AI at the speed of light just became a possibility (techxplore.com)

Researchers at Aalto University have demonstrated "single-shot" tensor computing that performs complex AI arithmetic in a single propagation of coherent light, a paper appearing in Nature Photonics (2025). Led by Dr. Yufeng Zhang and Prof. Zhipei Sun, the team encodes digital values into the amplitude and phase of optical fields; as these light fields interfere and combine, they naturally execute matrix and tensor multiplications—operations underlying convolutions and attention layers—without stepwise electronic processing. By using multiple wavelengths the approach generalizes to higher-order tensors, and because the computation happens passively during light propagation, no active electronic switching is needed. The result is effectively instantaneous, massively parallel tensor math at the speed of light, promising orders-of-magnitude improvements in latency, throughput and energy efficiency compared with GPU-based platforms constrained by electronic bottlenecks. Technically, it maps numbers to optical degrees of freedom and leverages coherent interference for linear algebra primitives central to deep learning, making it compatible with a range of optical platforms and amenable to on-chip photonic integration. The team estimates 3–5 years to integrate with existing hardware ecosystems, positioning optical tensor processors as a practical path toward next-generation AI/AGI accelerators that dramatically reduce power and scale limits of today’s systems.