Key takeaways
- NVIDIA announced Spectrum-XGS Ethernet on August 22, 2025 as a scale-across addition to its Spectrum-X platform.
- NVIDIA says the technology uses distance-aware congestion control, precision latency management and end-to-end telemetry to connect distributed data centers.
- The company says Spectrum-XGS nearly doubles NCCL performance, but its announcement does not publish a test configuration, baseline or independent result.
- The underlying Spectrum-X platform comprises Spectrum-X switches and ConnectX-8 SuperNICs. NVIDIA does not say in the announcement that every existing Spectrum-X deployment can add XGS through software alone.
- CoreWeave will be among the first to connect its data centers with Spectrum-XGS; NVIDIA does not provide a deployment date or production benchmark.
Scale-out and scale-across are different jobs
Spectrum-X is NVIDIA’s Ethernet platform for AI infrastructure. The company describes Spectrum-XGS as an addition that extends that platform beyond scale-out, which links compute within a data-center fabric, to “scale-across” connections between separate facilities.
That distinction matters. A long-distance link introduces more delay than a local fabric, and NVIDIA is selling XGS as a way to manage that distance while presenting several sites as a larger computing environment. “Giga-scale AI super-factory” is NVIDIA’s term for that design goal, not an independently defined performance class.
What NVIDIA has disclosed
NVIDIA says the Spectrum-X platform comprises Spectrum-X switches and ConnectX-8 SuperNICs. For Spectrum-XGS, the announcement names algorithms that dynamically adapt the network to the distance between facilities, along with three features:
- auto-adjusted distance congestion control;
- precision latency management; and
- end-to-end telemetry.
The release does not provide a bill of materials for an XGS deployment. It also does not say that XGS is merely a software update for every existing Spectrum-X installation. Describing XGS as “fully integrated into” Spectrum-X establishes the product relationship, but not an automatic upgrade path.
How to read the performance claims
NVIDIA states that Spectrum-XGS nearly doubles the performance of the NVIDIA Collective Communications Library (NCCL) for multi-GPU and multi-node communication across geographically distributed clusters.
That is a vendor performance claim. The announcement does not disclose the number or type of accelerators, link distance, transport capacity, workload, comparison baseline or absolute NCCL throughput. It therefore cannot support a universal “2×” expectation for every multi-site deployment, or a direct claim that XGS doubles a particular Spectrum-X configuration.
NVIDIA separately says the broader Spectrum-X platform provides 1.6× greater bandwidth density than off-the-shelf Ethernet. The release does not establish that 1.6× as a measured XGS long-distance result, so the two figures should not be combined.
What CoreWeave has committed to
NVIDIA identifies CoreWeave as one of the first companies that will connect its data centers with Spectrum-XGS. CoreWeave’s quoted statement describes the intended result as a unified supercomputer.
Neither statement is a field report. NVIDIA’s release gives no site names, rollout date, completed deployment, workload result or production throughput. The accurate reading is that CoreWeave announced an adoption plan, not that a pilot or production network had already proved NVIDIA’s claims.
Spectrum-X versus the XGS addition
| Question | Spectrum-X platform | Spectrum-XGS addition |
|---|---|---|
| Primary scope in NVIDIA’s announcement | AI Ethernet scale-out | Scale-across between distributed data centers |
| Named hardware | Spectrum-X switches and ConnectX-8 SuperNICs | Integrated into Spectrum-X; no separate bill of materials disclosed |
| Named controls | Not detailed in this announcement | Distance congestion control, latency management and telemetry |
| Quantified claim | 1.6× bandwidth density versus off-the-shelf Ethernet | Nearly 2× NCCL performance; test details not disclosed |
| Public customer evidence | Existing platform | CoreWeave plan; no production benchmark in the announcement |
| Availability statement | Existing platform | “Available now” as part of Spectrum-X |
Bottom line
Spectrum-XGS is best understood as NVIDIA’s scale-across extension to Spectrum-X, not as a separate switch generation and not as a proven drop-in software upgrade. The announced controls directly target the problems created by distance, while the NCCL figure and the CoreWeave plan show NVIDIA’s intended performance and market direction. Buyers still need topology, distance, workload and production data before translating those claims into a deployment forecast.
Sources
This article was researched and fact-checked against the following sources:
- NVIDIA Introduces Spectrum-XGS Ethernet to Connect Distributed Data Centers Into Giga-Scale AI Super-Factories | NVIDIA Newsroom (nvidianews.nvidia.com)
- NVIDIA Corporation - NVIDIA Introduces Spectrum-XGS Ethernet to Connect Distributed Data Centers Into Giga-Scale AI Super-Factories (investor.nvidia.com)
- NVIDIA Launches Spectrum-XGS Ethernet to Link Distributed Data Centers - Converge Digest (convergedigest.com)
- HPCwire - Since 1987 – Covering the Fastest Computers in the World and the People Who Run Them (hpcwire.com)
- Nvidia's new Spectrum-XGS aims to turn multiple data centers into 'one gigantic GPU' - SDxCentral (sdxcentral.com)
