The fundamental challenge of the AI era isn’t just about how many GPUs you can stuff into a rack. It’s about where you find the power to turn them on. As clusters balloon to tens of thousands of units, we’re hitting a wall where single data centers simply can’t provide the megawatts required. As a result, we’re seeing the rise of scale-across networking, a strategy where a single AI cluster is geographically fragmented across multiple facilities.
Nokia is a company that is leading the way in solving these infrastructure challenges. We got to hear about their approach recently during Networking Field Day 40.
The Physical Reality of Scale-Across
In a typical scale-out scenario, you’re connecting servers within a single room using short cable runs. Scale-across is a different animal entirely. We’re talking about distances of 10 kilometers or more, where the network must bridge separate buildings while convincing the GPUs they’re still sitting next to each other. Taken together, these distances introduce latency and bursty traffic patterns that would choke a standard data center switch.
Traditional switches are built for speed and low latency in short bursts, utilizing shallow buffers. When you stretch those links across a city, those shallow buffers overflow almost immediately. In practice, this leads to dropped packets and synchronization issues that can bring a massive AI training job to a screeching halt. To solve this, Nokia utilizes the Broadcom Jericho chipset family, specifically the Jericho 3, because it offers the deep buffering necessary to handle the physics of long-distance data center interconnects.
Engineering for the Power Gap
Since the primary driver for scale-across networking is a lack of power, it’s a bit ironic to solve the problem with power-hungry networking gear. Nokia has taken a different path here by focusing on radical hardware efficiency. Their 7250 IXR chassis, which houses these Jericho chips, consumes up to 30% less power than competing platforms in its class.
This efficiency isn’t marketing magic. It’s the result of some very specific, grounded engineering choices. Nokia completely avoids retimers, which are signal retransmitters that normally live on the PCB. By eliminating them, they reduce heat and power draw while simplifying the board to a single PCB, which inherently increases reliability. They’ve also moved away from standard perforated metal faceplates in favor of honeycomb meshes. These meshes offer a 90% open area for airflow while maintaining electromagnetic interference shielding. It’s a masterclass in building mission-critical hardware that stays cool without needing a dedicated power plant just for the fans.
Automation and the Digital Twin
Managing a network that spans multiple sites is a configuration nightmare if you’re doing it line-by-line. Nokia’s Event-Driven Automation, or EDA, moves the goalposts by using intent-based abstraction. Instead of telling the switch how to move a packet, you tell the system what the workload needs. This is particularly vital because a scale-across network needs to treat external flows differently than internal ones to maintain performance.
The real value in EDA is the digital twin capability. In my experience, nobody wants to test a new routing policy on a live $500 million GPU cluster. Nokia provides digital twins that allow operators to simulate their entire fabric, including the exact configuration rules, in a virtual environment. You can run an entire data center’s worth of nodes on a laptop to validate your design before touching a single piece of production hardware.
The Silicon Strategy
Nokia isn’t religious about their silicon, which is a refreshing change of pace in this industry. They use Broadcom Tomahawk for standard scale-out and Jericho for the deep-buffer scale-across roles. However, they also keep their proprietary FP silicon in the mix. These are fully programmable Network Processing Units that provide a flexible packet pipeline. If your AI workload requires high-touch packet manipulation that off-the-shelf silicon can’t handle, the FP NPU gives you the ability to program exactly how that packet is treated.
Ultimately, this comes down to the concept of Nokia Validated Designs. They aren’t just selling a switch and wishing you luck. They are pre-testing the entire stack, which includes optics, cables, switches, and the actual third-party GPU servers, to ensure the whole thing doesn’t fall over when the first training job starts.
Bringing IT All Together
The shift to scale-across networking is an admission that the power grid is the new bottleneck for AI. Nokia’s approach works because it acknowledges that distance changes the rules of the game. By combining deep-buffer Jericho silicon with power-efficient hardware design and digital twin automation, they’ve built a blueprint for clusters that can’t fit in one building. If you’re planning to split your compute across the metro area, you’d better have a network that understands the difference between a two-meter patch cable and a ten-kilometer fiber run. Hardware reliability and validated architectures are the only things standing between a successful model and a very expensive pile of idle silicon.
To learn more about Nokia and their AI networking solutions for scale-across designs, make sure to check out their website at https://Nokia.com. To see their entire presentation from Networking Field Day, be sure to head over to their presentation appearance page here.
