In a significant breakthrough for energy-efficient computing, Microsoft researchers have unveiled a novel analog optical computer (AOC) capable of dramatically improving energy efficiency for artificial intelligence inference and combinatorial optimization tasks—potentially boosting performance per watt by up to 100 times. The development, led by a team at Microsoft Research in Cambridge in collaboration with researchers from the University of Cambridge, was published in Nature on September 3, 2025. As global demand for AI and data-intensive computing grows, the energy consumption of traditional digital computers has become a pressing concern. Conventional systems rely on electronic signals switching between binary states, which leads to high power usage and bottlenecks in data movement between memory and processors—known as the von Neumann bottleneck. To overcome these limitations, scientists have explored alternative computing paradigms, with analog optical computing emerging as a promising frontier. Microsoft’s new AOC leverages the physical properties of light to perform computations in a fundamentally different way. Unlike general-purpose digital computers, this AOC is a specialized system designed to accelerate two critical tasks: AI inference and complex optimization problems. Instead of relying on discrete electronic transitions, the AOC uses continuous optical phenomena. Light from a microLED array represents input data, which passes through a network of lenses and a spatial light modulator (SLM). Each pixel on the SLM can be individually programmed to modulate the intensity of the light, effectively performing matrix-vector multiplication—the core operation in many AI and optimization algorithms. The modulated light is then focused onto a photodetector array, where the summation step is completed in parallel. This entire process happens in real time at the speed of light, enabling massive parallelism without the energy cost of digital signal conversion. Analog electronic circuits handle nonlinear activation functions, feedback loops, and iterative refinement—completing the computation pipeline. A key innovation is the “rapid fixed-point search” mechanism. In AI inference, the system iteratively converges toward a stable output state, or “fixed point,” which corresponds to the model’s prediction. For optimization problems, the fixed point represents the optimal or near-optimal solution. This iterative nature makes the system robust to the inherent noise of analog computing, as each iteration pulls the result toward the correct attractor state. The team validated the AOC’s capabilities through real-world applications. In collaboration with Microsoft’s Health Futures group, they used the AOC’s digital twin—a software model simulating the hardware’s behavior—to accelerate MRI image reconstruction. The results suggest a potential reduction in scan time from 30 minutes to just 5 minutes, with major implications for patient care and healthcare efficiency. In finance, Microsoft partnered with Barclays to tackle complex transaction settlement problems—classic combinatorial optimization challenges involving legal, credit, and timing constraints. In a test case with 46 transactions and 30 constraints, the AOC hardware successfully found the global optimum, outperforming previous quantum computing attempts on the same problem. The prototype also demonstrated success in machine learning tasks, including image classification using the MNIST and Fashion-MNIST datasets, as well as nonlinear regression. Results showed over 99% consistency between the AOC’s output and simulations from its digital twin, proving that models trained in digital environments can be effectively deployed on analog optical hardware. Currently, the prototype handles up to 256 weights, with potential for scaling to 4,096 through advanced techniques. Researchers believe larger systems—capable of managing hundreds of millions or even billions of weights—could be built using modular designs. The system’s components, such as microLEDs, optical lenses, and image sensors, are based on commercially available, mass-produced technologies found in smartphones and consumer electronics, lowering manufacturing costs and enabling scalable production. The team envisions a future system composed of 25 modular units, capable of processing 100 million weights with a power draw of around 800 watts. Such a system could achieve 400 peta-operations per second (Peta-OPS), delivering over 500 trillion operations per watt (TOPS/W)—a performance level more than 100 times more efficient than today’s leading GPUs for comparable tasks. While challenges remain in scaling and integration, the team is optimistic. The combination of high efficiency, low-cost components, and proven performance in real-world scenarios positions analog optical computing as a transformative technology for the next generation of AI and optimization systems.