ARIA’s Scaling Inference Lab Puts Lumai’s Optical Computing to the Test

The UK’s Advanced Research and Invention Agency (ARIA) will deploy Lumai’s Iris inference servers—a company spun out of the University of Oxford that performs part of AI’s calculations with light instead of relying solely on transistors—at its Scaling Inference Lab. It will be the lab’s first optical inference platform; the lab is backed by £50 million from ARIA and aims to test alternative computing architectures inside a real commercial data center. Lumai claims its technology can cut power consumption by up to 90%, a figure the trial will need to confirm.

Lumai’s optical computing in 30 seconds

  • ARIA will deploy Lumai’s Iris servers at its Scaling Inference Lab, operated by the nonprofit CommonAI.
  • The lab is backed by £50 million in UK public funding.
  • Iris handles the prefill phase of inference with light and leaves decoding to silicon.
  • Lumai claims up to 90% lower power consumption and 16x greater efficiency per inference—figures not yet independently verified.

Lumai announced the deal on November 3, 2026, from Oxford. ARIA is a UK public body created by law to fund high-risk scientific and technological projects. The Scaling Inference Lab, run by CommonAI CIC, aims to give startups and research groups an open testbed in a market dominated by large, closed systems, complete with standard power, cooling, and rack integration.

How the Iris server computes with light

According to Lumai, Iris encodes numbers as light intensity and solves matrix operations—the foundation of language models—using passive optical components such as lenses, in three-dimensional space. The company says this lets it perform millions of matrix operations in a single optical cycle, with power consumption barely increasing as performance scales up.

The company doesn’t position it as a GPU replacement. Iris handles prefill, the compute-intensive phase in which the model processes the input request, while conventional silicon handles decoding, the token-by-token generation of the response. The lab will test a multi-node, rack-scale system running production-sized large language models.

The figures come from Lumai: up to 90% lower data-center power consumption, speeds roughly 50 times higher, around a tenth of the total cost of ownership of a GPU-based system, and at least 16 times greater energy efficiency per inference than a silicon-only system. The announcement doesn’t specify exactly what configuration these numbers are measured against, nor does it include independent testing—which is precisely what the lab is meant to provide. It isn’t the only project of its kind in Europe: IONOS and Q.ANT recently opened a pilot to test photonic computing in their cloud.

Energy and sovereignty: the UK context

The underlying argument is energy. “Power, not demand, now limits how fast AI can be deployed,” says Xianxin Guo, co-founder and CEO of Lumai, who argues that multi-node testing in a commercial data center is where new architectures prove whether they’re ready. David Lewis, ARIA’s Chief Science and Technology Officer, says Lumai’s platform is “the most mature optical compute platform we evaluated.” Gavin Ferris, CEO of CommonAI, ties it to the UK’s sovereign capacity: “As the UK builds out its sovereign AI compute requirements, it has to do so within the power it has.”

The announcement doesn’t specify which data center will host the deployment, how many servers will be involved, the timeline, or the value of the deal. It also doesn’t mention Lumai’s funding rounds. It does note that the company spun out of optics research at Oxford in 2021, won the Falling Walls science breakthrough of the year award in 2025, was part of Intel Ignite’s first London cohort, and launched Iris Nova, its first product available for evaluation, in April 2026.

Frequently asked questions

What is optical computing?

Performing calculations with light instead of relying solely on electronic transistors. In Lumai’s case, AI’s matrix operations are solved using beams of light and lenses.

Does Lumai replace GPUs?

No. According to the company, Iris complements silicon: it handles the prefill phase of inference, while GPUs or other chips handle decoding.

Are the energy-savings figures verified?

No. The 90% savings figure and the 16x efficiency gain come from Lumai. The trial at ARIA’s Scaling Inference Lab should allow these to be independently evaluated.

Sources:

This article is also available in Spanish: “El laboratorio de inferencia de ARIA prueba la computación óptica de Lumai,” on revistacloud.com.

Scroll to Top