Nvidia launches Spectrum-XGS to build global AI factories
By introducing ‘scale-across’ networking, Nvidia’s Spectrum-XGS nearly doubles multi-node communication performance, creating giga-scale AI factories optimised for hyperscale infrastructure.
American technology company Nvidia has unveiled Spectrum-XGS Ethernet, a new networking technology designed to connect multiple data centres into unified giga-scale AI factories.
With AI demand skyrocketing, single facilities are hitting limits in power and capacity, creating the need for infrastructure that can operate across cities, nations and continents.
Spectrum-XGS extends Nvidia’s Spectrum-X Ethernet platform, introducing what the company calls a ‘scale-across’ approach, alongside scale-up and scale-out models.
Integrating advanced congestion control, latency management, and telemetry nearly doubles the performance of the Nvidia Collective Communications Library, allowing geographically distributed data centres to function as one large AI cluster.
Early adopters like CoreWeave are preparing to link their facilities using the new system. According to Nvidia, the technology offers 1.6 times greater bandwidth density than traditional Ethernet and features Spectrum-X switches and ConnectX-8 SuperNICs, optimised for hyperscale AI operations.
The company argues that the approach will define the next phase of AI infrastructure, enabling super-factories to manage millions of GPUs while improving efficiency and lowering operational costs.
Nvidia CEO Jensen Huang described the development as part of the AI industrial revolution, highlighting that Spectrum-XGS can unify data centres into global networks that act as vast, giga-scale AI super-factories.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
