Artificial intelligence infrastructure is expanding rapidly, with total spending reaching $318 billion last year, according to IDC. Beyond increased investment, the way data flows through these systems is beginning to show the growing pains. Moving data between processors is just as resource-intensive as computing itself, and in many cases, it's starting to determine the performance of the entire system.

This shift is evident within data centers. The ever-increasing volume of data transferred between xPUs, memory, and storage devices is putting immense pressure on bandwidth and power budgets. Traditional data transfer methods—electrical interconnects—are becoming increasingly strained as systems expand across racks and clusters.

It is in this context that silicon photonics (SiPh) and co-packaged optics (CPO) have emerged. By placing optical connections closer to the chip, electrical paths are shortened, thereby increasing interconnect density and integration. This shift is significant as systems become increasingly complex and efficient data transfer becomes crucial.

New restrictions: Data movement

Artificial intelligence data centers are expanding in all directions at an unprecedented pace. There are more chips in packages, more nodes in racks, more racks in clusters, and the volume of data transfer between data centers is constantly increasing. This growth has led to a dramatic increase in east-west traffic; data is no longer simply flowing in or out, but is continuously exchanged between GPUs and accelerators.

For decades, electrical connections have played a vital role, but they are increasingly becoming bottlenecks, resulting in what the industry calls "copper walls." Indeed, transmitting more data over greater distances requires significantly more power and makes signal integrity more difficult to maintain, especially outside of a single rack. For reference, scale-up data center architectures require less than 2 pJ of power per bit and latency of less than 1 μs.

The industry is responding. Fiber optic connections are appearing more frequently in data centers, and as data rates continue to climb, the growth rate of fiber optic interconnects is outpacing the number of processors they support, with demand growing rapidly. The expansion of computing power is only one aspect of the story; the efficiency of data transfer in distributed systems is equally crucial. Achieving high energy efficiency in data transfer can save the power resources required for computing, optimize processor performance, and maximize the performance of cloud-based artificial intelligence.

Why Silicon Photonics is on the Rise

Silicon photonics technology is gaining increasing attention because it aligns perfectly with the future direction of data centers. Based on semiconductor CMOS technology, it leverages silicon—the simplest and most easily manipulated material in the semiconductor industry—allowing optical functions to be more tightly integrated into computing, rather than being treated as a separate component.

Optical connections can handle significantly more data than electrical connections and are not limited by distance. This is especially important as systems scale up, particularly as companies struggle to meet ever-increasing bandwidth demands while avoiding excessive power consumption.

Investment is driving this growth, with the silicon photonics market projected to grow at approximately 25-30% annually, reaching $10 billion by the end of this decade. This forecast stems directly from the growing demand for artificial intelligence and cloud infrastructure, coupled with increasing pressure on data transmission speeds.

The application of silicon photonics technology is also shifting significantly. Components previously located at the periphery of systems, such as pluggable optical devices, are now moving closer to the processor. Co-packaged optical devices are a prime example, integrating optical and electronic components into a single package closer to the GPU (xPU) or switching ASIC, thus reducing data transmission distances.

This is partly because silicon photonics is not entirely unfamiliar territory. It builds upon the same manufacturing and supply chains as the semiconductor industry, facilitating its large-scale deployment.

In summary, these factors make silicon photonics no longer a niche solution, but a core component of next-generation AI infrastructure.

As AI systems scale and become increasingly distributed, efficient data transmission has become a critical constraint, not an afterthought. Silicon photonics offers a viable path for data centers to meet bandwidth demands while avoiding power overload—a challenge that traditional electrical solutions struggle to address on a large scale. The next generation of data centers is emerging. They are more integrated, have shorter electrical paths, and increasingly rely on optical connections to ensure smooth system operation. The future trajectory depends on how these systems are built and scaled in practice. AI infrastructure can remain sustainable if data can flow freely without adding extra complexity or energy consumption.