The surge in artificial intelligence isn’t just about more powerful processors; it’s about how those processors link together. According to this article from Wired.com, the real bottleneck today lies in the networking inside and between chips and racks in data centers.
Traditional electrical interconnects (wires, copper traces, electronic signaling) are reaching physical and power-efficiency limits as the amount of data moved per chip and between servers grows dramatically. To keep pace with the demands of massive AI model training and inference, companies are turning to photonics, using light (photons) instead of electrons, to move data faster and more efficiently.
Start-ups such as Lightmatter, Celestial AI, and PsiQuantum are developing silicon-photonic engines to link chips in three-dimensional stacks and server racks, aiming to deliver a “photonics future” for AI hardware. Meanwhile, legacy players such as NVIDIA and Broadcom, aware of the significance of networking, are making acquisitions and pushing into new interconnect architectures.
But the transition isn’t immediate or guaranteed. Photonics demands novel packaging, new manufacturing processes, and integration with existing electrical systems, all of which raise cost, complexity, and risk. The article suggests that while the potential payoff is huge—massive speed gains and lower latency—the returns may take years to materialize consistently in volume.
For engineers working on AI infrastructure, the shift means design focus is expanding: not just how fast a chip computes, but how fast it communicates. The networking layer has become a crucial piece of the AI-stack puzzle.