ARIA brings Lumai's optical AI servers into a commercial data-centre testbed

Lumai's optical inference servers are heading into a publicly funded test of emerging AI hardware in a commercial data-centre environment. The Oxford company announced on November 3 that Iris will be the first optical computing platform evaluated in ARIA's Scaling Inference Lab, run by nonprofit CommonAI CIC. Lumai claims potential reductions of up to 90% in data-centre power consumption. The programme will test how its equipment performs as part of an operating system of racks, memory, networking and cooling.[1][2]

ARIA already backs the lab with £50 million. Its published plan is to assemble new rack-level pilots every six months, bringing at least one unproven component into a system that users can actually run. This is a different environment from a component-level demonstration. A new processor must work with networking, memory, power delivery, software and surrounding accelerators, not merely perform a mathematical operation efficiently on a test bench. Standardising the setting also allows comparisons against competing systems and helps purchasers separate theoretical gains from whole-system results.[2]

Where the light does the work

Large-language-model inference is commonly separated into two phases. In 'prefill', the system processes the incoming prompt and prepares the information needed to begin answering. In 'decode', it generates successive output tokens while repeatedly reading model memory. Lumai targets the highly parallel matrix computations in prefill. It encodes numerical values as light intensity and uses optical components, including lenses, to perform transformations. Conventional silicon still handles work such as decoding, memory and system control. Describing Iris as a wholesale GPU replacement would therefore misunderstand the architecture.[1]

Lumai says its hybrid approach can deliver at least sixteen times better energy efficiency per inference than silicon-only systems, and it suggests roughly one-tenth the total cost of ownership under favourable circumstances. Those metrics require separate denominators. Energy per request differs from total site power, particularly when a server is idle or output decoding dominates. Cost of ownership additionally includes rack space, maintenance, networking, lifetime, software support and utilisation. A system could be exceptionally efficient at prefill and still yield a smaller improvement for a workload that spends most of its time producing long answers.[1]

The workload mix will determine how much of that advantage reaches a customer. Long prompts make prefill a larger part of the job; lengthy generated answers place more weight on decoding. Comparable trials therefore need to state model size, input and output lengths, numerical precision, batching, latency and the boundary at which electricity is measured. Measuring the full system also captures memory, data movement and other costs that a calculation-only benchmark leaves out.[1][2]

Why public money is supporting a hardware contest

The lab addresses the gap between a working prototype and a cloud customer's purchasing decision. Novel accelerators must fit existing software, cooling and servicing arrangements before their mathematical efficiency becomes useful. ARIA's £50 million programme makes some of that integration work a shared resource. The UK AI Hardware Plan commits at least another £20 million to expanding the lab, adding to the agency's original allocation.[2][3]

Lumai introduced Iris Nova for evaluation in April 2026. The November announcement advances the next stage: planned deployment of Iris in a commercial-environment programme. ARIA's selection is a meaningful route to operational testing, although no independent whole-system results for the new installation accompany the announcement. Those results should show which workloads benefit, and whether energy savings survive the cost of running the complete service.[1][2]

Report an error or suggest a correction