Vision
The next layer of compute is light
AI is no longer limited by arithmetic. It is limited by how fast, and how cheaply, data moves between the chips doing the arithmetic. LightScale is building that layer in light: switches that reconfigure at packet cadence, links that carry die-to-die traffic over light at any reach, and one programming model across both. The same photonics point further, to optical compute and to switching for quantum computing.
The problem
Compute kept scaling. The network didn't
Where the time goes
Models outgrew the chip years ago. A training step or a served token now runs across thousands of accelerators, and a growing share of every step is spent waiting on the network: gradients to reduce, experts to exchange tokens, activations to arrive. The arithmetic finishes early and sits idle.
Where the energy goes
Today's fabrics are electrical switches joined by optical cables. Every hop turns light into electrons and back, and every lane that leaves a package becomes a serial link at several picojoules per bit. Multiply that by every port in a cluster and moving data becomes a power budget of its own.
And the shape of that network is fixed the day the racks are cabled. Every parallelism strategy wants a different topology, so a fixed fabric picks one compromise at build time, and every workload pays for it on every step.
What we believe
Light for the data path, software for its shape
The switch stays optical
A photon that enters a switch should leave it as a photon. Scale1 reroutes any input to any output entirely in light, on a non-thermal switching matrix built to scale radix without the thermal load or per-port power overhead of conventional reconfigurable optics.
Distance stops costing energy
Photon1 keeps its electronics at logic speed and lets light carry the distance, so the energy per bit barely moves whether the far die is a millimetre or a rack away. One link serves the package, the board and the rack.
The fabric follows the workload
Hyperlane exposes topology, routing and link assignment to software and reconfigures at packet cadence, fast enough to reshape the network inside a training step. Each phase gets the network it wants, and the code never notices.
What we're building
One platform, from the die to the pod
Scale1
Any-to-any switching at packet cadence, entirely in light. The 64 × 64 pilot ships in Open Rack V3 and as a 1U dev kit, with higher port counts to follow.
Explore Scale1 Optical die-to-die interconnectPhoton1
A chiplet beside a GPU or memory controller that carries its die-to-die traffic over light, in the package, across the board or across the rack, with the host's protocol layers unchanged.
Explore Photon1Pod1
Full racks of Scale1 switches, connected by an all-optical fabric. Dedicated switching capacity for AI infrastructure, shipped one rack at a time.
LUMA
The software layer above the hardware. One API spans switching and compute, PyTorch first, beneath the framework's collective and runtime layers, so existing code runs unchanged.
Explore LUMAThe horizon
A fabric that is light end to end
-
Now
Scale1 in the rack
The SXNR-001-64 pilot, 64 × 64, in Open Rack V3 and as a 1U dev kit, with LUMA above it, PyTorch first.
-
Next
More ports, whole racks
Higher port counts, then Pod1: racks of Scale1 switches joined by an all-optical fabric, so a pod is dedicated switching capacity that lowers into place.
-
Then
Light between dies
Photon1 brings the same optics inside the package, so reticle-sized dies, their memory and the racks around them work as one system, with the host's protocol layers unchanged.
-
Beyond
Light beyond the interconnect
The same photonics reach into compute itself. Optical compute, where the matrix operations at the heart of AI run in light. And quantum computing, where information already travels as photons, so a switch that routes them in light, with no conversion to electrons, is what joins quantum processors into systems. LUMA spans switching and compute from Scale1 onward, so the fabric keeps changing underneath and the code written today keeps running.
The next layer of compute, starting with Scale1
Photonic switching at packet cadence, optical links that hold their energy at any reach, and one programming model across both. The first pilot is under way.