The new Nvidia Quantum-X and Spectrum-X silicon photonics switches with co-packaged opticsNvidiaBuilding any complex system is a sum of its parts, but the most foundational element is the thing that binds them together. In brick, its mortar; in wood, its nails; and in data centers, its networking. Last year, Nvidia CEO Jenson Huang declared that the data center is the new unit of compute. This means an entire data center should be thought of as a single system, just as a system-on-chip (SoC) or a server. Combining all those individual elements, including processing, memory, and storage, into a single platform requires a complex, high-performance, low-latency network.Disclosure: My company, Tirias Research, has consulted for Nvidia and other companies mentioned in this article.In 2024, Nvidia introduced the new Blackwell GB200 GPU AI accelerator to be combined with the Grace CPU in the new NVL72 rack server configuration. While the GPU and rack configuration were, and remain industry leading solutions, the real star of the story was the NVLink switch that allowed all 72 GPUs to be interconnected and to act as a single GPU, essentially a foundation for the server and rack. Because it was all in the same rack, it is referred to as a scale up solution. In 2025, Nvidia is looking to scale out the networking throughout the data center with new co-packaged optical networking solutions.According to Nvidia, it has worked with a number of industry partners to develop the Quantum-X InfiniBand and the Spectrum-X ethernet silicon photonics networking switches with co-packaged optics on the networking module. In addition, Nvidia is introducing three new liquid-cooled optical networking switches. Because of the high networking demands of AI data centers, especially AI factories, optical networking is already common. While it provides a significant boost in performance and reduction in latency, it comes at a cost in terms of power, space, complexity, and money. According to Nvidia, an AI factory can use up to 2.4 million optical transceivers using up to 24MW of power, which can be upwards of 10% of the overall data center power consumption.New photonic switch systems from Nvidia.NvidiaAdvantages of the new Nvidia photonic switch system over a traditional switch using optical ... [+] transceivers.NvidiaWhen comparing the Spectrum-X silicon photonics networking switches to traditional optical transceivers, Nvidia claims a 4x reduction in the number of lasers, a 3.5x increase in power efficiency, a 63x improvement in signal integrity, a 10x increase in network resiliency, and 1.3x decrease in network deployment time. These improvements are critical to meet the demands of agentic AI, which may require hundreds of times more resources than generative AI workloads spread across data center resources. The new Nvidia networking solutions will allow data centers to continue scaling with the demands of AI.While the announcement is a boost to Nvidia, it is a milestone for the industry. Co-packaged optic solutions have been under development since 2000 but have not entered mass production because of technical and manufacturing challenges, such as fiber coupling and light source integration. While the industry agrees that co-packaged optics are inevitable, the likelihood of a mass-produced solution was still viewed by many as several years away. However, with the assistance of its partners, Nvidia claims to have overcome the challenges and is ready for volume production starting later this year. The Nvidia partners include Browave, Coherent, Corning, Fabrinet, Foxconn, Lumentum, Senko, SPIL, Sumitomo Electric, TFC, and TSMC.In addition to the new networking solutions, Nvidia made a plethora of other announcements, including:Nvidia announced the next three generations of data center GPUs - Blackwell Ultra, Rubin, and Rubin UltraThe new DGX Spark (formerly Project Digits) and DGX Station systems for AI developersThe Blackwell DGX SuperPOD, an out-of-the-box AI factoryA number of new models/libraries and software solutions for developing and maximizing agentic AI solutionsFiber optic cables connected to the front of the Nvidia silicon photonic switch systemNvidiaEven with all these other announcements, networking was still the highlight of Jensen Huangs keynote and GTC for the second year in a row. Not only does it improve the performance efficiency of the data center, but it also advances a technology that is critical to the entire industry. While slated for rack-to-rack scale out connectivity today, it will likely be targeted for internal rack scale up networking in the future as the industry pushes the limits of copper interconnects.