The AI supercluster doesn't fit in one building anymore
DriveNets and WhiteFiber say they deployed the first commercial AI supercluster across separate data centers, betting long distance bandwidth can hold up under real workloads.
DriveNets and WhiteFiber say they deployed the first commercial AI supercluster across separate data centers, betting long distance bandwidth can hold up under real workloads.
For years, AI builders had a single rule: keep the GPUs in one building. The internal links between servers in one data center move data roughly a thousand times faster than anything you can run between sites, and AI superclusters, the multi-thousand-GPU pools that train frontier models, depend on that bandwidth. The deployment DriveNets and WhiteFiber are announcing says it can retire that constraint. A commercially operated cluster whose GPUs live in separate data centers, stitched together by orchestration software, behaves as one logical supercluster. (DriveNets announcement, Telecompaper coverage)
WhiteFiber, a North American AI-infrastructure provider that builds and leases GPU capacity, is the customer of record. DriveNets, a networking-software vendor, supplies the orchestration layer that lets physically separated nodes act as one training and inference pool. DriveNets calls the design "scale-across"; the label is the vendor's framing, not an industry term, and "distributed cluster" reads cleaner. (TelecomTV coverage)
Today, superclusters sit in single buildings because spreading nodes across data centers collapses the bandwidth between GPUs and stalls training. Long-distance optical links carry orders of magnitude less bandwidth than the internal fabrics that bind a single-site cluster, and the gap shows up as GPUs waiting on each other. If orchestration software plus long-distance optical links can close that gap, siting decisions loosen. A builder could put part of the cluster near cheap power and part near users, fold in stranded GPU capacity in a different region without re-architecting, or expand past the power or cooling ceiling of any single building. Geography becomes a planning variable instead of a constraint.
The deployment is one engineering bet: that node-to-node bandwidth and latency across long-distance links can hold up under real AI training workloads. The press release and its trade re-reports do not disclose the link distance, the per-node bandwidth, the interconnect type, the workload mix, or whether the cluster is running production training or still in commissioning. Independent measurement under real training, not vendor benchmarks run on favorable traffic, is what would convert the announcement from a hypothesis in progress into a working design.
What to watch next. Three numbers matter: link distance, bandwidth per node under training traffic, and how the orchestration recovers when one site drops out or a long-distance link flaps. Until those land in independent hands, the deployment is a test rig with the wiring in place but the verdict still pending.