Dipole Labs boosts computing capacity by 25% in AI data centers through optical switching

Summary

Dipole Labs, founded by two Harvard postdoctoral researchers and participating in the YC Summer 2026 program, is developing AI-controlled optical circuit switches. This system aims to achieve up to 25% additional useful computing capacity from existing AI cluster hardware by keeping inter-rack data within the optical domain.

Why it matters

AI data centers are becoming increasingly network-dependent. Expensive GPUs sit idle while waiting for data to pass through electrical conversion layers. If hyperscale-level sub-microsecond optical reconfiguration can be achieved, idle silicon can be converted into productive computing capacity without the need for new chips. The company claims this gain equates to a value of approximately $1 billion annually at the 1 GW scale.

Highlights

• Optical domain switching with sub-microsecond reconfiguration eliminates the electrical round-trip path that adds latency, heat, and power consumption between racks.
• The AI-native control layer learns GPU communication patterns and synchronizes switch reconfiguration with training/inference processes, turning the network into an active accelerator.
• The founders bring a decade of research experience in photonics and quantum optics (Harvard, ETH Zurich, EPFL; publications in Nature and Nature Materials) to this field.

Critical take

The 25% increase mentioned in the headline is not a measured result from a deployed system but a projection at the design stage. The company is still in the prototype and design partnership phase. Therefore, the figure should be evaluated as a targeted value rather than a proven capability. The 'AI-native control layer,' which must predict and track the highly dynamic and non-stationary GPU communication patterns in real time, remains an open engineering problem. Additionally, no references to benchmarks or third-party validation are provided.