Physics Practical Magic

A simple 'shape' change just made future AI computer chips 100 times more durable.

April 15, 2026

Original Paper

Increased endurance of nonvolatile photonics enabled by nanostructured phase-change materials

arXiv · 2604.08637

The Takeaway

The best candidate for future AI chips—phase-change materials—usually wears out after being used a few thousand times, making them impractical for real-world use. This paper shows that by simply 'nanostructuring' the material into tapered and segmented shapes, researchers reduced the energy loss by 94% and boosted the chip's lifespan to over 100 million cycles. This solves a major 'hardware bottleneck' that has kept light-based AI chips in the lab for years. For you, this means the 'AI in your pocket' is finally becoming possible—chips that can process complex neural networks using almost no power and lasting for a decade without wearing out.

From the abstract

The rapid rise of artificial intelligence, and in-memory computing has reinvigorated research on scalable, energy-efficient, and reconfigurable photonic hardware. Non-volatile phase-change materials (PCMs) are attractive, as they offer large refractive index contrast, wavelength-scale footprints, and zero static power consumption. However, current PCM-based electrically controlled photonic devices are plagued by high insertion loss and low endurance. One prevalent hypothesis for these material l