Photonics Is Becoming the Nervous System of AI Clusters

Optics is moving into system architecture

I would not reduce this development to another investment trend. By September 2026, photonics had become an answer to a physical constraint: electrical connections between accelerators and switches are starting to limit the scaling of AI clusters.

NVIDIA’s strategic partnership announcement with Lumentum cites a $2 billion investment. At the time, the funding was linked to research, future capacity expansion and work at a new U.S. factory. Reuters and CNBC also reported an NVIDIA investment of the same size in Coherent.

Another sign of the shift is Marvell’s $3.25 billion acquisition of Celestial AI. The key asset is not merely a collection of optical components, but photonic fabric technology—an effort to bring light-based data transfer deeper into the computing system. At the same time, Defiance launched PHOT, Europe’s first photonics UCITS ETF.

The technical direction is clear: the industry is moving from 400G and 800G networks toward 1.6T infrastructure. Optical links provide the required bandwidth over distances where electrical signals become harder to maintain because of loss, signal quality and power consumption. Co-packaged optics goes further by placing optics closer to the networking chip and shortening the electrical path.

Industry materials cite potential network energy savings of up to 40% with co-packaged optics. I see that as a benchmark rather than a universal outcome: the collected sources do not provide a comparable laboratory assessment of latency, cost and power across different architectures.

Compute is no longer the main constraint

The bottleneck is shifting from individual GPUs to data exchange between them. A fast accelerator offers limited value when collective operations are constrained by the network and every delivered bit requires more energy.

That is why interest in Lumentum, Coherent, Celestial AI, POET Technologies and Lightwave Logic makes sense. But investments and ETFs validate demand for the theme, not the readiness of any particular technology for large-scale deployment. Xanadu also works with photonics, although quantum computing is a different application class.

I would look first not at peak speed, but at cost per bit, thermal behavior, laser reliability, serviceability and compatibility with existing 400G and 800G networks. Integration will determine whether 1.6T becomes standard infrastructure or remains an expensive showcase.

Photonics is no longer an exotic edge technology in AI. The harder unanswered question is how quickly optics can move from strategic investments to predictable operation inside real-world clusters.

We previously covered how confidential compute is reshaping AI inference infrastructure, privacy, and operating costs. That shift complements the move toward optical interconnects as AI systems demand faster, more efficient hardware at scale.