← Back to feed

DriveNets has deployed what it calls the industry’s first commercial AI supercluster spanning two geographically separated data centers — not in a lab, not in a proof of concept, but in production, for a paying customer.

The customer is WhiteFiber, an AI infrastructure provider running Project Redwood. Two WhiteFiber facilities sit 52 miles apart. DriveNets’ AI Fabric connects them as a single logical GPU cluster at 111.2 Tbps with 0.9ms guaranteed latency.

Why physical distance matters

GPU clusters are typically designed for colocation: every GPU in the same building, connected by InfiniBand or NVLink over a few meters. The physics of AI training demand tight coupling — collective operations like AllReduce pass billions of parameters across every GPU simultaneously, and any added latency compounds into throughput loss.

The assumption built into most AI infrastructure is that you need everything in one room. That assumption shapes how data centers are sited, how power contracts are structured, and how much redundancy is physically possible.

Project Redwood breaks that assumption in production. At 52 miles, the round-trip latency budget would normally be prohibitive. DriveNets managed it by designing around AI workload traffic patterns rather than general-purpose datacenter networking: the AI Fabric is optimized for the collective communication patterns that dominate training runs, not for general TCP/IP packet forwarding.

Numbers

• Interconnect: DriveNets AI Fabric
• Bandwidth: 111.2 Tbit/s validated at production scale
• Latency: 0.9ms guaranteed round-trip
• Distance: 52 miles (approximately 84 km)
• Deployment: Live, commercially operational

What this enables

Geographic separation changes the reliability calculus for large training runs. A single-site GPU cluster goes down if that site loses power, cooling, or network. A multi-site supercluster survives the failure of one location if the interconnect stays up.

It also opens siting options that single-facility clusters don’t have. Two smaller, cheaper sites in different grid zones can replace one large site in a constrained location. Power availability, land costs, cooling water, and permit timelines vary dramatically across even short distances. Splitting a cluster across 52 miles lets operators arbitrage those variables.

The state of scale-across

Scale-across AI networking — linking clusters across physical sites — has been discussed at the research level for years. The major hyperscalers have done internal experiments. What DriveNets and WhiteFiber have done is move from internal experiment to commercial production, which is a different category of claim. Production deployment means the architecture had to survive real workloads, real traffic patterns, and real failure modes.

The 111.2 Tbps figure is not a theoretical peak; it’s a validated measurement on live infrastructure. That’s the number customers pay against.

Whether this generalises to the largest training clusters — which may require petabit-scale interconnects, not terabit — is an open question. But as a proof point that commercial AI superclusters can operate across tens of miles, Project Redwood is now the reference deployment.