For the better part of two decades, the pluggable optical transceiver was the workhorse of data center networking. A compact, hot-swappable module, insert it, pull it, replace it. Elegant in its simplicity. The industry built a global supply chain around it. Then AI happened.

AI workloads are far more demanding that traditional internet traffic. They require massive dataflows between GPUs and very low latency, to keep computations synchronized. The shift from training a language model on a few thousand GPUs to orchestrating a cluster of half a million has exposed a fundamental physical constraint: copper can no longer move data fast enough, or cheaply enough in power terms, to keep the compute fed. That constraint pushed optical technology out of its old home in long-distance telecom networks and into the middle of the AI accelerator stack. And it has given a once niche packaging concept called co-packaged optics a very large spotlight indeed.

At modest data rates, an electrical trace on a circuit board is the cheapest, most reliable way to move a signal a few centimeters. But push the line rate past roughly 100 gigabits per second per lane, and two things happen at once. Signal loss over distance rises sharply, and the correction electronics needed to recover a clean signal at the other end, equalisers, retimers, digital signal processors, start burning more power than the data itself is worth. Industry engineers have taken to calling this the I/O wall: compute density keeps compounding through advanced packaging, but the electrical links feeding that compute do not scale at the same rate.

In practical terms, a high speed electrical signal can degrade meaningfully within about a meter of ordinary circuit board material. That is shorter than the distance from a switch chip to the front panel of the very box it sits in. Once the reach a signal needs to travel exceeds what copper can carry cleanly, the only sensible option is to convert it to light.

Strip away the marketing language and every optical interconnect performs the same basic conversion. Electrical data arrives from a chip. Driver electronics condition that signal. A laser supplies a steady stream of light. A modulator encodes the data onto that light. Fiber, or increasingly an on-chip waveguide, carries the signal to its destination, the physical handoff point shown in Figure 2. A photodetector converts the light back into an electrical current, and receiver electronics recover the original bits.

None of this is new physics. What is new is where the industry needs this chain to work reliably, at what volume, and at what cost. A long haul telecom link can afford an exotic, hand-tuned laser because it serves an entire city's traffic. An AI cluster needs thousands of these links per rack, manufactured to a consistent standard, serviced by technicians who are not laser physicists. The chain is genuinely only as strong as its weakest link, and in commercial deployment that weak link is rarely the laser itself. It is coupling loss, thermal drift, and packaging yield at scale.

Optical fiber has connected data centers to one another for years without controversy, that is the easy part. The interesting engineering, and the interesting business opportunity, is in how far into the system optics can be pushed. Today's deployed technology links racks and switch fabrics with pluggable modules at the front panel. The next step moves optical engines onto the board itself, shortening the electrical hop before conversion. Beyond that, some designs place a small photonic chip right inside the processor's own package, next to the processor itself. And at the switch layer, co-packaged optics places the optical engine directly next to the switching silicon, cutting the electrical path to a few millimeters instead of tens of centimeters.

Each step closer to the compute buys back power and bandwidth density, and each step also makes the system harder to build, test and repair in the field. That trade-off, not the underlying physics, is what is currently being fought over at conferences, in vendor roadmaps, and inside hyperscaler procurement teams.

The commercial numbers are starting to reflect the urgency. One market research estimate puts the optical interconnect segment for AI data centers over 30 billion dollars by 2035, growing at fourteen percent a year as co-packaged optics moves from prototype to deployed infrastructure. Vendors are not waiting for that curve to arrive. Marvell has already expanded its optical DSP portfolio to cover 400G, 800G and 1.6T class links for AI fabrics, and 800G deployments are scaling through 2026 while 1.6T moves from roadmap slide to shipping product. Future roadmaps already point toward 3.2 terabits per second per fiber.

It is not an unqualified success story. Analysts covering the AI chip supply chain have flagged optical interconnect testing and silicon photonics manufacturing capacity as an emerging bottleneck in their own right. Advanced packaging, the precise process of connecting fiber to a chip, and optical testing capacity are all finite. They are now competing for the same limited resources as the accelerators they are meant to feed.

For procurement teams reading a switch vendor's spec sheet this year, the practical takeaway is not that plugables are dead. Copper still works in plenty of places, and even the most aggressive co-packaged optics roadmaps keep copper inside the rack. It stays there because it is cheaper and easier to service. Hybrid architectures, part copper and part optics, are the honest near term reality. But the direction of travel is clear. Laser fabs, packaging houses, and fiber attach specialists are all being pulled into a new design cycle, which is now being set by AI scaling.