Solving AI Data Center Power Constraints With Scale-Across Networking
Running out of power at an AI data center no longer limits AI scaling. See how WhiteFiber Continuum™ links two sites into one GPU cluster.
Every AI infrastructure buildout eventually runs into the same wall: power. Not square footage, not chip supply, power. Once a data center's power envelope is exhausted, adding more GPUs and CPUs to serve a growing workload simply isn't possible at that site, no matter the budget behind it.
According to DriveNets, which provides the network fabric for WhiteFiber Continuum™, that wall doesn't have to be the end of the road.
Why AI Data Center Power Constraints Become a Hard Stop
A data center's power budget is fixed by its facility, its utility connection, and its cooling capacity. When that budget is spent, it's spent. Teams facing this constraint are typically left with two options: wait, often months or years, for additional power to come online at the same site, or stand up a second, separately managed deployment elsewhere and accept the coordination overhead that comes with running two clusters instead of one.
Let's frame the alternative directly: use another data center, with its own available power, and scale the network across the two. Add GPUs at the second site, and both GPU clusters work as a single cluster, not two clusters pretending to be one. As DriveNets puts it, done right, the result is as good as deploying everything in the same data center to begin with.
How WhiteFiber Continuum Puts This Into Production
WhiteFiber Continuum is a live example of this approach. It links WhiteFiber's QTS ATL1 (Atlanta) and QTS SUW1 (Suwanee) facilities, 83 kilometers (52 miles) apart, over 12 dark fiber strands leased from Zayo. A Ribbon Communications DWDM system lights the fiber with 170 channels of 800G ZR+ coherent optics. The result, validated in production: 136 Tbps of aggregate bandwidth at 0.9 milliseconds of round-trip latency across the 83km span, within 8% of the physical limit for light in fiber over that distance.
The network fabric runs on DriveNets AI Fabric, built on Broadcom's Jericho3-AI and Ramon3 chipsets, using scheduled cell-based Ethernet rather than conventional hash-based load balancing. HBM-buffered leaf switches sit at the boundary between each site's internal fabric and the inter-site link, absorbing the burst of traffic that happens when every GPU sends its gradients at once, before that burst can turn into congestion or packet loss. On the storage side, WEKA NeuralMesh replicates data intelligently, pulling exactly what a job needs to the site where its GPU allocation lands, instead of keeping both sites in constant lockstep.
Each facility runs 256 NVIDIA H200 GPUs, 512 total, active-active. Both sites carry live production traffic at the same time, so power that would otherwise sit unused at a second site becomes usable capacity for the same cluster.
Watch: DriveNets on Breaking the Power Ceiling
Dudi Cohen of DriveNets, which supplies the AI network fabric portion of WhiteFiber Continuum, explains how scale-across networking solves one of the biggest constraints in AI infrastructure: power. Watch here
In DriveNets' words: "When you run out of power in your data center, you are locked because you cannot add more GPUs and CPUs in order to serve your workload. Now we have a solution."
Beyond Today: Where This Is Headed
Two sites is where Continuum has been validated, not where the architecture stops. The Jericho3-AI chipset already deployed in Continuum has headroom for a single cluster of up to 32,000 GPUs at 800Gbps each, 62 times the 512 GPUs running across ATL1 and SUW1 today. The next constraint isn't the architecture. It's how many sites, and how much power, WhiteFiber is willing to wire together.
That expansion is expected to take the form of a hub-and-spoke or mesh configuration, combining smaller, modular data center capacity into one larger fabric. Looking further out, Broadcom's newly shipping Jericho4 generation points toward lossless RDMA over distances beyond 100km and full line-rate MACsec encryption on every port. Neither capability is in Continuum today, since Jericho4 is a newer chip generation than what's currently deployed, but both represent a credible path toward solving AI data center power constraints at greater distances.
Who Runs Into AI Data Center Power Constraints
Power constraints hit some AI infrastructure buyers harder than others. They tend to matter most for:
Frontier model labs and AI research teamswhose training runs have outgrown a single site's power envelopeHyperscaler and enterprise AI teamsstuck in an interconnection queue, unable to add megawatts at their current siteGPUaaS and NeoCloud providersthat need to promise elastic scale and uptime to their own customers without building the multi-site engineering themselvesTelco and edge operatorslooking to monetize existing distributed infrastructure, central offices and fiber investment, as GPU capacity
FAQ
What causes AI data center power constraints?
A data center's power budget is fixed by its facility, utility connection, and cooling capacity. Once that budget is spent, more GPUs and CPUs can't be added at that site to serve a growing workload, regardless of available funding.
How does scale-across networking solve AI data center power constraints?
Instead of waiting for more power at the original site, scale-across networking connects a second data center with its own available power and runs the network across both, so GPUs added at the second site work as part of the same single cluster rather than a separate deployment.
How far apart can data centers be and still solve power constraints this way?
WhiteFiber Continuum is capable of reaching 100 km while current production deployment runs at 52 miles (83km), the distance it has been built and validated at today.
How does WhiteFiber Continuum address power constraints?
By utilizing available fractional power at data centers and connecting them together with scale-across networking made for higher performance GPU clusters.
Is this approach to solving AI data center power constraints available today?
Yes. WhiteFiber Continuum has been commercially available since September 23, 2026, built in collaboration with DriveNets and WEKA.
.svg)
