Building a data center for AI isn’t just about racking as many NVIDIA H100s as your power envelope allows. You’re building a massive, interconnected computer where the network is the backplane, and if that backplane isn’t right, your expensive GPUs spend half their lives waiting on packets. Most legacy networking tools weren’t built for this kind of East-West intensity or the multi-tenant requirements of modern AI factories. Netris changes the math here by treating physical networking like a cloud resource through their Network Automation, Abstraction, and Multi-tenancy (NAAM) platform.
Operators can finally stop treating switch configurations like precious, hand-crafted artifacts. Netris acts as the brain that sits above the hardware, managing the entire life cycle of switches from NVIDIA or Arista through a distributed agent system. It handles everything from the front-end management network to the complex quad-plane architectures required for massive GPU clusters. This shifts the focus from managing individual boxes to consuming a network fabric through VPCs and virtual networks. We had the opportunity to see all the things that Netris is capable of doing recently during Networking Field Day 40.
Validation through Digital Twins
Building a network for a few hundred million dollars’ worth of GPUs means you can’t afford a week of troubleshooting after the gear arrives. Netris uses digital twins to let you build and interact with your entire topology in a simulated environment like CloudSim or NVIDIA DSX Air before a single cable is plugged in. I’ve seen teams spend days hunting for a missing upstream connection or a storage integration mismatch that could’ve been caught in minutes during a simulation. By the time the physical hardware is powered on, the controller already knows exactly what the world should look like. When the physical matches the virtual, you know it was time well spent.
Taken together, these simulations serve as the definitive blueprint for the automation algorithm. When the switches finally boot up, Netris uses zero-touch provisioning to push the pre-validated configurations immediately. If a technician accidentally swaps two cables, the system doesn’t just fail; it tells you exactly which port is miswired. In practice, about 5% of cables in a new deployment are typically flipped, and having a system that flags these errors instantly saves a staggering amount of time.
Scaling the Edge with SoftGate
Providing AWS-like services like NAT and Layer 4 load balancing in a private AI cloud is notoriously difficult because virtual machines just can’t handle the throughput. Netris solves this with SoftGate, which is a software gateway running on bare metal rather than inside a performance-sapping VM or container. It uses an XDP-accelerated data plane written in low-level C to move traffic at line rate. This gives you the multitenant isolation of a VPC without the latency penalties that would otherwise choke an AI workload.
To keep this horizontally scalable, SoftGate uses a proprietary implementation of the Maglev algorithm. Most load balancers struggle to scale because they waste resources syncing state between nodes, but Maglev allows each SoftGate server to be completely stateless. As a result, you can start with a small cluster and add servers as your CPU or RAM utilization spikes. Each node independently calculates the correct destination for a packet, ensuring consistent routing across the whole cluster without any inter-node chatter.
The Netris NAAM Advantage
The complexity of AI networking demands a move away from manual CLI work toward a consumption-based model. Netris provides this by abstracting the physical hardware into VPCs, which lets you reassign GPU clusters between tenants on the fly. It automatically handles the heavy lifting of EVPN BGP and VXLAN, ensuring that even if different tenants use overlapping IP addresses, their traffic remains perfectly isolated. In practice, this means your network team spends less time on VLAN tags and more time on architecture.
The goal is to make the physical network as invisible and reliable as the virtualized resources in a public cloud. By integrating DPUs and hardware-accelerated gateways directly into the automation flow, Netris removes the manual bottlenecks that usually slow down cluster expansion. Taken together, these tools turn the network into a predictable utility.
Bringing IT All Together
Manual switch configuration is the single greatest risk to AI cluster uptime and should be replaced by a centralized automation controller that treats the fabric as a single entity. You should never be “discovering” a routing loop or a cabling error on live hardware. You need to validate the entire topology in a digital twin to ensure the network is revenue-ready the moment it’s powered on. Netris provides this capability as well as critical infrastructure needs like load balancing and NAT through XDP-accelerated gateways running on bare metal to maintain the necessary throughput. It’s the modern way to build a network that will be ready to go when the GPUs show up on the dock.
For more information about Netris and their AI networking solutions, make sure to check out their website at https://Netris.io. If you want to see their entire presentation at Networking Field Day, you can head over to the event appearance here.
