LightScale builds ultrafast photonic network switches for AI data center interconnects, using non-thermal optical switching matrices that reconfigure at packet cadence. The company's Hyperlane fabric enables topology-on-demand networking—reshaping network architecture within a single training step—while its LUMA software layer exposes switching and compute through one API. The initial product, Scale 1, is a 64×64-port photonic switch pilot designed for bandwidth-heavy AI workloads like training and inference serving.
Funding
Funding not disclosed
Founders
Product
Problem
AI workloads impose extreme communication demands on data center networks, but conventional electrical switches and fixed optical fabrics force a compromise: they either consume high power, add latency, or lock operators into a single network topology that poorly matches the varied communication patterns of distributed training, expert routing, and inference. Rapidly changing workload phases—gradient reduction, all-to-all token exchange, pipeline streaming—each prefer different network shapes, yet traditional fabrics cannot adapt quickly enough to serve them all efficiently.
Solution
LightScale delivers ultrafast photonic switching for AI interconnects, enabling optical networks that reconfigure at packet cadence rather than as slow control-plane actions. Its Scale 1 switch uses a non-thermal photonic switching matrix integrated at the device level, avoiding the thermal load and per-port power overhead of conventional reconfigurable optics. The Hyperlane architecture exposes topology, routing, and link assignment to software, letting operators reshape the fabric inside a training step—not just between jobs—so each workload phase gets the network shape it wants. The LUMA software layer provides a single API spanning switching and compute, so developers write once and run across all hardware generations without per-device rewrites or changes to existing PyTorch code.
Target Audience
Primary customers are AI infrastructure operators—including cloud providers and enterprises running large-scale training, inference serving, and MoE routing—as well as developers building distributed machine learning systems who need adaptive, low-power network fabrics.
Features
- Non-thermal photonic switching matrix integrated at device level, scaling switch radix without per-port thermal overhead
- Packet-cadence reconfiguration fast enough to change topology within a collective operation, not just between runs
- Hyperlane exposes topology, routing, and link assignment to the control plane, with real-time congestion and link-health feedback into reconfiguration decisions
- 64 × 64 pilot switch (SXNR-001-64) in 1U chassis with 96-core ARM CPU, 256 GB DDR5 RAM, and dual ConnectX-7 200/400G NICs; higher port counts up to 768 × 768 planned
- LUMA ships as a library that slots beneath PyTorch's collective and runtime layers, requiring no new operators, rewrites, or model changes
- Supports a broad range of network topologies including ring, mesh, 2D/3D torus, hypercube, fat-tree, dragonfly, butterfly, Slim Fly, Jellyfish, and Xpander