A New Era for AI: Microsoft Unveils an Optical Computer That’s 100x More Efficient

Microsoft Research’s Cambridge team has unveiled the Analog Optical Computer (AOC), a light-based system assembled from off-the-shelf components such as smartphone camera sensors, Micro LEDs, and lenses, which demonstrated speeds and energy efficiency surpassing GPUs by a factor of 100 and earned publication in Nature.

After four years of development, the team merged optics with analog circuits within a single loop, executing matrix operations in light while handling non-linear functions electronically, with each iterative computation taking about 20 nanoseconds.

In collaboration with Barclays Bank, the system solved a financial settlements optimisation problem in seven iterations, while in healthcare it transformed MRI data into an optimisation framework for compressed sensing imaging, reconstructing 32 × 32 brain slices and enabling a digital twin that simulated 200,000 MRI variations.

The study suggests scan times could be reduced from 30 minutes to five, easing patient burden, and notes that AOC’s fixed-point search is well suited to iterative convergence in models like Deep Equilibrium Networks, achieving nearly 99% accuracy against digital simulations on several benchmarks.

Microsoft estimates that a mature AOC could reach 500 TOPS per watt, vastly outperforming NVIDIA’s H100 at roughly 4.5 TOPS per watt, with potential expansion to 100 million to 2 billion weights as a low-power infrastructure for AI inference. The project is led by Francesca Parmigiani, Jiaqi Chu, and Jannes Gladrow, with Hitesh Ballani emphasising the aim of releasing an open digital twin to engage a wider research community.