NVIDIA doubles down on photonics

NVIDIA is spending billions on a part of the AI stack that most people never see. Not the GPU. Not the server. The links between them. In March, the company said it would invest $2 billion in Lumentum and $2 billion in Coherent, two major photonics suppliers. On March 31, it added another $2 billion investment in Marvell as part of a broader NVLink Fusion partnership. The pattern is the story. NVIDIA has decided that optical networking is no longer a side component of AI infrastructure. It is now one of the main constraints on how far these systems can scale. That shift matters because AI data centers have changed shape. A few years ago, the hard part was getting enough compute into a rack. Now the hard part is keeping ever larger clusters fed with data while power use, heat, and signal loss stay under control. Copper traces and pluggable optics still work, but they become clumsy as clusters spread across more racks and more buildings. NVIDIA has been saying this for a while. At GTC in March 2025, it introduced Spectrum-X and Quantum-X photonics switches and said co-packaged optics could cut networking power sharply while making giant GPU fabrics more resilient. What looked like a product roadmap now looks like a capital plan. The Lumentum and Coherent deals show what NVIDIA thinks it needs. Both agreements, announced on March 2, were not just equity checks. They also included multiyear purchase commitments and future capacity rights for advanced laser and optical networking parts. NVIDIA is not merely funding research. It is trying to lock in supply. Lumentum said the money would support R&D, future capacity, operations, and a new U.S. manufacturing buildout. Coherent said its agreement would expand supply, deepen joint development, and advance U.S.-based manufacturing. When a company prepays for lasers, it is telling you that lasers are becoming strategic. That makes the Marvell deal easier to read. Marvell is not a pure photonics company, but it sits in the middle of the same bottleneck. Its role in the new partnership is to provide custom XPUs and NVLink Fusion-compatible scale-up networking, while NVIDIA provides the broader rack-scale system around them. The two companies also said they would collaborate on silicon photonics. So the Marvell investment is not separate from the optics push. It extends it. NVIDIA wants custom silicon inside its interconnect ecosystem, and it wants the optical layer underneath that ecosystem to mature on its timetable. The deeper point is that NVIDIA is trying to control the distance between chips. Its dominance began with accelerators. Then it expanded into networking with InfiniBand, Ethernet, NVLink, DPUs, and switches. Now it is moving into the physics of how signals leave one package and reach the next one without wasting too much power on the way. That is where photonics enters the picture. Light is not replacing compute. It is being recruited to keep compute from choking on its own scale. Investors noticed because this is how a bottleneck becomes visible. Marvell shares jumped after the March 31 announcement. The earlier Lumentum and Coherent deals also lifted optics names across the sector. Markets were reacting to the money, but the money was reacting to something more basic: AI systems are now large enough that networking components once treated as plumbing have become core infrastructure. That is why the most revealing line in these announcements was not about GPUs at all. It was about future capacity rights for laser components.