Rebuilding AI from first principles with superconducting optoelectronic networks

Rather than accelerating the matrix operations performed by GPUs, at Great Sky we build a new AI computing architecture that implements neural networks directly in physical hardware using superconductors and light. Synapses, dendrites, and neurons are physical devices embodied in analog Josephson circuits rather than numbers computed digitally using a conventional Von Neumann computer architecture. Information moves between these computing nodes via optical signals at the single-photon level — a feat made possible by superconducting detectors. We refer to these systems as superconducting optoelectronic networks (SOENs). The hardware incorporates local, distributed memory that’s accessed as processing occurs without the need for calls to RAM. Optical communication allows each neuron to make direct, dedicated connections to many thousands of synapses without the constraints of a shared, digital routing fabric. Our modeling indicates the total cost of ownership per output token is 300x less than Nvidia GB200 NVL72 at equivalent capability, a figure combining measured device energies with explicit scaling assumptions, including 4K refrigeration overhead. Because GPU attention scales quadratically with the sequence length while our architecture’s cost scales linearly, the advantage widens for more data-intensive modalities like video, and with input resolution. Since our engineering team began work in early 2025, we’ve completed four tape-outs, with our latest chip implementing a high-speed recurrent neural network used for dynamical system analysis and control tasks. We’re now partnering with imec to accelerate our manufacturing roadmap.