications represent a revolutionary step in AI infrastructure design. By overcoming the limitations of electrical interconnects, CPO technology opens the door to building larger, faster, and more energy-efficient AI clusters. It is no longer just about the power of individual GPUs, but how they communicate together. In the future, artificial intelligence will be powered, literally, at the speed of light.
Frequently asked questions
Q: What is silicon photonics?
A: It is a technology that uses silicon to create optical components capable of transmitting data using light. This is faster and more energy-efficient than using traditional copper wires to transmit electrical signals.
Q: What is the main advantage of co-packaged optics (CPO)?
A: The main advantage is a dramatic reduction in power consumption (from 30 watts to 9 watts per port) and increased reliability by integrating the optical engine directly next to the switch chip, minimizing signal loss and component count.
Q: When can we expect to see this technology in the market?
A: Nvidia plans to release its first CPO-based switches in early 2026.
