A Different Answer to the Power Problem
Every hyperscaler capacity conversation in 2026 eventually arrives at the same wall: power. Lumai, a company spun out of the University of Oxford in 2021, is pitching an answer that sidesteps the wall rather than climbing it. Its Iris Nova system is an optical AI accelerator built specifically for data center inference, and it has just gone live at the Scaling Inference Lab in the UK, a facility built to test new AI hardware against real commercial workloads rather than lab benchmarks.
Lumai director of product Phillip Burr frames the motivation in blunt economic terms: "Power is the biggest challenge." His follow-up is the thesis the whole product rests on: "Data centers have a fixed power budget, and inference demand is growing faster than the available power." That framing matters because it treats power as the hard constraint on inference capacity, ahead of raw compute availability, which lines up with what GlobalData and IDC have separately found about grid access becoming the binding bottleneck across the industry.
How Optical Inference Actually Works
Iris Nova is a hybrid design, not a full replacement for electronic computing. It encodes numbers as light intensity and uses passive optics, lenses in effect, to perform matrix multiplication, the core math behind transformer models. Everything else, memory, control logic, and the nonlinear functions that sit between matrix operations, stays electronic. Burr says each optical cycle can complete what he calls millions of operations using the same underlying math a transistor-based GPU would otherwise perform.
Critically, Lumai is not claiming optical compute can do everything a GPU does. The company is explicit that training is a poor fit for the technology today, since training requires higher numerical precision and continuous weight updates that optical hardware is not built for. The pitch is narrower and, because of that, more credible: optical compute for one specific, well-defined slice of the inference pipeline, applied where its physics genuinely has an advantage rather than stretched to cover every workload a data center runs. That restraint is itself a signal worth noting, since plenty of exotic-hardware pitches over the past decade have failed precisely because they overpromised on generality.
Why Lumai Only Wants Half the Job
That slice is prefill, the phase of inference where a model processes an incoming prompt and its context before generating any output. Prefill is compute-bound, meaning raw arithmetic throughput determines how fast it runs, which is exactly where Burr says optical hardware has an advantage. Decode, the phase that generates output tokens one at a time, is limited by memory bandwidth instead, so faster matrix math barely moves the needle there. Lumai leaves decode on GPUs entirely.
The operational model is disaggregated serving: prefill and decode run on separate hardware, coordinated through standard tooling. Lumai says Iris works with PyTorch and integrates with the vLLM and SGLang serving frameworks already common in production inference stacks. Burr's pitch to infrastructure buyers is direct: "Moving prefill off GPUs also frees that GPU capacity for decode," turning a hardware swap into a capacity multiplier for the GPUs an operator already owns.
No Liquid Cooling Required
One detail stands out against the rest of this issue's power and cooling theme: Iris Nova runs in standard air-cooled racks. Burr states it plainly: "Iris Nova operates in standard air-cooled racks and does not need liquid cooling." That is a meaningful operational claim if it holds up, since most new high-density AI hardware, from GB300 racks to the modular liquid-cooled units competitors are shipping, assumes direct liquid cooling as a baseline requirement, not an option.
If optical inference hardware can deliver meaningful throughput without the plumbing, water sourcing, and retrofit cost that liquid cooling demands, it changes the calculus for operators retrofitting older air-cooled facilities rather than building new liquid-ready sites from scratch. That is speculative until independent data exists, but it is the kind of claim worth tracking precisely because it cuts against where the rest of the industry is headed. A large installed base of air-cooled legacy facilities sitting idle for lack of a liquid-ready upgrade path would suddenly look like usable capacity again, which is a meaningfully different story than another vendor simply promising faster chips.
The Deployment Is the Disclaimer
The honest caveat here is that this is a vendor-sourced story at an early stage. Lumai has published no benchmark, throughput, or power consumption figures beyond its own claims, and the Scaling Inference Lab deployment exists specifically to generate that validation data, which is still pending. The lab is operated by the non-profit CommonAI CIC and backed by ARIA, the UK's government-funded Advanced Research and Invention Agency, which underwrites exactly this kind of unproven-but-promising hardware bet.
That is not a criticism of the technology so much as a statement of where it sits in its lifecycle. Optical computing for AI has been a research topic for years without a commercial deployment at this stage of maturity. Iris Nova getting onto live commercial workloads, even in a test lab, is a genuine step forward. It is not yet evidence that the economics work at the scale a hyperscaler or large enterprise would need.
What Enterprise Infrastructure Teams Should Take From This
For now, Iris Nova is best treated as an early signal worth tracking rather than a line item in next year's infrastructure budget. Enterprise teams evaluating inference infrastructure spend should watch the Scaling Inference Lab results when they publish, specifically looking for independently measured power-per-token figures on prefill workloads compared against current GPU baselines under comparable load. Until that data exists, any power-savings claim coming from Lumai should be treated as a vendor projection to monitor, not an input to plug into a real capacity or budget model this year.
The broader value of this story is what it says about where hardware innovation is heading. With grid delays stretching data center delivery timelines by roughly two years and gas turbines sold out through 2031, approaches that reduce power draw per unit of compute are getting serious attention from labs and funders that would have ignored them three years ago. Enterprise buyers should expect more optical, photonic, and exotic compute pitches over the next 12 months, and should build a simple evaluation checklist now, covering independently verified throughput, power draw, and cooling requirements, so the eventual winners are easy to separate from the noise once the funding announcements start arriving faster than the proof points do.
