SK hynix has outlined a roadmap for co-packaged optics, or CPO, that could extend the company’s role in AI infrastructure beyond HBM and into the high-speed optical links connecting processors, memory and large-scale computing systems.

The research, published in Nature Electronics under the title “Co-packaged optics for high-performance computing and artificial intelligence,” addresses a problem becoming increasingly important as AI clusters grow: computing power is advancing faster than the networks that move data between computing resources.

HBM helped overcome one of the first major bottlenecks of the generative AI era by placing high-bandwidth memory close to AI accelerators. But as AI infrastructure expands from individual accelerators to thousands of GPUs and HBM stacks operating across racks and pods, the bottleneck is shifting.

The next challenge is moving enormous amounts of data between them.

Compute Is Scaling Faster Than Connectivity

According to the research highlighted by SK hynix, computing performance is increasing by roughly threefold every two years, while interconnect bandwidth is growing by only about 1.4 times over the same period.

The widening gap creates what researchers call a “bandwidth wall.”

Even if individual GPUs become substantially faster, the performance of an AI cluster cannot increase at the same pace if data cannot move quickly enough between processors, memory and other accelerators.

The problem becomes particularly significant once signals leave a semiconductor package.

Traditional copper-based electrical connections remain economical over short distances, but higher speeds and longer transmission distances increase signal loss and require additional circuitry. That raises power consumption and latency — two increasingly critical constraints for next-generation AI data centers.

CPO offers a different approach.

Instead of relying on long electrical connections, the technology brings optical transceivers, or optical engines, much closer to processors and integrates them within the same package.

Electrical signals travel only a short distance before being converted into light, allowing most of the remaining journey to take place optically.

This could enable AI systems to maintain high bandwidth density and energy efficiency as connectivity expands from chips to racks and eventually to entire computing pods.

A Target of More Than 100 Tb/s per Node

The significance of the Nature Electronics paper lies not only in identifying CPO as a potential solution but also in defining targets for future AI infrastructure.

The researchers envision systems delivering more than 100 Tb/s of bandwidth per node while consuming less than 1 pJ per bit and maintaining chip-to-chip latency below 10 nanoseconds.

They also map a packaging path from 2D integration to 2.5D interposers and eventually 3D heterogeneous integration.

The direction is clear: optics is moving progressively closer to computing silicon.

That evolution has parallels with the rise of HBM.

HBM improved AI computing efficiency by bringing memory physically closer to processors and dramatically increasing the bandwidth between them. CPO seeks to apply a similar principle at a larger scale by reducing the distance traveled by high-speed electrical signals and using optical links for longer connections.

If HBM addressed the bandwidth problem around the processor, CPO could help address the next bottleneck across the AI system.

What Happens When Memory Is Connected by Light?

The longer-term vision described by the researchers goes further than connecting processors optically.

It extends optics directly into the memory architecture.

Under an “optics-centric” architecture, optical interposers could directly connect pools of computing resources, such as XPUs, with large pools of memory.

Instead of requiring memory resources to remain tightly coupled to individual accelerators, multiple processors could potentially access shared memory resources through extremely high-bandwidth optical connections.

Such an architecture could become increasingly important as AI models continue to grow.

Physical packaging imposes limits on how many processors and memory devices can be placed close together. Optical connectivity offers a potential path toward separating those resources physically while allowing them to operate as part of a much larger computing system.

In that scenario, the defining metric of an AI data center would no longer be only how powerful each GPU or HBM stack is.

The ability to make thousands of processors and memory devices behave like one enormous computing resource could become equally important.

Why This Matters for SK hynix

The paper also offers a glimpse into how SK hynix is thinking about its position in the next phase of AI infrastructure.

The company has become one of the central beneficiaries of the generative AI boom through HBM, which has emerged as a critical memory technology for AI accelerators.

But increasing memory bandwidth alone cannot eliminate every bottleneck in a rapidly expanding AI cluster.

The next stage requires memory devices, controllers, optical components, processors and advanced packaging technologies to be designed together rather than optimized independently.

That is why the researchers emphasize “co-design.”

The study was conducted by researchers from SK hynix, the University of Virginia, the University of Illinois Urbana-Champaign, Nanyang Technological University, MIT and Yonsei University. Seunghoon Hong of SK hynix’s AI Infra team and University of Virginia professor Kyusang Lee participated as corresponding authors.

Lee said faster computing chips alone cannot improve overall system performance if data movement between chips fails to keep pace, describing optical transmission as one of the most promising ways to overcome the physical limitations of copper interconnects.

The transition will not happen overnight.

Low-power photonic integration, coherent protocols, packaging and reliability technologies still present significant engineering challenges. Researchers nevertheless see optical interconnect technology as moving beyond laboratory demonstrations and toward the early stages of industrialization.

For SK hynix, that transition could carry strategic significance.

HBM established memory as one of the defining technologies of the first phase of the AI infrastructure boom. Optical interconnects could help determine how the next generation of those memory and computing resources are assembled into much larger systems.

The AI race is therefore expanding beyond the competition to build faster chips.

The next battleground may be the architecture that connects them — and increasingly, that connection could be made with light.

Yeon Choul-woong More by this reporter