AINEWS
Photonic optical circuit switching (light based, dedicated on demand paths, like 1960s telephone switchboards) gives AI training the timing stable connections that best effort internet routing can't deliver at scale.
The biggest bottleneck in next-gen AI is no longer chips or raw bandwidth. It is the network fabric inside the data center, the part that moves gradients and activations between thousands of GPUs while a training run is in progress. The fix being trialed at scale looks like a 1960s telephone switchboard: dedicated, always-on connections you set up on demand, instead of best-effort packet routing shared across the modern web.
The mechanism is called photonic optical circuit switching (OCS), and a Spanish startup named iPronics just raised $125 million to scale it for AI data centers, the company announced on September 2, 2026. The Register's exploratory feature on the same day reads the same trend as a switchboard comeback, this time etched into silicon photonics.
To understand why, it helps to translate the plumbing. A modern data center runs on packet switching: every message is chopped into small chunks, each chunk is addressed independently, and routers along the way do their best to deliver the chunks quickly and in order. That works for web traffic, video calls, and email, where the workload is bursty and statistically multiplexed. An AI training run is the opposite. Thousands of accelerators need to exchange partial results in lockstep, hundreds of times per second, in patterns shaped by collective operations like all-reduce, the synchronized averaging step that lets a distributed model train as one. Drop a packet or add a few microseconds of jitter, and the whole cluster waits. The traffic looks less like email and more like a scheduled conference call: deterministic, latency-sensitive, and unhappy with shared best-effort paths.
Photonic OCS reframes the fabric as a programmable grid of optical paths. Instead of forwarding packets through a packet switch, an operator sets up a dedicated optical circuit between two endpoints, runs the workload over that circuit, and tears it down when the job changes. The switching happens in silicon photonics, where light beams on a chip are steered by tiny actuators to reconfigure the network on millisecond timescales. The system behaves like a crossbar telephone exchange, except the operators are schedulers and the wires carry laser light.
iPronics is the most visible commercial bet on this idea at the moment. The company calls its product the "World's First Silicon Photonics Optical Circuit Switch" for AI-driven data centers, a claim that, like most "World's First" announcements, is company-stated and has not yet been independently benchmarked against rivals. The September raise follows an earlier €20M Series A, roughly US$20M at typical exchange rates of the time, announced on the company's site, and is, per the company's product page and about section, aimed squarely at hyperscaler and large enterprise AI builds.
The engineering argument for the switchboard revival is real, but it is not settled. Photonic OCS trades flexibility for raw speed and power efficiency on the optical paths it does support; a circuit that is set up for a training job is not a circuit that is being used for something else. Schedulers have to plan ahead, and workloads that do not behave like collective AI training do not benefit. The orthodoxy being challenged, the "just add more bandwidth and let the routers sort it out" school, has decades of investment behind it and is not standing still. Hyperscalers are also building their own fabrics, from NVLink-style proprietary interconnects (NVIDIA's direct GPU-to-GPU links) to in-house photonic research, which may end up displacing third-party OCS in the largest deployments.
What changes if the switchboard revival is right? The cost, location, and identity of the next AI buildout shift. Data centers optimized for collective AI traffic value fiber plant and scheduler software over raw bisection bandwidth, the maximum data a network can move per second between any two halves of the cluster. A handful of operators with the capital to lay down programmable optical fabric pull further ahead of everyone renting best-effort cloud. The bottleneck moves from "did the GPU ship?" to "can the network keep up with the GPU?" That is the case The Register is floating, and the case iPronics is now funded to test.
The first independent signal will come from deployments, not press releases. A hyperscaler that publishes latency, jitter, and total-cost-of-ownership numbers for a photonic OCS fabric at production scale would turn the "may hinge on" into "is hinging on." The iPronics deployment slate, and the rivals that name names this year, are where to watch.