Photonic chips race toward an "energy cure" for AI — but can light really tame the power appetite?

Big-money bet and a clear angle

Nvidia (英伟达) reportedly poured $4 billion in early March to secure future capacity from two optical-communication giants, Lumentum and Coherent, locking their high-end photonic chip and module output for the next three years. The move highlights a strategic bet: as generative AI workloads explode and datacenter power bills soar, photonics — using photons instead of electrons to compute and interconnect — is emerging as the most talked-about route to break the AI “energy curse.” Can light replace heat-generating transistors at scale? Investors and chipmakers are starting to act as if the answer might be yes.

Research breakthroughs multiply

Academic papers and demonstrations from around the world are feeding the optimism. Researchers at the University of Sydney reported in Nature Communications (DOI: 10.1038/s41467-026-68648-1) a wafer-scale, ultra-compact photonic chip that performs neural-network operations with nanostructures only tens of micrometres across; reported experiments show high accuracy on biomedical-image tasks while consuming far less energy and latency than electronic accelerators. Separately, Xidian University (西安电子科技大学) published in Optica a two-chip photonic neuromorphic system that reportedly implements reinforcement learning purely in the optical domain — overcoming earlier hybrid designs that had to convert light to electricity for nonlinear steps. It has been reported that the EU-funded HAETAE project and other groups project up to an order-of-magnitude improvements in energy efficiency if fully optical processing can be commercialized.

From interconnects to full optical brains — supply chains and politics

The industry shift is not limited to compute. It has been reported that MicroLED and silicon-photonics interconnects can cut transmission energy to roughly 5% of equivalent copper solutions, pushing datacenter operators toward optical fabrics. That transition has commercial and geopolitical implications: Western chip OEMs are locking down component supply amid export controls, reshuffled supply chains and rising demand for advanced packaging. Nomura Securities has estimated advanced photonic chip capacity could grow more than 80% year‑on‑year in 2026 — a milestone year that many in the industry view as the tipping point for large-scale silicon-photonics deployment.

Caution: lab success ≠ instant revolution

Promises are large but scaling remains hard. Many high-profile demonstrations are single-task, lab-scale systems; manufacturing yield, integration with existing electronic ecosystems, and durable nonlinear optical elements are real engineering hurdles. It has been reported that while pure-photon computation can slash certain sources of energy loss, hybrid architectures and cooling, testing and control will still matter. So will regulatory and trade dynamics as nations race to secure photonics supply chains. The question now is not whether photonics can perform — research says it can — but whether industry can turn those flashes of light into a practical, global answer to AI’s runaway power bill.