As artificial intelligence models grow larger and more computationally demanding, even the biggest data centers are starting to hit their limits. NVIDIA believes it has found a solution: Spectrum-XGS Ethernet, a new networking technology designed to link multiple AI data centers into what it calls “giga-scale AI super-factories.”
Announced ahead of the Hot Chips 2025 conference, the technology addresses one of the biggest challenges facing the AI industry: how to scale computing power when one building is no longer enough.
The Challenge: Scaling Beyond a Single Facility
AI models now require such enormous processing power that a single facility often cannot meet the demand. Traditional data centers face hard limits on physical space, power capacity, and cooling systems. Until now, the only option was to build entirely new sites—a costly process complicated by the difficulty of networking multiple centers together.
Standard Ethernet infrastructure has been a bottleneck, plagued by latency, jitter, and inconsistent performance over long distances. These issues prevent complex AI workloads from being distributed effectively across sites.
NVIDIA’s Answer: Scale-Across Technology
NVIDIA’s Spectrum-XGS introduces what it calls a “scale-across” approach, complementing existing strategies of scaling up processors or scaling out within a single site.
Key innovations include:
- Distance-adaptive algorithms that optimize performance across long connections
- Advanced congestion control to avoid data bottlenecks
- Precision latency management for predictable performance
- End-to-end telemetry for real-time monitoring and optimization
According to NVIDIA, these improvements nearly double the performance of its Collective Communications Library, which coordinates data exchanges between GPUs and computing nodes.
Early Adoption: CoreWeave Steps In
Cloud infrastructure provider CoreWeave will be among the first to deploy Spectrum-XGS.
“With NVIDIA Spectrum-XGS, we can connect our data centers into a single, unified supercomputer, giving our customers access to giga-scale AI,” said cofounder and CTO Peter Salanki.
This rollout will provide a critical test of whether the technology delivers on its promises in real-world conditions.
Why It Matters for the AI Industry
NVIDIA has been steadily expanding its networking portfolio, including the original Spectrum-X platform and Quantum-X silicon photonics switches, signaling that it sees infrastructure as a core bottleneck in AI growth. CEO Jensen Huang described the new era as an “AI industrial revolution,” with giant-scale computing factories forming the backbone.
If successful, Spectrum-XGS could change how AI infrastructure is designed. Instead of building massive single facilities that strain power grids and real estate markets, companies could distribute operations across multiple smaller centers while maintaining supercomputer-level performance.
Considerations and Open Questions
Still, the technology faces hurdles. Physics sets limits on data transfer speeds, and distributed centers introduce challenges in data synchronization, fault tolerance, and regulatory compliance that go beyond networking.
NVIDIA says Spectrum-XGS is available now as part of its Spectrum-X platform, though pricing details remain undisclosed. Adoption will depend on whether it proves more cost-effective than simply building larger standalone sites.
The Bottom Line
If NVIDIA’s technology works as intended, it could unlock faster AI services, lower costs, and more powerful applications by allowing distributed facilities to function as one. But until deployments like CoreWeave’s demonstrate results, the industry will be watching closely to see whether Spectrum-XGS can turn the vision of giga-scale AI super-factories into reality.
