When Copper Stops Scaling: Co-Packaged Optics and the Future of AI Infrastructure

Artificial intelligence has revived an old problem in computing: moving information. For decades, the semiconductor industry focused primarily on computation itself. Transistors shrank, clock speeds increased, and GPUs accumulated staggering numbers of parallel cores. But modern AI systems are increasingly constrained not by the arithmetic inside the chips, but by the movement of data between them.

A modern AI training cluster may contain tens of thousands of GPUs, all communicating continuously. Every generated token, every gradient update, and every synchronization step depends on an enormous web of interconnects carrying information between processors, switches, and memory. As these systems scale, the energy required to move bits through the data center has begun to rival the energy required to compute them. And this is where physics begins fighting back.

Copper and the physical limits of electrons

The modern data center was built on copper. Copper is cheap, manufacturable, mechanically robust, and naturally compatible with electronic systems. For short distances, electrical interconnects remain extraordinarily effective. But at the signaling speeds demanded by modern AI infrastructure, copper no longer behaves like an ideal wire. It behaves like a lossy radio-frequency transmission line.

As frequency increases, several physical effects begin to dominate [1]. The first is resistance: current flowing through copper dissipates energy as heat, and at higher frequencies this worsens because of the skin effect, where current crowds near the surface of the conductor, reducing the effective cross-section available to carry the signal. The surrounding dielectric also begins absorbing energy, and nearby traces couple electromagnetically, producing crosstalk, reflections, and timing distortion.

At the highest data rates, copper channels can lose tens of decibels over only a few meters. Digital signal processing can compensate, but it consumes substantial power. The industry has entered a feedback loop where increasing bandwidth requires increasingly complex electronics simply to preserve signal integrity.

Why light carries data better

Optical communication solves many of these problems by moving information onto light instead of electrical current.

The difference begins with carrier frequency. Electrical systems typically operate in the gigahertz regime; optical carriers operate hundreds of thousands of times higher, near 200 terahertz. This enormous frequency space allows vastly larger communication bandwidths [2].

Optical fiber is also an extraordinarily efficient medium. Standard single-mode fiber has attenuation near 0.2 dB per kilometer, effectively negligible compared to copper at comparable bandwidth densities. Information can travel long distances without repeated amplification, regeneration, or equalization.

Optics offers another advantage: multiplexing. A single fiber can simultaneously carry many independent wavelength channels through wavelength division multiplexing (WDM), each wavelength acting like an independent lane of traffic through the same strand of glass. Dense WDM has long been the foundation of long-haul telecom, and these techniques are increasingly moving into hyperscale computing infrastructure.

This transition has also revived debates [3] around coherent optics inside the data center. Traditional datacenter links have largely relied on intensity modulation and direct detection (IMDD), which is simpler and lower power. Coherent systems encode information onto both the amplitude and phase of light and require substantial digital signal processing. In long-haul telecom, the added complexity is justified by spectral efficiency and reach; inside AI data centers, DSP power consumption itself becomes a major constraint.

The result is a growing architectural tension: the industry increasingly needs optical communication bandwidth, but it must deliver it within extremely constrained thermal and power budgets.

Co-packaged optics

Even today, many 'optical' data center links remain partially electrical systems.

In traditional architectures, optical transceivers sit at the front panel of a switch. Electrical signals must still travel significant distances across the printed circuit board between the switch ASIC and the optical module. At modern bandwidths, these PCB traces are becoming one of the dominant sources of loss and power consumption.

Co-packaged optics attempts to eliminate this bottleneck. Instead of placing optics at the edge of the system, it integrates optical engines directly beside the switch ASIC. The goal is simple: minimize how far electrons travel before becoming photons.

This reduces electrical trace lengths, lowers equalization requirements, improves bandwidth density, and potentially cuts total interconnect power substantially. In many ways, co-packaged optics is less about improving optics than about escaping the limitations of electrical signaling.

But integrating optics beside some of the hottest chips ever manufactured introduces a new class of engineering problems. Switch ASICs already dissipate enormous heat, while lasers, modulators, and photonic devices prefer stable temperatures. For this reason, most CPO architectures keep the laser source external to the package, housed in the switch chassis or front panel, feeding continuous-wave light into the co-packaged optical engines via fiber, so that laser degradation or failure does not require replacing the entire switch assembly. Packaging complexity rises dramatically when fibers, photonics, and high-power electronics must coexist in the same module, and reliability, serviceability, and manufacturing yield become critical concerns.

The challenge is no longer merely transmitting light. It is integrating optics into an environment historically designed entirely around electronics.

The datacenter becomes an optical machine

A growing ecosystem is now racing to define this transition. Broadcom has emerged as a leading supplier of co-packaged optical switches, shipping 51.2 Tbps CPO systems that integrate silicon photonics beside switch ASICs and reduce optical interconnect power by 70% compared to pluggable transceivers [4]. NVIDIA has made a major push into silicon photonics, announcing co-packaged optical networking integrated with its Quantum and Spectrum switches at GTC 2025 [5]. Marvell, Coherent, and Lumentum are building the broader photonic and DSP ecosystem, while startups such as Ayar Labs and Lightmatter push optical interconnects directly into the ASIC package, targeting chip-to-chip communication at the scale of individual processors and accelerators, not just network switches.

Not all of the most consequential technologies here are the most obvious. VCSELs, microLEDs, multimode optics, polymer waveguides, and even plastic optical fiber remain relevant because hyperscale systems care as much about manufacturability and cost as raw performance.

The modern data center is beginning to resemble less a room full of computers and more a large optical communication system attached to accelerators. For decades, computing scaled by shrinking transistors. The next era may scale by minimizing the distance electrons travel before becoming light.