ARIA to test Lumai’s optical AI servers in commercial data centre with £50M lab

What the partnership involves

ARIA’s Scaling Inference Lab will deploy Lumai’s Iris servers—the startup’s light-based AI hardware—in a commercial data centre for real-world testing. The lab, funded by £50M from ARIA and operated by CommonAI, will assess Lumai’s claims that its optical computing approach can reduce data centre power use by up to 90% and improve energy efficiency by 16 times per request compared to traditional silicon-based systems.

Lumai, spun out of Oxford University in 2021, focuses its technology on the 'prefill' stage of AI processing, where a model first interprets a user’s prompt. The company’s hardware works alongside standard chips to enhance performance while cutting energy consumption.

Why this testing matters

Lumai’s claims have yet to be independently verified, making ARIA’s testing a critical step for validating—or challenging—its potential. If proven, the energy savings could address a major bottleneck in AI infrastructure: the rising power demands of data centres. This is particularly relevant as competitors like Nvidia invest heavily in light-based connections within data centres, though their focus differs from Lumai’s full-stack optical computing approach.

The Scaling Inference Lab already hosts Oriole, another UK startup developing light-based AI chip connections. Fractile, a UK-based inference-chip startup, is also raising funding at a $6.5B valuation, reflecting broader innovation in AI hardware across the country.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

For founders and operators in AI infrastructure, ARIA’s test is a signal to watch closely. Lumai’s approach—if validated—could redefine expectations for energy efficiency in data centres, a growing pain point as AI workloads expand. While optical computing is not new, its application in AI inference remains unproven at scale. A successful trial could accelerate adoption of light-based systems, pressuring incumbents to explore alternatives to silicon-dependent architectures. Conversely, if the results fall short, it may reinforce the dominance of established players like Nvidia, which are already integrating optical elements into their solutions.

Related on Lazyfounder

• Nvidia
• Fractile
• More technology news