Dipole Labs aims optical circuit switches at the dead time GPUs spend waiting on the network

Between racks today, a transfer starts as light in a fiber, becomes electricity at a packet switch, and becomes light again on the way out. Each of those conversions consumes power, adds components and generates heat, according to runtimewire's account of the architecture ^[5]. An optical circuit switch skips the conversion by opening a direct light path between two selected endpoints ^[6].

The price of skipping it is that a circuit is a commitment. The path has to be established and changed as communication patterns move across the cluster ^[7]. Reconfigure too slowly and short, irregular transfers suffer. Reconfigure fast without capable control software and the switch spends its advantage connecting the wrong machines ^[7].

So the control layer decides whether the optics pay for themselves. Dipole pairs its hardware with software designed to learn GPU communication patterns and reconfigure the optical connections around training and inference workloads ^[3]. In its YC launch materials the company says the switches are targeting sub-microsecond reconfiguration at large port counts ^[8]. For that target to mean anything in a specific cluster, two things have to hold. The collectives have to hold one communication pattern long enough that a reconfiguration is cheap next to it, and the pattern has to repeat closely enough for a controller to anticipate the next one. runtimewire, whose primary source is Y Combinator, reports the figure as a target and gives no measured switching time, port count or insertion loss ^[8].

A company this young would not have published measurements yet. Dipole was founded in 2026 ^[1] and presented at the Summer 2026 batch's Demo Day on September 10th ^[9], so at most about eight months separate the company's first day from the pitch ^[17].

Deepankur Thureja and Gabriele Pasquale met as postdoctoral researchers at Harvard, after building quantum and nanophotonic devices at ETH Zurich and EPFL ^[15]. Thureja led a 2022 Nature paper on electrically tunable confinement of neutral excitons in two-dimensional semiconductors, reaching confinement lengths below 10 nanometers ^[12], and ETH Zurich gave him its 2024 ETH Medal for the thesis ^[13]. Pasquale's Nature Materials paper in January 2025 reported electrical detection of light's chirality in a two-dimensional material, work EPFL recognized with its 2025 IBM Prize ^[14]. runtimewire's own caveat is the right one. Those papers establish that the founders have spent years building and measuring difficult optical devices, and packaging, yield, insertion loss, thermal behavior and control electronics can erase a photonic device's theoretical advantage before it reaches a rack ^[16]. Confinement below 10 nanometers is evidence about fabrication discipline.

The commercial premise is that AI spending increasingly depends on extracting more work from installed accelerators, and that workload-aware optical switching can improve utilization without adding another row of GPUs ^[4]. TechCrunch put Dipole among the nine companies early-stage investors cited most often from the batch ^[10]. Cisco Investments lists Dipole in its portfolio. The amount and structure of the financing are undisclosed ^[11].