NVIDIA’s $6.5 billion investment spree into photonics since March 2024 is not merely a diversification play—it is a strategic retreat from the physical limits of electrical interconnects. As AI models double their computational appetite every 18 months, copper-based data pathways have hit a wall in bandwidth density and energy efficiency. The solution? Light. By backing Ayar Labs, Coherent, and Corning, NVIDIA is orchestrating an end-to-end optical infrastructure stack—from chiplet-level I/O to data center cabling—aimed at keeping Moore’s Law on life support.
Ayar Labs sits at the heart of this transformation. Its TeraPHY optical I/O chiplets, co-developed with Marvell and integrated into NVIDIA’s roadmap, deliver 200 Gbps per lane with one-third the power of conventional SerDes. Crucially, Ayar’s technology enables direct optical links between GPUs without external switches, bypassing PCIe and NVLink bottlenecks entirely. I judge that by 2027, over 40% of high-end AI training clusters will adopt some form of on-package or board-level optical interconnect—and Ayar is positioned to become the de facto standard setter for this new architecture.
Yet silicon photonics alone cannot carry the load. Generating, modulating, and receiving light demands precision lasers and ultra-low-loss waveguides—domains where Coherent and Corning dominate. Following its merger with II-VI, Coherent now controls a commanding share of indium phosphide (InP) laser production, the backbone of high-speed optical transceivers. NVIDIA’s stake ensures supply chain resilience for these critical components. Meanwhile, Corning leverages its mastery in specialty fiber optics: its newly launched TXF™ Extreme fiber reduces attenuation by 30% compared to standard single-mode fiber and supports mode-division multiplexing (MDM), perfectly aligning with the dense, short-reach topology of NVIDIA’s GB200 NVL72 systems.
This shift reflects a deeper paradigm change in AI infrastructure. The old data center triad—CPU, DRAM, Ethernet—is giving way to a new trinity: GPU, HBM, and optical interconnects. In thousand-GPU-scale deployments, electrical latency and power draw are no longer tolerable. Internal NVIDIA estimates suggest that if Blackwell Superchips relied solely on copper interconnects, signaling alone would consume over 40% of total system power. Optical links can slash that to under 15%. The implications extend beyond efficiency: a 100MW AI data center saving 25% on interconnect power frees up 250 million kWh annually—enough to power a mid-sized city’s street lighting.
Significant hurdles remain. Silicon photonics lacks mature high-volume manufacturing yields. Coherent’s lasers cost three to five times more than copper alternatives. Corning’s advanced fibers require new installation protocols and tooling. Perhaps most critically, the ecosystem remains fragmented: Ayar champions UCIe-compatible optical interfaces, while rivals like Lumentum push alternative co-packaged optics (CPO) implementations. NVIDIA’s massive capital injection is effectively forcing convergence before standards solidify—locking in architectural control early.
There’s a darker implication: this optical revolution may deepen the concentration of AI infrastructure power. Only hyperscalers like NVIDIA, Microsoft, or Meta can afford the upfront costs of deploying next-gen optical networks. Smaller cloud providers risk being priced out of the AI race altogether. Optical interconnects are fast becoming a gatekeeper technology—not an option, but a prerequisite.
As the chip war shifts from transistors to photons, victory may hinge not on who designs the fastest GPU, but on who controls every nanometer of the light path between chips. The alliance of Ayar, Coherent, and Corning is laying that photonic track. The critical question is whether this track leads to an open ecosystem—or another walled garden ruled by a few.