Meridian Cloud Selects Aria Networks for Inference-Ready AI Cluster Designs

Meridian Cloud has selected Aria Networks as the networking partner for its dedicated AI inference cloud, scheduled to launch in the second quarter of 2027. The companies plan to develop standardized, validated network designs that make cluster deployments more repeatable and provide clearer engineering specifications, cost estimates, and production timelines.

Announced November 5, the partnership brings Aria into the planning process before equipment is installed. Its scope includes sizing the back-end GPU fabric, front-end and storage networks, mapping topologies and cabling, and coordinating supply and delivery with Meridian’s deployment schedule.

The first joint reference designs are expected to be available to Meridian customers and facility partners in Q2 2027. The announcement describes a development and deployment program, rather than a completed rollout of Meridian’s new inference cloud.

Making Each New Cluster Less of a New Engineering Project

Meridian’s approach is to run a lean central team supported by specialist partners across engineering, compute, deployment, and networking. Aria will help turn that work into a library of reference designs, each matched to a class of facility.

The idea is to start with a site’s available power and match it to a configuration whose networking requirements have already been validated. That would allow Meridian to establish specifications and costs before construction begins, instead of reworking the fabric as a separate engineering exercise at every location.

For a dedicated-cloud customer, the value of that standardization is predictability. A cluster involves more than the number of accelerators ordered: cable routes, switch capacity, host connections, storage traffic, and the timing of equipment delivery all affect when the system can begin useful work.

“Aria’s Forward Deployed Engineers have been embedded with our team from day one, working with us on network design, supply and deployment planning, and validating every cable, link and configuration before a customer job ever runs,” said Vic Rose, CEO of Meridian Cloud.

Why Inference Puts the Network in the Critical Path

Inference is the process of running a trained model to produce an answer. In a distributed deployment, parts of that work and its supporting data may move between machines, making communication time part of the response the user experiences.

One example is disaggregated inference, where prompt processing and token generation run on separate resources. The system may need to move the model’s key-value cache—the intermediate attention data associated with the request—between them. Congestion along that path can hold up work even when the GPUs themselves are available.

Aria’s technical discussion of inference networking emphasizes correlating host activity with network telemetry. A port that appears healthy in a periodically refreshed dashboard can still experience short disturbances that affect application performance.

The partnership release points to congestion events, degrading optics, and host-side problems that can be missed by tools polling every few seconds. Detecting those events requires a different level of visibility from simply knowing whether a link is up.

For an inference business, the consequence is commercial as well as technical. Time spent waiting for data reduces the useful output of hardware that continues to consume power and incur ownership costs. The relevant operating measures include response latency, useful throughput, and cost per million tokens—not link speed alone.

What Aria’s Deep Networking Adds

Aria describes Deep Networking as a co-designed system combining switching hardware, Aria Operating System, fine-grained telemetry, and specialized agents. The Meridian announcement specifies 800G Aria Switch platforms and a standards-based Ethernet fabric intended to remain open to different accelerator vendors.

The company claims telemetry at 100 to 10,000 times the resolution of traditional tools across switches, optics, and hosts. That is a measurement-resolution claim, rather than a claim that Meridian’s applications will run 100 to 10,000 times faster.

Aria’s product overview provides additional hardware context: its 800G switch has 64 800GbE ports in a two-rack-unit chassis, supports RoCEv2 and Layer 2/Layer 3 networking, and includes a dedicated telemetry processor. RoCEv2 enables remote direct memory access over Ethernet, a capability used to move data efficiently between machines in high-performance clusters.

The same product page describes correlating GPU-host and switch data, selecting diagnostic tools through specialized agents, and executing guided resolution workflows. This puts operational visibility alongside the hardware design: the network must be observable and maintainable after it is commissioned.

The companies have not published joint Meridian deployment benchmarks or a measured cost-per-token reduction for this partnership. Those outcomes will depend on the workloads and configurations deployed, as well as how the network behaves under production traffic.

Meridian’s Dedicated Inference Model

On its website, Meridian positions its offering around reserved capacity matched to a customer’s workload, with options spanning NVIDIA and AMD hardware as well as specialized inference accelerators. It also describes managed orchestration and bare-metal operating models.

Those are descriptions of the broader service proposition, not confirmation that every advertised configuration is part of the initial Aria reference-design library. The partnership announcement does not list the first facility locations, fixed cluster sizes, or customer pricing.

Meridian’s focus on dedicated capacity also changes the planning problem. Instead of relying solely on shared availability, a customer needs a system designed for its expected traffic, model characteristics, and service requirements. Repeatable network designs can make that planning easier while still leaving room for workload-specific choices.

A Partnership Focused on Repeatable Execution

Aria’s role extends from design to hands-on deployment, with lessons from individual sites intended to inform later installations and software improvements. The objective is to make network engineering a reusable part of Meridian’s rollout rather than a recurring source of uncertainty.

The near-term milestone is concrete: joint reference designs and the planned cloud launch in Q2 2027. Beyond that, the partnership’s test will be whether validated designs and finer operational visibility translate into clusters that arrive on schedule and deliver consistent inference performance once customer workloads begin running.