ARIA to Test Lumai Optical Computing in UK AI Infrastructure Lab
The UK’s Advanced Research and Invention Agency (ARIA) will deploy optical inference servers developed by Oxford-based Lumai in its Scaling Inference Lab, giving researchers a chance to assess light-based AI processing in a data-centre environment. The system, called Lumai Iris, is designed to perform parts of AI inference using beams of light rather than relying entirely on electronic calculations in silicon chips.
The trial puts Lumai optical computing into a live infrastructure testbed, where its performance can be examined alongside conventional systems. Lumai says the technology could reduce data-centre power consumption by up to 90% and deliver processing speeds up to 50 times higher. Those figures are company claims and will need to be assessed under the lab’s testing conditions.
How Lumai optical computing works
Lumai’s Iris servers use a three-dimensional optical architecture to carry out matrix operations, a core part of AI workloads. The system encodes numerical information in light and uses optical components to perform calculations. This approach is intended to handle compute-intensive tasks without requiring every operation to pass through conventional silicon transistors.
The technology is designed particularly for the prefill stage of large language model inference. During prefill, a model processes the input prompt before generating its first response token. This stage is compute-intensive, while the later decode stage is more dependent on memory bandwidth. Lumai’s approach is therefore designed to work alongside conventional processors, rather than replace silicon across the entire inference process.
ARIA lab will test the technology at rack scale
ARIA’s Scaling Inference Lab is operated by CommonAI CIC and backed by £50 million in funding. It provides an open environment where emerging AI hardware can be integrated into rack-level systems and tested against commercial alternatives. The goal is to assess how new technologies perform under practical workloads, including their impact on power use, cost and overall system performance.
Lumai’s Iris servers will be the lab’s first optical inference platform. ARIA’s chief science and technology officer described the company’s system as the most mature optical compute platform evaluated by the agency, while stressing that the lab will provide a way to test its potential independently at rack scale.
Energy efficiency is the central test
Data-centre electricity demand has become a major constraint on AI expansion. As companies deploy larger models and serve more inference requests, operators need ways to increase computing capacity without proportionally increasing energy consumption. Optical processing is one possible route, although its practical value depends on performance across complete systems, not just individual calculations.
Lumai says its Iris architecture could achieve at least 16 times better energy efficiency per inference when used for prefill alongside silicon-based hardware for decode. The company also claims the servers could reduce total cost of ownership to around one-tenth of conventional GPU architectures. These are not yet independently established outcomes from the ARIA deployment.
The evaluation will help determine whether optical computing can move beyond specialised demonstrations and contribute to production-scale AI infrastructure. The results will matter not only for Lumai, but also for efforts to diversify the hardware used to run AI systems and ease pressure on data-centre power supplies. Readers can find the company’s official Lumai Iris product information for further technical details.
