The conversation around artificial intelligence often revolves around the chips that power these advanced systems. However, there is a lesser-known bottleneck emerging within data centers, specifically concerning how these chips communicate with each other. As AI clusters expand to include tens or even hundreds of thousands of processors, the traditional copper wiring that has been the backbone of computer connectivity for decades is reaching its limits.
The solution to this bottleneck lies in silicon photonics, and the investment in this technology is rapidly gaining momentum. Two companies, in particular, are poised to benefit significantly from this shift.
Why Copper is Struggling in AI Clusters
Training a modern AI model requires thousands of chips to function as a single massive brain, constantly transferring vast amounts of data between them. As more chips are added to the cluster, the amount of data traffic flowing through the connecting wires increases. This is where copper wiring begins to falter. At the high signaling speeds demanded by these systems, a passive copper cable can only maintain a clean signal for a short distance before degradation occurs. To compensate for this, more electrical power needs to be supplied to preserve the data integrity, leading to increased heat generation and higher energy costs.
In smaller server setups, this may not pose a significant issue. However, in large warehouse-sized clusters spanning rows of racks, it becomes a physical barrier known as the “copper wall.” When the bottleneck is limited by centimeters and watts, adding more copper is not a viable solution.
How Silicon Photonics Overcomes the Barrier
Silicon photonics provides a solution by transmitting information as light pulses through fiber optics instead of electrons through metal wires. Light can travel further, carry more data, and consume less power over longer distances, making it ideal for large AI clusters. Co-packaged optics represent the cutting edge of this transition, where optical components are integrated directly next to the switch chip rather than being externally connected. This integration minimizes power loss and maximizes bandwidth within the same space.
While copper wiring is not becoming obsolete, its role is shifting towards shorter connections within a package. Optics, on the other hand, are taking over at the board and rack scale. The financial investment in this transition is substantial, with the optical interconnect market for AI data centers expected to grow significantly in the coming years.
Investing in Silicon Photonics Companies
Two companies, Coherent and Lumentum, are well-positioned to capitalize on the shift towards silicon photonics. Coherent is a leader in optical technology for AI data centers, with a growing demand for its data center transceivers. Lumentum supplies lasers and optical components for data center networking and is expanding its capacity to meet the rising demand for lower-power optics and tighter silicon photonics integration.
While these companies offer promising opportunities for investors, it’s essential to consider the cyclical nature and volatility of the optical component market. Valuations are high due to the enthusiasm surrounding AI, and competition in the optics industry is fierce.
In conclusion, the transition from copper to light-based communication is not just a possibility but a necessity as AI clusters continue to expand. Investing in companies like Coherent and Lumentum offers a focused way to participate in this technological shift, albeit with higher risks and potentially higher rewards. It is crucial to approach these investments with an understanding of the optical business’s inherent volatility and the long-term trend towards silicon photonics in AI data centers.
