NVIDIA (NVDA), the global leader in AI computing, attempts to solidify its supremacy in the AI infrastructure market with its latest announcement of a $4 billion investment in photonics, the technology of using light instead of electricity to move data. Amid Meta’s announcement of four new AI chips, which challenge NVIDIA’s market dominance, the firm is moving upstream in the photonics supply chain by splitting the investment equally between Lumentum and Coherent, two leading photonics companies, to prep for the shipping of its next-generation Vera Rubin architecture in mid-2026.
This move is tailored to eliminate the critical networking bottlenecks currently faced by massive data centers, driven by the need to connect hundreds of thousands of GPUs for parallel processing. By integrating advanced optical interconnects directly into its silicon roadmap, NVIDIA attempts to break this throttling issue and push the ‘unit of compute’ from individual chips to entire rack-scale AI factories.
Bridging the Interconnect Wall
As AI models advance to the trillion-parameter scale, the traditional copper-based interconnects pose a bottleneck in moving data between chips, which is being bridged by CPO (Co-Packaged Optics). While this revolutionary technology allows for the direct integration of optical engines onto the switch ASIC substrate, the traditional pluggable optical transceivers are too power-hungry, making them unable to scale to meet the bandwidth needs of future AI data centers.
NVIDIA’s Spectrum Ethernet switches, which integrate 32 silicon photonic engines, can deliver up to 409.6 Tb/s of aggregate bandwidth. By placing optical engines (photonic chiplets) directly next to the switch silicon, CPO can eliminate this hurdle of traditional copper wiring, reduce power consumption and heat production, and increase data bandwidth density. Also, the new design uses fewer lasers, which is essential for massive AI platforms.
By joining hands with Lumentum and Coherent, NVIDIA is aiming for 5 times more power reduction and 10 times more network resilience, for energy savings and enhanced performance. This can be driven by advanced photonics, including Lumentum’s CW lasers and Indium Phosphide (InP) based photonic integrated circuits, combined with External Laser Source Pluggable (ELSP) modules. Coherent will be collaborating on silicon photonics engines that power the new transceiver form factor, enabling CPO to scale up to millions of GPUs.
Vera Rubin: The Birth of the Rack-Scale Organism
In the AI infrastructure, which is drastically altering the terrain of the tech world, NVIDIA’s goal is primarily to demolish the interconnect wall by solving it with silicon photonics. With this, it also aims to introduce the HBM4 memory, which can provide 22TB/s bandwidth, and the custom Vera CPU, which is made to replace the previous Grace architecture for faster reasoning.
The NVIDIA Vera Rubin NVL72, unveiled at CES 2026, is designed as a rack-scale system, where 72 Rubin GPUs and 36 Vera CPUs act as a single organism. With the HBM4 memory, the capacity per card increases to 288GB, enabling the full NVL72 rack to provide 20.7 TB of HBM4 capacity.
Named after the astronomer Vera Rubin, this platform focuses on accelerating agentic AI and deep reasoning tasks. Targeted at large-scale AI factories and industrial AI applications, this architecture represents a pivotal advancement in AI infrastructure.
The platform provides 5x higher inference performance and 3.5x faster training compared to the Blackwell architecture. It also introduces third-generation Confidential Computing, which secures data across CPU, GPU, and NVLink Domains.
From a market outlook perspective, while NVIDIA is aiming to enhance its infrastructure capabilities and reduce its reliance on external suppliers, these strides in the AI sector amid geopolitical tensions are raising concerns about geopolitical risk tax. Along with Lumentum and Coherent, NVIDIA has also made another massive investment in cloud partners Nebius and Coreweave, strengthening its coordination to ensure immediate deployment of the Vera Rubin infrastructure upon release.
