Reframing AI acceleration: from electrons to photons in the Fourier domain

The rapid growth of AI workloads is increasingly constrained by system power and bandwidth rather than arithmetic throughput. Convolution-heavy pipelines are limited by data movement and the cost of large linear transforms. Photonic Fourier-domain approaches offer a distinct path by implementing the Fourier transform through passive propagation, shifting the scaling bottleneck away from clocked digital switching. Building on the previously demonstrated photonic joint transform correlator (pJTC) prototype [1], this invited paper focuses on how pJTC-type accelerators can scale and integrate into practical systems. We first establish a system-level energy and throughput landscape, contrasting electronic FFT-based convolution with photonic Fourier optical convolution and formulating an end-to-end break-even condition that accounts for E/O and O/E interfaces, loss-driven optical power, thermal stabilization, and peripheral overheads. We then position pJTC relative to other integrated photonic compute styles and identify why convolution-specialized Fourier optics provides a clear scaling axis, while remaining sensitive to mixed-signal and packaging constraints. Finally, we recast prior WDM scaling projections through a packaging-aware lens, identifying key breakpoints and proposing a two-dimensional (2D) Fourier-lens pJTC architecture targeting spatial scaling (patent pending). These results provide a roadmap-oriented blueprint for when and how photonic convolution accelerators can complement electronics in edge and cloud AI.