Computing with light: photonics comes for AI's interconnect bottleneck.

The dirty secret of frontier AI training is that the chips are rarely the bottleneck — the wires between them are. As clusters scale to hundreds of thousands of accelerators, copper interconnects hit hard physical limits on bandwidth, reach, and power. The industry's answer is to replace electrons with photons, as close to the silicon as possible.

The current benchmark for that transition: Lightmatter's March 2026 demonstration of 1.6 terabits per second over a single fiber, achieved by pairing its Passage co-packaged optics chiplet with Qualcomm's 112G SerDes and multiplexing 16 wavelengths of light down one strand — a claimed 8x bandwidth-density improvement over existing optical packaging. By June 2026, Lightmatter had joined NVIDIA's NVLink Fusion ecosystem, a strong signal that co-packaged optics is entering the mainstream of AI cluster design rather than remaining a specialty technology.

Photonics is also moving from moving data to computing on it. In April 2025, Nature published a photonic processor running real neural workloads — ResNet vision models, BERT language models, and deep reinforcement learning agents — at accuracy approaching conventional electronics. Analog computation with light, long a laboratory curiosity, now executes the same architectures the rest of the industry ships.

This intersection is the one to watch. The models the field most wants to train — continual-learning world models, brain foundation models over petabyte-scale neural recordings — are interconnect-hungry in a way current fabrics cannot economically serve. We think the labs that own their optical stack will set the pace of the next scaling era, the way the labs that owned GPU clusters set the pace of the last one.