Run neural-network inference at a fraction of the energy, on silicon photonics
Great Sky builds photonic processors that carry inference on light instead of electrons. For data-center operators running large inference fleets, that means far lower power draw and heat per token served, at the throughput your models already demand.
Inference fleets are running into a power ceiling
Every transformer layer you serve multiplies matrices in transistor logic. Each transistor switch dissipates heat, draws current, and loads capacitance. Multiply that by billions of inferences per day and you have a data-center power problem that scaling GPU hardware alone cannot solve.
The bottleneck is not compute speed. It is energy per multiply-accumulate operation. Moving those operations into optical interference on silicon waveguides changes the physics of the cost equation.
See the PhysicsInference in three optical steps
Every matrix multiplication in your model runs as light propagating through a silicon waveguide mesh.
Weights encoded in light
Modulator banks set the amplitude and phase of coherent light to represent your neural network weight values. No electrons, no resistive switching.
Waveguide mesh performs dot products
Light propagates through a Mach-Zehnder interferometer mesh. Optical interference computes matrix-vector products at the speed of light, passively, without switching transistors.
Photodetectors read the result
Photodetector arrays convert the optical output back to electrical signals. On-chip ADCs digitize the inference result and hand it back to your CUDA stack.
Your GPU stack stays. Our chip handles the matrix math.
The Great Sky photonic processor slots into a PCIe Gen 4 x16 slot alongside your existing GPU nodes. There is no model recompilation, no inference framework replacement. The SDK intercepts matrix-multiplication calls and routes projection layers to the photonic co-processor transparently.
- Half-height half-length form factor, fits standard 1U and 2U server slots
- CUDA 12+ compatible, existing CUDA driver stack unchanged
- Handles MLP fully-connected layers and transformer attention projection
- Operating wavelength 1310 nm, compatible with standard SMF-28 fiber
Built by photonics researchers
Photonics researcher who spent a decade studying optoelectronic devices before recognizing the inference energy crisis as the right application for silicon photonic computing. Founded Great Sky in 2023 to build that bridge.
Silicon photonics IC design specialist with a PhD from KTH and prior experience in waveguide layout and MZI optimization at two photonic ASIC firms. Co-founded Great Sky to take that device expertise from transceiver applications into inference acceleration.
PhD researcher in photonic computing theory whose doctoral work on phase noise in MZI mesh networks informs the GSK-1's precision architecture. Connects the academic photonic computing literature to the practical constraints of production silicon.
Join the evaluation program
We are shipping evaluation units to a limited number of qualified data-center operators. If you run large inference fleets and want to measure the energy impact directly, reach out. This is a research collaboration, not a SaaS sign-up.