An experimental all-optical recurrent neural network handled computing tasks at up to 80 GHz with accuracy above a purely linear model, surpassing conventional CPU clock rates that have plateaued near 5 GHz.
The system keeps linear operations, nonlinear activation and memory entirely in the optical domain, avoiding electronic bottlenecks that have limited faster real-time processing of ultrafast signals.
In tests, it classified noisy waveforms with 97.5% accuracy at 10 GHz, 92% at 50 GHz and still beat random guessing at 120 GHz; its nonlinear advantage held through 80 GHz.
The prototype also classified microresonator soliton states with 95.6% accuracy in under 100 ns, forecast simple time series at up to 10 GHz, and generated MNIST-style sevens from quantum noise.
Researchers said integrated thin-film lithium niobate components could push the architecture toward terahertz rates, though current training still relies on digital electronics and some tasks remain hybrid electro-optic.
With training still bottlenecked by electronics, how far are we from truly autonomous, all-optical AI systems?
As optical AI solves the energy crisis, what new ethical safeguards are needed for terahertz-speed intelligence?
Beyond Silicon: How 100 GHz All-Optical Computers Are Redefining the Future of Computing
Overview
For years, CPU clock speeds have been stuck around 5 GHz due to physical and architectural limits like the breakdown of Dennard scaling and the von Neumann bottleneck, which made it hard to boost performance by simply increasing speed. As a result, the industry shifted focus to multi-core designs. In 2026, researchers broke this barrier by demonstrating an all-optical computer running at over 100 GHz, using light instead of electricity to process information. This breakthrough overcomes old bottlenecks and opens the door to much faster and more efficient computing, marking a major leap forward in computer technology.