Summary

• Lumai Iris will become the first optical inference platform evaluated in ARIA's Scaling Inference Lab, operated by CommonAI CIC.
• The architecture uses optical processing for the compute-intensive prefill stage of AI inference, while conventional silicon handles token generation.
• Lumai claims major energy and cost reductions, but the laboratory's independent testing has not yet established those results.

Oxford computing company Lumai will supply optical AI inference servers to a British research facility designed to test emerging computing architectures under commercial data centre conditions.

The company’s Iris hardware has been selected as the first optical inference platform for the Scaling Inference Lab, backed by the Advanced Research and Invention Agency (ARIA) and operated by CommonAI CIC. Researchers will assess its performance using real AI workloads and infrastructure incorporating conventional rack, power, and cooling systems.

Lumai claims its approach could reduce data centre electricity consumption by up to 90% in suitable deployments, with further improvements in the cost of operating AI services. Those figures remain supplier projections rather than results independently established by the laboratory.

The planned deployment addresses a practical obstacle facing companies developing alternatives to conventional AI accelerators: hardware that performs well during a laboratory demonstration may behave differently when incorporated into a complete computing system.

Optical processing targets the first inference stage

Language-model inference involves several kinds of computation, with different demands on processing capability and memory. During the initial prefill stage, a model processes the incoming prompt and its associated context before beginning to generate a response.

This stage involves substantial parallel mathematical processing, including matrix operations that conventional accelerators perform electronically. Lumai instead uses light travelling through optical components to execute selected calculations, encoding information through optical intensity.

The company argues that passive optical components can perform some of these operations with considerably less energy than electronic circuits. However, an optical computing system still requires conventional hardware for memory, data conversion, control, and other functions.

Consequently, the electricity needed for an optical calculation is not equivalent to the total energy consumed by a server delivering an AI service. The complete system must be measured before meaningful comparisons can be made with established computing products.

Lumai’s Iris architecture is designed specifically for prefill rather than the entire inference process. Conventional silicon accelerators continue handling the decode stage, when a language model generates successive output tokens.

Decode typically places greater demands on memory bandwidth and the movement of information between processing components. Separating the two stages allows specialised hardware to concentrate on operations for which its architecture may offer an advantage.

However, the division also introduces integration requirements. Data must move between optical and electronic processing systems efficiently, while software must coordinate the work without introducing delays that outweigh improvements in individual calculations.

Testing moves beyond isolated components

ARIA’s Scaling Inference Lab was established to examine whether emerging computing technologies remain commercially useful when deployed as complete systems.

Supported by £50 million from ARIA, the laboratory is being delivered by CommonAI CIC as an open testing environment for new accelerators, memory systems, interconnects, and other computing technologies.

Its programme involves constructing experimental computing racks and comparing their performance with existing commercial systems. ARIA’s published approach envisages a new rack approximately every six months, each incorporating an unproven technology that can be evaluated through practical AI workloads.

These evaluations extend beyond processor speed. Researchers must account for the electricity consumed by servers, cooling requirements, networking, software compatibility, reliability, and the ability to sustain performance across different applications.

Such factors can substantially affect the economics of deploying computing equipment. A processor that completes a particular mathematical operation efficiently may provide limited commercial benefit if integration introduces additional energy consumption, latency, or operating costs.

For Lumai, the assessment will involve Iris servers operating in a commercial data centre environment. The laboratory is intended to provide independent evidence about their behaviour at rack scale rather than relying exclusively on benchmarks selected by the technology supplier.

David Lewis, ARIA’s chief science and technology officer, described the purpose of the evaluation: “The Scaling Inference Lab provides the means to test that potential openly and independently at rack scale.”

The programme provides a route for emerging hardware developers to demonstrate capabilities without first securing the resources required to build a complete commercial computing installation themselves.

Energy projections face independent examination

Lumai claims its hybrid optical and silicon approach could deliver at least 16 times better energy efficiency per inference than conventional silicon-only systems under suitable conditions.

It also projects that the technology could reduce data centre power requirements by as much as 90% and bring total ownership costs to approximately one-tenth of those associated with conventional GPU architectures.

These estimates should not be interpreted as independently verified results. Their commercial relevance depends on the models being processed, the proportion of inference work handled optically, system utilisation, and the infrastructure needed to support the servers.

Total cost of ownership is particularly sensitive to deployment assumptions because it incorporates equipment expenditure, electricity, cooling, maintenance, supporting infrastructure, and the useful life of the installation.

A lower energy requirement for one processing stage may not produce an equivalent reduction in total system costs, particularly where other components continue to operate at conventional power levels.

The laboratory’s findings should help establish whether the claimed benefits remain visible across complete workloads, rather than individual operations selected to demonstrate the hardware’s strongest performance.

Britain’s computing constraints add commercial pressure

The evaluation comes as Britain seeks to expand domestic AI computing capacity while facing pressure on electricity networks, development sites, and access to specialist hardware.

More efficient inference equipment could allow operators to deliver additional computing within existing power limits. That would be valuable where electrical connections or cooling infrastructure restrict the capacity available for new installations.

However, improvements in efficiency do not necessarily reduce overall electricity consumption. Lower processing costs can encourage organisations to run more AI workloads, potentially increasing aggregate demand even as individual tasks become less expensive.

Alternative accelerators also need support from software frameworks, system integrators, and equipment operators. Commercial adoption requires dependable maintenance, compatible applications, and evidence that performance remains stable outside controlled demonstrations.

Lumai previously introduced Iris Nova as an evaluation product in April 2026. The ARIA deployment will extend that development through a more demanding examination of system-level performance.

The laboratory has not yet published results establishing the projected energy savings or ownership-cost reductions. Its testing will provide evidence about whether optical processing can become a practical complement to conventional AI infrastructure and which inference workloads are most suitable for the technology.