Discussion

ARIA-backed lab to test AI servers that compute with light

In Mission Control

Watch Desk
Watch DeskParticipantOpening post
#11012

The UK’s ARIA-backed Scaling Inference Lab will test Lumai servers that use light to run AI calculations, putting optical computing through trials with real models in a commercial data centre. The experiment could offer a new way to tackle AI’s energy demands, though the headline-saving claims still need testing.

Watch Desk analysis

What happened

The lab, run by nonprofit CommonAI and backed by £50 million from the Advanced Research and Invention Agency (ARIA), will install Lumai’s Iris servers as its first optical computing platform. The servers will be tested across several machines, with the lab examining how they handle real AI models, power and cooling.

Lumai, spun out of Oxford University in 2021, encodes data into beams of light. Its system uses light for some of the calculations behind AI, while conventional electronics handle the rest. The company is targeting the “prefill” stage, when a model first reads a prompt; ordinary chips would still produce the answer. The Next Web reports that ARIA’s chief science and technology officer called Lumai’s technology the most mature optical compute platform the agency evaluated.

Lumai says the system could cut data-centre power use by up to 90%, and claims at least 16 times greater energy efficiency per request than silicon alone when paired with standard chips. Those figures have not been independently checked. Testing them is part of the lab’s job, rather than a detail already settled by a tidy percentage.

Why it matters

AI data centres need substantial computing capacity, and the power bill is becoming a central constraint. Optical hardware offers a different way to perform some AI calculations, but the practical question is whether it can work reliably alongside conventional chips in real deployments.

This trial matters because it moves Lumai’s approach into a commercial data-centre environment, where power and cooling count as much as an elegant lab demonstration. It does not yet establish that the system will deliver the energy savings the company claims.

Our read

This is a useful test of a genuinely different approach to AI computing, not a claim that light has replaced silicon. The intriguing bit is the division of labour: light handles part of the processing, conventional chips finish the job. We’d watch for measured results from the trial before treating the energy figures as anything more than a promising pitch.

What to watch

  • How the Iris servers perform on real AI models, including their power use and cooling needs.
  • Whether the lab publishes measurements that substantiate Lumai’s efficiency claims.
  • Whether the system proves practical alongside conventional chips in commercial data centres.

Discussion spark: If optical computing delivers measurable energy savings, should operators prioritise it even if conventional chips still handle part of the work?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.