While discussions about energy consumption and scalability in artificial intelligence infrastructure are growing, Lumai announced the Lumai Iris system, which can infer a large language model with billions of parameters in real time. The company announced that Lumai Iris Nova, the first server in the Iris family, has been opened to hyperscaler, neo-cloud companies, enterprise customers and research institutions for the evaluation process.
Lumai Iris uses a light-powered optical computing approach instead of traditional silicon-based processing architectures. According to the company’s statement, the system offers higher speed in inference workloads, higher execution efficiency and up to 90 percent lower energy consumption compared to traditional GPU architectures. Light-powered AI server from Lumai The Lumai Iris family consists of servers called Nova, Aura and Tetra.
Iris Nova, the first member of the family, is available for evaluation as of today. The system focuses on the new era of artificial intelligence, where the emphasis is on running large language models in the real world rather than training them. As artificial intelligence services reach more users, the inference load increases rapidly and data centers face tougher limits on power consumption, cooling and scaling. Futuristic Circuit Board Render With Bokeh Effects – Technology Related ConceptLumai introduces optical computing architecture at this point.
According to the company, Iris servers use photons instead of the traditional silicon-based processing approach to accelerate inference workloads. This structure carries out the basic mathematical operations of artificial intelligence models more efficiently and gives a higher result in terms of performance per watt, especially in data centers. The company states that the artificial intelligence industry is now shaped not only by model training but also by running models on a large scale.
Therefore, inference capacity, latency, energy consumption and cost have now become more critical in data center planning. According to data from the International Energy Agency, global data center power demand will double by 2030. Lumai positions the Iris family as a new process approach against exactly this energy wall. In traditional silicon architectures, each new generation comes with more power, more cost and more demanding thermal requirements.
According to Lumai, optical computing opens a different way to overcome this limitation. The company’s CEO and co-founder, Dr. Xianxin Guo said that while the industry is moving into the inference era, it is also on the verge of the post-silicon era. Guo stated that by moving the computational paradigm from electrons to photons, an order of magnitude increase in performance and significant energy savings could be achieved.
Lumai’s technology emerged from optical research conducted at Oxford University. The company has developed an architecture that uses light within a three-dimensional volume to counter the two-dimensional limitations of traditional chips. This approach enables large-scale spatial parallelism. Thus, millions of transactions can be executed simultaneously. Lumai states that this provides low-cost and high token throughput inference, especially in computation-intensive workloads.
Lumai Iris Nova runs real-time inference on Llama 8B and Llama 70B models. The server uses a hybrid processor architecture. In this structure, digital processing units manage the system control and software side, while the optical tensor engine performs basic mathematical operations. Lumai states that thanks to this hybrid approach, the system can be integrated more smoothly into existing data center infrastructures.
The company’s technology also stands out in the prefill phase of decoupled inference architectures. The prefill process is known as the stage where the model processes input tokens and requires high computational power, especially in long contexts. Lumai states that its optical transaction structure can process tokens at a high scale and with high efficiency at this stage. The British government-supported Advanced Research and Invention Agency also drew attention to Lumai’s optical transaction approach.
Suraj Bramhavar, ARIA Program Director, said the demand on existing AI processors necessitates an urgent search for alternative means of scaling. Bramhavar noted that Lumai is among the companies showing that optical processors could be one of those avenues, and ARIA is pleased to work with the company to explore moving beyond the traditional digital computing paradigm. Lumai Iris Nova is currently available for evaluation.
The next systems in the Iris family will extend to larger hyperscale and enterprise deployments with higher performance and greater efficiency. After Nova, the company will expand the Iris family with Aura and Tetra models. Lumai was founded in 2021 as a company born from the optical research of Oxford University. The company focuses on optical computing technologies that reimagine AI infrastructure for the inference era.
Lumai is developing systems that promise faster inference, higher execution efficiency and up to 90 percent lower energy consumption compared to traditional GPU architecture. The company also received the Falling Walls Award for Science Breakthrough of the Year 2025. He took part in the first cohort of Intel Ignite in London and won the “Best Overall Technology” award at the OCP Future Technologies Symposium. Lumai CEO Dr.
Xianxin Guo is among the graduates of the Royal Academy of Engineering’s Shott Accelerator program. The company’s CTO, Dr. James Spall was included in the 2025 Photonics 100 list.
Comments
You can write your views about this story. Comments may be moderated according to site settings.