The most expensive GPU in an AI data centre is surprisingly unproductive when it is waiting for data. As models and clusters grow, the network connecting processors becomes as important as the processors themselves. Meta and Oracle are responding by adopting NVIDIA’s Spectrum-X Ethernet platform for their next generation of AI infrastructure.
Oracle intends to combine Spectrum-X with its Vera Rubin architecture to connect millions of GPUs in large AI factories. Meta is integrating Spectrum Ethernet switches into the Facebook Open Switching System, or FBOSS, its internal platform for managing networks at scale. Both moves place an open, flexible networking layer at the centre of immense compute systems.
NVIDIA chief executive Jensen Huang calls Spectrum-X the nervous system for “giga-scale AI factories.” The metaphor is useful: adding more computational muscle will not help if information cannot move through the body quickly and predictably.
AI Workloads Expose Ethernet’s Weaknesses
Conventional Ethernet can lose effective capacity when data flows collide and congestion forms around particular links. NVIDIA says standard implementations typically achieve about 60 percent throughput, while Spectrum-X can reach up to 95 percent effective bandwidth. Adaptive routing and telemetry-based congestion control steer traffic away from hotspots and help several workloads share the infrastructure without disrupting one another.
That predictability matters in distributed training, where many GPUs work on portions of the same model and frequently exchange results. A delayed connection can leave expensive processors idle. Faster, steadier networking improves training and inference performance while increasing the return on the underlying GPU investment.
Spectrum-X combines switches and SuperNICs and supports multiple network operating systems. Meta will retain FBOSS, while alternatives include Cumulus, SONiC, and Cisco’s system through partnerships. For hyperscalers, that openness offers a way to introduce AI-specific performance without surrendering the operational environment built around existing networks.
One AI System Can Span Multiple Sites
NVIDIA’s modular MGX design lets partners combine CPUs, GPUs, storage, and networking in different configurations while retaining a common architecture across hardware generations. Inside a rack, NVLink provides scale-up connectivity; Spectrum-X enables scale-out growth across racks. Multiple facilities can then be connected through dark fibre or additional MGX-based switches.
NVIDIA distinguishes between Spectrum-X for networking inside a data centre and XGS for communication between sites. The hardware foundation is shared, but algorithms are tuned for different distances. The objective is to make geographically separated facilities behave more like one large AI supercomputer.
This approach aligns with Meta’s need to train increasingly large models while delivering services to billions of users. Oracle likewise expects Spectrum-X and Vera Rubin to accelerate the training and deployment of customer models. NVIDIA says Vera Rubin should become commercially available in the second half of 2026, with Rubin CPX expected by the end of that year.
Power Must Scale With the Network
Moving data efficiently is only part of the challenge. NVIDIA is working “from chip to grid” as rack density and overall power demand climb. A planned shift to 800-volt direct-current delivery is intended to reduce heat loss, while power-smoothing technology could cut peak requirements by as much as 30 percent. That would allow more compute within the same power footprint and soften sudden demands on the electrical grid.
The company is collaborating across the supply chain: Onsemi and Infineon on components; Delta, Flex, and Lite-On at rack level; and Schneider Electric and Siemens on data-centre designs. Software co-design adds another efficiency layer through FP4 kernels, Dynamo, TensorRT-LLM, and techniques such as speculative decoding.
Meta and Oracle’s adoption shows that AI infrastructure is becoming a systems problem rather than a shopping list of accelerators. GPUs, network algorithms, operating software, racks, cooling, and grid connections must behave as one machine. In the trillion-parameter era, the competitive advantage may belong not simply to whoever owns the most chips, but to whoever keeps those chips supplied with data and productive.
