Lumai Productizes Lens-Based Optical Computer for AI Inference
Key Takeaways
- ▸Lumai's optical computing system successfully runs billion-parameter AI models, the first commercial demonstration at scale
- ▸Iris Nova offers 50× GPU performance with 90% power reduction, directly addressing data center power constraints
- ▸Roadmap to cluster-scale deployment (Iris Tetra) by 2029 with 100 TOPS/W efficiency and 1 exaOPS capacity
Summary
British startup Lumai has productized its lens-based optical computer, marking the first successful demonstration of optical computing at scale for billion-parameter AI models. The system uses 1,024 laser light sources and an electronic display to perform matrix multiplication optically, claiming 50× the performance of today's GPUs with a 90% reduction in power consumption.
The Iris Nova inference server, containing a single first-generation optical engine, will be offered to hyperscale customers for evaluation. The system offloads 90% of Llama workload to the optical domain while a CPU handles non-linear operations and accuracy-sensitive parts of the algorithm. Lumai's approach uses industry-standard customized components rather than requiring new materials, addressing scalability concerns that have limited previous photonic computing solutions.
The company has announced an ambitious roadmap: Iris Nova deployments in test clusters by end of 2026, followed by Iris Aura (multi-engine rack systems), and Iris Tetra (cluster-scale deployment planned for 2029) capable of delivering 100 TOPS/W and 1 exaOPS within a 10kW power budget. This iteration speed reflects Lumai's commitment to getting systems into customer hands for real-world evaluation.
- Hybrid architecture with CPU + optical engine provides flexibility for accuracy-sensitive operations while maintaining 90% optical workload for Llama
Editorial Opinion
Lumai's achievement represents a significant milestone in optical computing's journey from theoretical promise to practical commercial deployment. By successfully running industry-standard models like Llama at scale with demonstrated power efficiency gains, the company has moved beyond proof-of-concept into critical validation phase. The aggressive roadmap to 1 exaOPS within 10kW by 2029 is bold, but achieving it would fundamentally reshape data center economics for AI. Whether real-world deployments will match lab demonstrations and whether power efficiency alone will drive adoption across the industry remains to be proven.
