'Optics When You Must' Arrives for Data Centers

Previously in this series, we argued that the network, not the processor, has become the control point for AI performance and that co-packaged optics is the front line. This week, the physics behind why, from our CEO Matt Crowley's piece in Laser Focus World.

There is a rule that has guided datacenter design for years: copper where you can, optics when you must. For a long time, "when you must" stayed comfortably in the future. Use cheap copper until physics forces an upgrade. As AI clusters scale toward million-GPU factories, "when you must" has arrived, and it arrived faster than almost anyone planned for.

Network engineers have done extraordinary work pulling reach and bandwidth out of copper. But no amount of engineering overturns physics, and copper has two enemies that get worse at exactly the frequencies AI signaling now demands.

The first is the skin effect. As the electrical signal alternates faster, the current crowds into a thinner and thinner layer at the surface of the conductor. At around 53 gigahertz, a frequency common in current AI signaling, that layer is roughly 0.3 micrometers thick. The signal uses less than 1 percent of the wire, and the resistance climbs to more than 100 times its DC value.

The second is dielectric loss. At gigahertz frequencies, the molecules in a cable's insulation cannot keep up with the field as it reverses around them. The lag turns signal energy into heat. Put the two together, and over roughly 2 meters of high-quality cable at these frequencies, they can consume more than 90 percent of the signal's power budget.

This is not a manufacturing problem waiting for a better cable. It is a fundamental trade-off between bandwidth and reach, which is why copper reach is measured in meters while fiber reach is measured in kilometers.

Optics does not fight this fight. Photons have no electrons to crowd, far fewer ways to shed energy, and they stay organized over distance in a way copper cannot. Co-packaged optics (CPO) is the architectural answer: confine copper to the ultra-short chip-to-chip paths where it still wins, and hand everything, leaving the package to fiber. Power drops, reach extends, and density climbs right where the GPU needs it most.

In a few years, an all-copper AI datacenter will look as quaint as an all-copper long-haul internet. The pivot is not a matter of taste. It is what physics requires, and the teams reading the constraint correctly are already designing for it.

That settles whether optics. It does not settle how.

Read Matt Crowley's full piece in Laser Focus World. It carries the physics through to the consequences that decide datacenter ROI, from tail latency to GPU utilization and the economics they drive: "Optics when you must' arrives for data centers."

---

Next in the series, we take up the question that decides scale-up architecture: why the optical layer has to be multiplexed, and why adding fibers is a trap.