Oxford – November 03, 2026 -- The Advanced Research and Invention Agency (ARIA), a UK public R&D funding body, will deploy Lumai's Iris inference servers as the first optical inference platform in its Scaling Inference Lab. The deployment will test whether optical compute can cut data center power consumption by 90% in a commercial environment.
Lumai claims one-tenth total cost of ownership versus GPUs
The Scaling Inference Lab is run by CommonAI CIC, a non-profit focused on collaborative AI development. Researchers will use the Iris servers to evaluate optical compute under standard power, cooling and rack integration. Lumai says the servers could cut total cost of ownership to around one-tenth of conventional GPU architectures. The headline claims for the technology also include processing speeds up to 50x higher.
Iris runs production-scale large language models (LLMs). It performs AI computation within beams of light in three-dimensional space rather than on silicon transistors. Numbers are encoded as light intensity, and passive optical components such as lenses carry out the mathematical operations. Lumai says the energy used for the computation itself is negligible and barely rises as performance scales.
"Power, not demand, now limits how fast AI can be deployed," said Dr. Xianxin Guo, Lumai's co-founder and CEO.
Iris targets the compute-bound prefill stage and pairs with silicon for decode
Large-model inference runs in two phases: prefill, which is compute-bound, and decode, which is memory-bandwidth-bound. Specialized inference hardware has increasingly focused on decode. Lumai instead targets the matrix-heavy prefill stage before the first token is generated, performing millions of matrix operations in a single optical cycle.
Iris is not a silicon replacement. Paired with silicon for decode, it delivers at least 16x better energy efficiency per inference than silicon-only systems, according to Lumai.
ARIA calls Lumai the most mature optical compute platform it evaluated
David Lewis, ARIA's Chief Science and Technology Officer, said the technology could make inference more energy-efficient and less dependent on constrained global hardware supply chains. He said the lab allows that potential to be tested openly and independently at rack scale.
ARIA was created by an Act of Parliament and is sponsored by the Department for Business, Innovation, Science and Trade. The Scaling Inference Lab is backed by £50m from ARIA. It addresses the lack of an open platform for startups and research groups to showcase AI hardware, given the dominance of large, closed systems.
Power limits drive the UK sovereign compute argument
Gavin Ferris, CEO of CommonAI, said the UK must build sovereign AI compute capacity within the power it has. The lab will test technologies such as Lumai's on real workloads alongside today's systems. The release says UK capacity depends on getting more from limited power and grid capacity and on bringing UK-developed technology into real data centers.
The release estimates that serving the computation ambitions of a single leading AI lab with conventional accelerators alone would cost $100 trillion within five years. It would also require more than 1,000 GW of additional energy.
The ARIA deployment follows the April 2026 launch of Iris Nova, Lumai's first product available for evaluation and the first optical AI accelerator engineered for data center-scale inference. Lumai spun out of University of Oxford optics research in 2021.