A fibre-optic ring and the bet on analogue hardware

Energy efficiency is still the crux of computing for artificial intelligence. Optical hardware has long promised speed and low power consumption, but it runs into a long-standing obstacle: backpropagating the error through analogue components is hard to do. A study published on 5 November 2026 in Nature Computational Science by a group led by associate professor Kohei Nakajima of the University of Tokyo proposes an alternative. According to the authors, the method, called SS-DFA (state-skipping direct feedback alignment), does not need access to the intermediate states of the physical system to propagate the error. It is paired with a single-map architecture in which many units follow general rules instead of complex connections. As doctoral student Yongbo Zhang put it, the aim was to rethink the learning process itself rather than keep building ever larger and more elaborate models.

The physical demonstration uses an optoelectronic circuit. Information is converted into a light signal, processed along a 200-metre fibre-optic ring and turned back into an electrical signal with a minimum of electronic components. The university's press release states that, in simulations and in the optical experiment, the simplified design matched more complex neural architectures, and in some tests outperformed them, with far less computation. The release gives no figures or benchmarks. The abstract of the preprint released in January 2026, under a different title from the final one, reported a reduction of about 40% in computing resources, with better performance, when the scheme is integrated into modern architectures. The release specifies that these are "relatively basic" architectures. The published version may report different numbers.

The release presents the actual effects on data-centre energy, cooling and water use only as possibilities, and the study did not measure them. The technique also has clear limits for now. Kohei Nakajima himself noted that the scheme has not yet been tested on the largest models, language and multimodal ones in particular, and that further work on miniaturisation, stability and large-scale manufacturing will be needed before any practical use.

Letting a signal do the computing along two hundred metres of optical fibre is an undeniably elegant idea, but the distance between a lab demonstration and what data centres need in operation remains wide. Cutting model complexity is an interesting conceptual attempt, as long as we don't mistake a theoretical possibility for an efficiency already measured in the field. — Olya