Merging the mantras “the network is the computer” and “the data center is the computer,” Nvidia concludes that the network is the data center: A data center’s target workload determines the network it will employ. Whereas other companies favor a particular interconnect, Nvidia offers the Ethernet-based Spectrum X, NVLink, and InfiniBand. Spectrum X is Nvidia’s name for a combination of its Spectrum 4 Ethernet switch and BlueField 3 data-processing unit (DPU). Their support for RDMA over Converged Ethernet (RoCE) improves AI-cluster performance. Although standardized, InfiniBand is a low-latency networking technology available only from Nvidia, which the company supports with BlueField 3 and the Quantum 2 switch. NVLink is the company’s proprietary interconnect that supports chip-to-chip and network modes. A multilane serdes-based technology running a lightweight protocol, it resembles PCIe and provides high-bandwidth short-reach connectivity. Nvidia positions the Ethernet-based Spectrum X approach for cloud-based AI systems where multiple tenants employ AI accelerators like Nvidia’s GPUs to execute models of various sizes, including large language models (LLMs) and generative AI. By sticking with Ethernet, cloud service providers can take advantage of the popular protocol’s management-software ecosystem. For what the company calls AI factories, large clusters typically hosting only a few users (often just one) training the largest models, the company offers the combination of NVLink and InfiniBand. NVLink connects chips on the same board or in nearby boxes, and InfiniBand provides broader network connectivity. InfiniBand could also be used as an alternative to Ethernet for cloud services provided but only by well-resourced providers with their own management software.