Optical I/O is the practice of moving data into and out of a chip or package over light instead of over copper wires. Today almost all chip-to-chip communication uses electrical SerDes driving signals down metal traces, but copper attenuates high-frequency signals badly over distance, so electrical links are stuck with short reach and rising energy cost as data rates climb. Optical I/O converts the electrical bits to modulated light, sends them across an optical fiber or waveguide, and converts them back — trading copper's reach-and-energy wall for the near-lossless, high-bandwidth physics of photons. For large AI systems trying to wire together thousands of accelerators, it is increasingly seen as the way past the interconnect bottleneck.\n\n``svg\n\n``\n\nThe motivation is that electrical links are hitting a wall. A PCB trace or cable loses more signal the faster you push it, so beyond roughly a meter an electrical link needs heavy equalization and burns significant energy per bit — and the bandwidth you can cram through the edge of a package (the "shoreline" or beachfront) is capped by how many copper pairs physically fit. Light does not attenuate the same way: an optical fiber carries enormous bandwidth over meters to kilometers at low loss, and many wavelengths can share one fiber. Optical I/O attacks reach, bandwidth density, and energy per bit all at once.\n\nA link is a chain of electrical-to-optical conversions. On the transmit side, a modulator (often a compact silicon ring resonator, or a Mach-Zehnder modulator) imprints the electrical data onto a beam of light supplied by a laser. The modulated light travels down a fiber or on-chip waveguide. On the receive side, a photodetector (typically germanium on silicon) turns the light back into current, and a trans-impedance amplifier recovers the electrical bits. The laser light itself usually comes from an external laser source (ELS) rather than being generated on the die, because efficient lasers are hard to build in silicon.\n\nWavelength-division multiplexing is the bandwidth multiplier. Because light of different colors does not interfere, many independent data channels can ride the same fiber at once, each on its own wavelength, using an array of ring resonators tuned to different colors. This WDM trick is what lets a single fiber carry terabits per second, and it is central to why optical I/O achieves such high bandwidth per millimeter of die edge compared with copper.\n\nThe figures of merit are energy, shoreline density, and reach — not just raw speed. Optical I/O is judged on picojoules per bit (it must beat electrical SerDes to be worth the complexity), on shoreline bandwidth density measured in terabits per second per millimeter of die edge, and on reach. Where electrical links top out around a meter, optical links keep their signal over meters to kilometers, which is exactly what disaggregated, rack-scale systems need.\n\nPackaging is marching the optics toward the die. The progression runs from pluggable optical transceivers at the faceplate, to co-packaged optics (CPO) that place the optical engine right next to the switch or accelerator ASIC on the same substrate, to fully in-package optical I/O where the optical interface is a chiplet sitting beside the compute die. Each step shortens the electrical path to the optics, cutting energy and boosting density — which is why CPO and in-package optical I/O are among the most watched technologies for next-generation AI fabrics.\n\n| Element | Job |\n|---|---|\n| Modulator (ring / MZM) | imprint electrical data onto light |\n| Laser source (ELS) | supply the optical carrier |\n| Fiber / waveguide + WDM | carry many wavelengths far, at low loss |\n| Photodetector + TIA | convert light back to electrical bits |\n| Packaging (pluggable→CPO→in-package) | move optics closer to the die |\n\nRead optical I/O through a beat-the-copper-wall lens rather than a faster-cable lens: the point is not simply speed but escaping the reach, energy, and shoreline-density limits that cap electrical SerDes at the package edge. Once the optical engine moves onto the package and light replaces copper for chip-to-chip links, bandwidth stops falling off with distance — which is precisely what lets an AI cluster grow from a board into a rack into a fabric without the interconnect becoming the bottleneck.\n