UK To Test Light-Based AI Servers That Promise Lower Power Consumption

Britain’s publicly backed Advanced Research and Invention Agency (ARIA) is preparing to test artificial intelligence servers that use beams of light to perform complex calculations, as demand for AI computing puts growing pressure on electricity supplies and data-centre infrastructure.

The agency has selected optical computing technology developed by Oxford-based startup Lumai for evaluation at its Scaling Inference Lab, according to an announcement made on November 3, 2026.

The project will test Lumai’s Iris inference servers in a commercial data-centre environment, allowing researchers to examine whether the technology can deliver meaningful improvements in processing speed, energy consumption and operating costs.

Lumai claims its technology could reduce data-centre power consumption by as much as 90% and offer substantial performance improvements over conventional computing architectures. However, these figures have not been independently established through the announced testing programme.

The development comes as governments and technology companies search for alternatives to increasingly expensive and power-intensive AI computing systems.

How Light Could Replace Some Electronic Calculations

Most AI systems today rely on electronic processors, including graphics processing units, or GPUs, to perform the mathematical calculations required to run large language models.

These processors use electrical signals moving through billions of transistors. Although highly capable, large clusters of such chips consume considerable electricity and require substantial cooling infrastructure.

Lumai is developing a different approach.

Its optical computing technology uses beams of light travelling through three-dimensional space to perform mathematical operations.

Instead of relying entirely on electrical circuits, the system encodes numerical information into light intensity and uses optical components, including lenses, to process it.

This approach is designed to perform many calculations simultaneously while reducing the energy needed for particular computing operations.

However, it does not eliminate conventional electronics.

Lumai’s servers combine optical components with silicon-based computing systems, allowing different parts of an AI workload to run on hardware suited to their requirements.

The company focuses particularly on a process called prefill, which occurs when an AI model reads and processes a user’s prompt before generating its response.

For example, when someone submits a lengthy document to an AI chatbot, the system must first process the supplied information before beginning its answer.

Lumai’s technology is designed to accelerate that initial calculation-heavy stage.

Traditional electronic processors remain responsible for other tasks, including generating subsequent output.

UK Research Lab To Test Technology Against Existing Systems

The planned evaluation will take place at ARIA’s Scaling Inference Lab, a £50 million initiative managed by the non-profit organisation CommonAI.

The laboratory was established to help emerging AI hardware developers demonstrate whether their technologies can operate effectively beyond controlled research environments.

Its approach involves integrating experimental hardware into complete server systems and testing them with practical AI workloads.

Lumai’s Iris servers will become the laboratory’s first optical inference platform.

Researchers will examine how the servers perform across multiple machines while operating within standard data-centre conditions, including conventional power supplies, cooling arrangements and equipment racks.

The testing will be important because performance demonstrated by an individual component does not necessarily translate into equivalent benefits across a complete computing system.

Real-world AI services must also handle memory access, communication between processors, software coordination and simultaneous requests from users.

These factors can affect processing speed and energy consumption even when a new computing component performs well during isolated experiments.

ARIA Chief Science and Technology Officer David Lewis described Lumai’s platform as the most mature optical computing technology evaluated by the agency.

The laboratory is intended to assess its potential independently, rather than accept the manufacturer’s performance estimates.

Can Optical Computing Reduce AI Infrastructure Costs?

Lumai claims its architecture could lower overall data-centre electricity requirements by up to 90% and reduce total ownership costs to approximately one-tenth of conventional GPU-based systems.

The company also reports that pairing optical processing with silicon-based systems could deliver at least 16 times greater energy efficiency per inference.

Its broader technology claims include the potential for processing speeds up to 50 times higher.

These measurements concern different aspects of the technology and should not be interpreted as interchangeable comparisons.

The announcement does not provide completed independent rack-level benchmark results establishing that such improvements can be achieved across commercial AI workloads.

That distinction is important because the electricity consumed by a complete data centre includes more than the calculations performed by its processors.

Memory systems, networking equipment, cooling infrastructure and power distribution also contribute to total consumption.

Consequently, reducing the energy required for one computing operation does not automatically produce an equivalent reduction in the electricity needed by an entire facility.

The Scaling Inference Lab will help determine whether Lumai’s claimed advantages remain significant when those additional requirements are included.

The project also has implications for Britain’s broader technology ambitions.

Countries developing domestic AI infrastructure face challenges involving expensive computing hardware, limited electricity-grid capacity and dependence on international semiconductor supply chains.

Lumai’s technology, which emerged from University of Oxford research and was developed into a company in 2021, represents one attempt to address those limitations through alternative computing architectures.

The company introduced Iris Nova, its first optical AI accelerator available for evaluation, in April 2026.

The ARIA programme now offers an opportunity to assess that technology in a more demanding environment before its potential for wider commercial adoption can be established.

What This Means For You:

More efficient AI hardware could eventually reduce the electricity and infrastructure costs of operating AI services. For now, the claimed savings remain unproven until independent testing establishes how the complete systems perform.