NVIDIA Builds a Network for Data Centres with Hundreds of Thousands of GPUs

NVIDIA introduced the Spectrum-6 networking platform for data centres containing tens or hundreds of thousands of GPUs. In a system of this size, the network is not a secondary component. It determines whether individual machines operate as one system or whether expensive hardware waits for data transfers. Spectrum-6 belongs to the Spectrum-X family and uses Ethernet. NVIDIA is competing with more proprietary networking technologies while relying on a standard already familiar to enterprise data centres. AI workloads, however, require more precise traffic management than normal office applications. When a large model is trained, thousands of processors continuously exchange parts of the computation. A brief congestion event or the failure of one component can slow the entire workload. In agent systems, the network also connects compute servers, storage, databases and external tools. Latency accumulates at every step. NVIDIA says the platform is designed for “hundreds of thousands of GPUs” and for gigawatt-scale data centres. That figure shows the scale of planned investment. One gigawatt is comparable to the output of a major power source and far exceeds the consumption of a normal enterprise computing facility. The economics of such projects depend on more than chip prices. Investors must secure grid connections, cooling, backup power, land, water and long-term electricity contracts. Networking equipment also adds its own consumption. Even a small efficiency gain can therefore be financially significant in continuous operation. The vendor emphasises resilience and the ability to reroute traffic during failures. That matters because an outage in a large computing cluster can waste many hours of ongoing training. Companies should nevertheless demand concrete data on availability, repairability and actual capacity under partial failure. For European investors, permitting is crucial. A data centre of this size affects the transmission grid, local infrastructure and water availability. A project may therefore be delayed not by chip shortages but by grid capacity, planning procedures or local opposition. Supplier concentration also creates risk. A customer that buys processors, networking, software and management from one company gains easier integration. At the same time, the future cost of switching to a competitor increases. Ethernet alone does not guarantee that the entire system can be replaced easily. Regulators will monitor energy consumption and competition. A supplier with a dominant position in GPUs can use the networking layer to strengthen its control over the entire data centre. Customers should therefore address access to operating data, compatibility and service terms in their contracts. Spectrum-6 is not a normal product for most companies. It will nevertheless affect them indirectly through cloud pricing. If the new network increases utilisation of expensive processors and reduces outages, computing prices may fall. If it requires more capital and locks operations to one vendor, part of the cost will be passed into model prices.

NVIDIA Builds a Network for Data Centres with Hundreds of Thousands of GPUs

The full article is available to users registered at Hard Skills