Analog optical computers achieved 94.6% balanced accuracy on mortgage approval classification from 5.84 million U.S. HMDA records, trailing XGBoost by 3.3 percentage points with 5,126 parameters including 1,024 optical (Berloff et al., arXiv:2604.13251, 2026). The digital twin isolated three accuracy-loss layers: binary encoding dropped all models to 89.4-89.6%, widening channels from 16 to 48 recovered only 0.5 points, and seven calibrated hardware non-idealities added zero penalty. Prior optical demonstrations remained confined to small image benchmarks such as MNIST via diffractive networks (Lin et al., Science, 2018).

The mortgage benchmark extends optical inference to real-world tabular financial data, exposing an architectural bottleneck over hardware fidelity that earlier coverage of photonic systems overlooked in favor of raw speed metrics. This connects to post-silicon development patterns including IBM's analog in-memory computing chips for inference efficiency (IBM Research, 2021) and Lightmatter's photonic processors targeting reduced energy per operation. GPU-centric AI infrastructure faces documented power scaling limits, with inference workloads projected to drive substantial electricity demand (de Vries, Joule, 2023).

By demonstrating viable accuracy on million-record datasets without hardware-induced loss, the work identifies encoding schemes and optical core design as immediate improvement targets. It fits an under-reported shift from digital silicon toward hybrid analog-optical-neuromorphic architectures that prioritize energy efficiency for sustained AI scaling beyond current GPU paradigms.