Researchers unveil an optical device that performs AI calculations faster and more efficiently than conventional electronics, targeting data center power demands.

As data centers worldwide grapple with surging electricity demand from AI workloads, researchers from Pennsylvania State University (Penn State) have unveiled a prototype light‑powered computer they say could dramatically cut the energy needed to run artificial intelligence.

The device, described in a new paper in Science Advances, uses an “infinity mirror”‑style loop of compact optical components to encode data into beams of light, and capture the resulting patterns with a microscopic camera — thereby performing AI‑style calculations faster and with far less power than conventional electronic systems.

The Penn State team has said the technology targets the core bottleneck in optical computing: achieving the nonlinear mathematical operations that underpin AI decision‑making, without relying on exotic materials or high‑power lasers. Instead, their technology setup repeatedly recirculates light through inexpensive, widely available optical elements similar to those found in LCD displays and LED lights, allowing the light patterns to “build up” the required nonlinear relationships over multiple passes.

In tests, the optical system matched the accuracy of fully nonlinear digital neural networks on standard image‑classification tasks, including 96.82% accuracy on the MNIST handwritten‑digit dataset under incoherent white light.

Projections cited by Penn State and other outlets suggest AI‑driven data centers could consume more than 13% of global electricity by 2028, underscoring the urgency of more efficient computing approaches.

The researchers view their prototype as a proof‑of‑concept for an optical module that could plug into existing computing platforms and offload the most energy‑intensive AI operations from power‑hungry graphics processing units. They estimate an industry‑ready, programmable version could emerge in roughly two to five years, depending on investment and application focus.

Other researchers believe such optical accelerators could one day complement traditional silicon hardware by executing vast numbers of operations in parallel at higher speeds and lower energy, helping ease the cooling and cost burdens now facing AI data centers.