Stronger Cyber Defense with Advanced CISO-Level Protection
Stronger Cyber Defense with Advanced CISO-Level Protection
Sophos Gold Partner
Monitoring & Support
Deployment Options
Nationwide Coverage
AI networking is the part of the network built specifically to move data between GPUs, not between regular servers. A training cluster can have hundreds of GPUs trying to exchange data simultaneously, and if the network between them stalls even briefly, the entire job waits, no matter how powerful the GPUs are. Arista’s AI Networking and Data Center Infrastructure Solutions are built around Ethernet rather than the older InfiniBand standard, using a purpose-built portfolio called Etherlink to keep that GPU-to-GPU traffic moving at the speed training actually needs. The scale of demand here is real: Arista’s AI center networking already passed its $1.5 billion target in 2025, and the company expects that number to roughly double again in 2026.
Netmate IT designs and deploys this kind of AI fabric from its Bur Dubai office, with engineers who size GPU clusters and tune congestion control settings against real training workloads rather than generic lab benchmarks. Arista is Netmate’s trusted infrastructure partner for this work, and the team’s focus starts with the UAE before extending into the GCC, Kenya, Nepal, and the wider Africa region, for any organization running AI workloads serious enough to need a network built for them specifically.
Etherlink is Arista’s portfolio built specifically for the back-end network that connects GPUs, not the front-end network connecting users to applications.
It’s standards-based Ethernet, so it doesn’t require a separate, proprietary fabric as InfiniBand does. A team that already runs Ethernet everywhere else in its data center doesn’t have to learn and maintain a second networking stack just for AI.
In 2023, Arista was a founding member of the Ultra Ethernet Consortium, which aims to promote the adoption of Ultra Ethernet Transport, an industry standard for data transport that is specifically designed for AI and high-performance computing traffic.
As Arista is in the bootstrapping phase, rather than just following along the beaten path after the fact, this will allow Arista to also provide hardware support for new UEC features in advance of the competition.
AI training traffic doesn’t behave like normal web or database traffic. It moves in bursts, with many GPUs trying to talk at once, and a single slow link can stall an entire job.
Arista EOS supports PFC, DCQCN, and ECN, along with packet trimming and packet spraying, all of which exist to keep that bursty traffic flowing smoothly instead of backing up behind one bottleneck.
Current AI fabrics run on Arista’s 800G R4 series and 7060XE7 series switches, giving clusters the bandwidth modern GPUs actually need to stay busy instead of waiting on data.
Arista has already signaled 1.6 terabit switching is coming next, which matters for anyone planning a cluster now that they don’t want to outgrow in eighteen months.
Based on Arista’s AVA machine learning platform, AI Analyzer provides a team with visibility beyond the switch level and down to the individual workload and NIC level.
When training is slow, this is how an engineer is able to determine if the issue is a particular GPU, a particular NIC, or it’s all about the network itself rather than guessing.
In older AI deployments, the storage, compute, and GPU cluster were typically served by separate and isolated networks.
The idea is so comprehensive that it is bundled into a single Ethernet-based architecture, which Arista calls the AI Center. That is, rather than three independent systems requiring three independent experts, one team is operating one network.
Netmate IT designs and deploys Arista’s AI networking fabric from its Bur Dubai office, with engineers who size GPU clusters, configure RoCE and congestion control settings, and tune the fabric against a client’s actual training workload instead of a generic benchmark. Getting cluster sizing wrong here is an expensive mistake, since the GPUs themselves cost far more than the network connecting them, and an undersized fabric means those GPUs sit idle waiting on data.
Netmate’s work in AI networking runs across the UAE first, then extends into the GCC, Kenya, Nepal, and Africa, supporting AI labs, cloud providers, and enterprises that need a network built specifically for how GPUs actually communicate. Whether a client is standing up its first small training cluster or scaling an existing one past a few hundred nodes, Netmate brings the same certified Arista design process to the project.
A research lab in Ajman could be training models using dozens of GPUs, and experiencing network bottleneck issues before the GPUs even have their fill! A cloud hosting provider in Dammam may be creating a GPU as a service service that desires to outgrow a handful of hundred nodes without having to completely redesign. A research centre in a University could be performing genomics or climate modelling tasks that require massive transfers between storage and compute every day.
A fintech company in Nakuru may be using its own sensitive data to train models which they are unable to send to a third-party cloud. Butwal has a telecom operator who could be developing an edge AI capability to process data as it’s created rather than sending it back to a far-off data center. In the end, the decision between Ethernet and InfiniBand, whether it works best with your organization, is going to come down to how practical it is, in your case: Does the team already have Ethernet deployed elsewhere, or does InfiniBand require them to get on their feet and deploy a second network to support AI?
A normal data center network is built for traffic between servers and users, which tends to be steady and predictable. AI training traffic moves in sudden bursts, with many GPUs exchanging large amounts of data at the same moment, and a training job can only finish once every single piece of data reaches every GPU that needs it. One slow link anywhere in that path can stall the entire job, which is a failure mode a regular data center network was never designed around.
RoCE, short for RDMA over Converged Ethernet, lets GPUs exchange data directly over an Ethernet network with much lower overhead than a standard connection. The newer version, RoCE v2, adds routing support so it works across larger networks, not just within a single rack. It's effectively what makes Ethernet fast enough to compete with InfiniBand for GPU traffic in the first place.
It's increasingly a real competitor rather than a future promise. Arista has publicly reported winning the majority of its large AI networking bids against InfiniBand in recent competitive deals, and the broader industry shift toward Ethernet-based AI fabrics backs that up. InfiniBand still has a place in certain specialized deployments, but for most organizations already running Ethernet everywhere else, it's no longer the automatic default it once was.
It's an industry group, founded in 2023 with Arista as one of its founding members, working to build a next-generation Ethernet transport standard specifically for AI and high-performance computing traffic, intended to eventually replace RoCE. For a business buying AI networking equipment now, working with a vendor deeply involved in setting that standard means a better chance the hardware purchased today keeps up as the standard matures, rather than needing replacement.
Arista's AI spine and Etherlink platforms are currently well suited to clusters of approximately 100,000 GPUs, ranging from a small research cluster to some of the biggest training deployments currently in operation. Therefore, sizing for a specific deployment should be done before ordering hardware, as the actual maximum size of a cluster depends on the models of switches selected and the form of the switch fabrics.
This is dependent on the cluster size and the type of GPU used: newer and more powerful GPUs create more network traffic per GPU as compared to their older counterparts. A few dozen older GPUs might perform very well with lower-speed switching, and if it is a bigger or newer deployment, it could be a true bottleneck without 800G. This is the type of sizing problem that should be solved in conjunction with an engineer before buying hardware.
When training is complete, AI Analyzer provides visibility down to individual workloads and network interfaces rather than just an overall traffic view on the switch. If a training job all of a sudden slows down, it gives an engineer a way to determine if it's a specific GPU, a specific NIC, or a real network problem rather than hours of ruling things out one by one.
