Lumai's Optical AI Servers Head to UK ARIA Lab After Disrupt 2026 Win
Disrupt 2026 Showcase winner Lumai will put its Iris optical inference servers into ARIA's £50m Scaling Inference Lab to test claims of 90% lower power use.
Key points
- Lumai, an Oxford University spin-out founded in 2021, won the Enterprise Group 2 award for best Showcase Stage pitch at TechCrunch Disrupt 2026 in October.
- On 3 November 2026, Lumai said the UK's Advanced Research and Invention Agency (ARIA) will deploy its Iris optical inference servers in the Scaling Inference Lab.
- The Scaling Inference Lab is run by CommonAI CIC and backed by £50 million from ARIA, and Iris is the first optical inference platform it will host.
- Lumai claims Iris can cut data centre power use by up to 90% and lower total cost of ownership to about a tenth of GPU systems, figures the lab deployment is meant to test.
- Iris handles the prefill stage of AI inference with light, while conventional silicon chips handle decode, the step that generates each new token.
Lumai, the Oxford optical computing startup that won a Showcase Stage award at TechCrunch Disrupt 2026, is moving its technology out of the demo hall and into a working data centre. On 3 November 2026 the company said the UK’s Advanced Research and Invention Agency (ARIA) will deploy Lumai’s Iris inference servers in its Scaling Inference Lab, a testbed built to find out whether new kinds of AI hardware can ease the power crunch facing data centres.
What happened?
According to Lumai’s announcement, ARIA will add Iris servers to the Scaling Inference Lab, which is backed by £50 million from the agency and operated by the non-profit CommonAI CIC. Iris is the first optical inference platform to join the lab.
Data Center Knowledge reported on 6 November that the deployment will place a cluster of Iris Nova servers in the lab. Lumai told the publication the aim is to validate the system further and gather data that will shape its next product for production use. Neither the announcement nor the report gave a cluster size, a start date or any benchmark results.
ARIA’s chief science and technology officer, David Lewis, said in the release that Lumai’s system was the most mature optical compute platform the agency had looked at. Gavin Ferris, chief executive of CommonAI, linked the project to the UK’s push for sovereign AI computing, saying the country has to build that capacity within the electricity it already has.
Why it matters
Lumai’s pitch is that power, rather than chips or customer demand, is now the main limit on AI growth. “Power, not demand, now limits how fast AI can be deployed,” co-founder and CEO Dr Xianxin Guo said in the announcement.
The company makes three big claims for Iris:
- Up to 90% lower power use in the data centre than conventional systems.
- A total cost of ownership of roughly one-tenth of a GPU-based setup.
- At least 16 times better energy efficiency per inference than silicon-only systems, when Iris is paired with standard chips.
These are Lumai’s own figures. The point of the ARIA deployment is to check them on real workloads at rack scale, so they should be treated as claims until the lab publishes results.
How does Lumai’s optical computer work?
Most AI chips push electrons through silicon transistors. Lumai instead encodes numbers as the brightness of light and uses passive optical parts, such as lenses, to do the matrix multiplications at the heart of large language models. The company says this lets it complete millions of operations in a single optical cycle, working in three dimensions rather than across a flat chip.
The design is hybrid. Optics handle the heavy maths, while memory, control and other functions stay electronic. Lumai aims Iris at the prefill stage of inference, when a model reads and processes the whole prompt before it writes anything. The decode stage, which produces the answer one token at a time and depends more on memory speed, is left to GPUs or other chips.
Phillip Burr, Lumai’s director of product, told Data Center Knowledge that long prompts and agentic AI workloads, where models repeatedly process large amounts of context, make prefill an ever larger share of the work. In a September release, Lumai said its own tests showed about 10 times more compute per watt than GPU equivalents on prefill tasks.
Other practical details from the coverage:
- Iris works with PyTorch and with the vLLM and SGLang serving frameworks.
- The servers run in standard air-cooled racks and do not need liquid cooling.
- Lumai says training AI models is not a good fit for optical computing today.
The road from Disrupt
Lumai was part of the Startup Battlefield 200 at Disrupt 2026, held from 13 to 15 October at Moscone West in San Francisco. TechCrunch named it the best Showcase Stage pitch in the Enterprise Group 2 category, one of six industry-group awards alongside Uhura Bionics, Krokos Bio, Rein Security, Lola Vision Systems and Maka. We covered all six in our report on the Disrupt 2026 Showcase Stage winners.
The company spun out of University of Oxford research in 2021. In April 2026 it launched Iris Nova, the first of three planned Iris servers alongside Aura and Tetra, and said Nova could run Meta’s Llama 8B and 70B models in real time. Nova has been available for evaluation by cloud providers, enterprises and research groups since then. ARIA had already backed Lumai’s work before this deployment.
The ARIA news is one of the clearest follow-ups so far from this year’s Battlefield cohort. The overall winner, gut-organoid company Intero Biosystems, took home the Disrupt Cup and $100,000 in equity-free funding.
What happens next?
The lab now has to show whether optical prefill delivers the savings Lumai promises once it runs alongside conventional hardware in a commercial data centre. Lumai says the data will feed into its next production product. No timeline for published results has been announced.
For founders hoping to follow the same path, TechCrunch has listed Disrupt 2027 in San Francisco but has not yet confirmed dates. We will add it to the PFM Today tech events calendar as soon as it does.
Frequently asked questions
What is Lumai?+
Lumai is an optical computing company spun out of the University of Oxford in 2021 and led by co-founder and CEO Dr Xianxin Guo. It builds Iris, a family of AI inference servers that use light to perform the matrix maths behind large language models.
What is ARIA's Scaling Inference Lab?+
It is a testbed backed by £50 million from the UK's Advanced Research and Invention Agency and run by the non-profit CommonAI CIC. It evaluates new AI hardware on real workloads in a commercial data centre setting, and Lumai's Iris is the first optical inference platform to join it.
Did Lumai win an award at TechCrunch Disrupt 2026?+
Yes. TechCrunch named Lumai the best Showcase Stage pitch in the Enterprise Group 2 category of the Startup Battlefield 200 at Disrupt 2026, which ran from 13 to 15 October in San Francisco.
Sources
- Lumai – ARIA's Scaling Inference Lab deploys Lumai's optical compute technology to help power AI era (3 November 2026)
- Data Center Knowledge – Can Optical AI Cut Data Center Inference Power Use? (6 November 2026)
- TechCrunch – TechCrunch Disrupt 2026 Startup Battlefield 200: Celebrating outstanding achievements
- Lumai – The Next AI Infrastructure Challenge Is Before the First Token (10 September 2026)
- IT Brief UK – Lumai launches optical Iris Nova server for AI inference
- Lumai
- ARIA
- Optical computing
- AI inference
- Disrupt 2026
PFM Today Newsroom · News Desk
The PFM Today Newsroom tracks developments across startups, AI, design, business and more. Stories are researched from primary sources and reputable outlets, written with AI assistance, and reviewed against our editorial policy before publishing. Read our editorial policy.