LEVIATHAN SYSTEMS

Networking

One AI Cluster Across Data Centers: Define the Inter-Site Network Handoff

Technical contentPublished Updated

In the fictional Alder/Birch proposal, Birch reserves transport shelf EX-TB1 without a matching receiving entry and lacks a defined continuation from fabric switch EX-BSW/P9 to provider panel EX-BDP/07. Establish whether the transport arrangement requires a customer shelf, then assign the local connection that reaches its handoff. These are different decisions: confirming a shelf is unnecessary leaves the continuation to be supplied and installed. Relate both sites' physical packages to the same inter-site service, and track local completion, provider delivery and workload acceptance separately.

Separate Birch's shelf question from its missing connection

EX-AC19 follows a buyer who intends one distributed training job to use resources at Site Alder and Site Birch. Both sites have issued rack and port records for their internal scale-out packages. The inter-site brief says connectivity included, leaving the actual transport arrangement unresolved.

Birch's draft elevation reserves shelf EX-TB1, but its receiving list contains no matching equipment. A provider notice identifies a handoff at panel EX-BDP/07 in a building meet-me room. The tenant cable list starts at fabric switch EX-BSW/P9 and contains no continuation to that panel.

The provider and customer network owners must first establish whether EX-TB1 belongs in the selected arrangement. If it does, its approved equipment package and supply responsibility need resolving. Separately, the customer must assign the omitted panel-to-fabric continuation. Buying a shelf would leave that connection unexplained; assigning a cable task would leave the equipment requirement unexplained. The buyer needs both answers before treating the proposal as a complete physical scope.

Follow the connection from Birch's fabric to Alder's fabric

At Birch, the local fabric connects the site's compute nodes. Its connection toward the other site reaches a local fabric endpoint, represented here by EX-BSW/P9. A separate local continuation must reach the confirmed provider handoff at EX-BDP/07. Beyond that boundary lies the provider's span or service, followed by the local continuation between Alder's handoff and Alder's fabric. Each object can have a different supplier, performing party and completion record.

This trace explains why the issued internal cabling records are insufficient: they describe the local fabric but omit the part that reaches the provider. It also explains why the provider's panel reference cannot settle Alder's local work. Alder needs its own confirmed endpoint and continuation, related to the opposite end of the same inter-site arrangement.

One logical edge on a fabric drawing can traverse several physical segments; it is not a cable quantity. Keep local cable identities, equipment assets and the provider service reference distinguishable while relating them through the confirmed handoffs.

The architecture terms describe communication roles within this picture. NVIDIA identifies NVLink and NVLink Switch as scale-up technology for GPU communication within servers and across a rack. Scale-out connects compute nodes through the selected cluster network. Scale-across adds communication between distributed data centers. On August 22, 2025, NVIDIA announced Spectrum-XGS Ethernet within the Spectrum-X platform, naming Spectrum-X switches and ConnectX-8 SuperNICs. That dated platform description gives the network owner an architecture reference; the project still needs its approved interfaces, equipment and service arrangement.

References: [N1][N2][N3]

Let the transport arrangement determine the shelf requirement

Birch's shelf question turns on what the customer receives. A dark-fiber arrangement supplies a span over which the customer needs supported transmission equipment. The appointed network/optical owner must determine that equipment from the actual span and interfaces. A separate transport shelf may or may not be part of the resulting design; the words dark fiber do not select EX-TB1.

Equinix provides specific examples. Fiber Connect uses leased dark fiber for connections within a metro area. Metro Connect offers optical Layer 1 and packet-based Layer 2 services and includes transmission equipment for the inter-IBX fiber connection. Its published scope includes a cross-connect at the customer demarcation, typically a cage fiber distribution panel; Extended Metro Connect terminates in a building meet-me room. Patch leads between the panels and customer equipment are excluded.

Those Equinix boundaries show how a provider can supply transmission equipment while leaving a local connection for another party. They do not identify the fictional provider or establish either product's suitability for Spectrum-XGS. The actual service return must resolve Birch's equipment responsibility and termination boundary.

An active-service handoff can therefore remove a customer shelf requirement while leaving the panel-to-fabric continuation outstanding. The network/optical owner remains responsible for the approved interfaces and complete optical path. City-to-city distance or connector appearance cannot select the optic. The colocation cross-connect guide linked below covers the detailed reconciliation of provider patching and remaining tenant work.

References: [E1][E2]

Apply the provider return to Birch and the matching Alder package

For EX-AC19's documentary resolution, assume the appointed provider and customer network owners confirm a managed transport handoff at EX-BDP/07 with no customer transport shelf required at Birch. The responsible document owners reissue the affected elevation and equipment list to reflect that decision. EX-TB1's missing receiving entry can now be understood as an unnecessary equipment expectation in the draft.

The same return leaves a real scope gap between EX-BSW/P9 and EX-BDP/07. The customer defines local package EX-LB1 for that continuation. Its physical performing party needs the approved interfaces, parts, route and patch positions for the current design. The customer must also identify the material supplier, delivery responsibility and permitted work area. Where the handoff is outside tenant-controlled space, the applicable customer/provider access and work release remain necessary inputs. This is a purchasing-scope explanation; installation methods come from the applicable OEM and site packages.

At Alder, the customer/network owner reconciles the local equipment-to-handoff package with the other end of the same transport record. Matching the packages means confirming that each reaches its applicable service termination. It does not mean copying Birch's equipment or cable quantities into Alder's scope.

The shared inter-site reference now lets the buyer trace the provider service to both local packages while retaining their separate cable and asset identities. A change to a returned endpoint can be assigned to the affected package owner, who must reconcile its applicable revision.

This return resolves the shelf expectation and defines the remaining local work. It creates no installed connection or passing test. Had the owners instead selected customer-supplied transmission over fiber, they would have needed the corresponding transmission equipment and its receiving, mounting and connection responsibilities before assigning that different scope.

Keep fabric suitability with the network owner

Distance adds propagation delay; queuing can add variable delay. That variation matters when a distributed workload waits on communication. Providing a physical path addresses one dependency, while its behavior under the intended workload remains a network and application question.

NVIDIA's September 9, 2025 technical explanation describes Spectrum-XGS using distance-aware, telemetry-based congestion control and adaptive routing. The distance relationship informs switch load balancing and the SuperNIC's injection rate: how quickly it sends traffic into the network. These controls address communication behavior across the fabric; their description supplies no reach, latency or job-performance guarantee for Alder and Birch.

The customer network owner therefore needs to establish which transport characteristics the selected fabric requires, where the required telemetry is available, and who confirms the exact hardware/software combination against the service and design revision. That decision belongs alongside the physical packages so an unresolved compatibility question remains visible to the buyer.

References: [N2]

Identify what remains in each of three closeout states

The same Alder/Birch connection can reach different reporting positions. Local handoff evidence should describe the equipment and endpoints actually installed, assembly and patch references, approved departures, agreed check records and open items. The provider's delivery evidence describes its own termination boundary. Compare each returned output with the scope it was meant to close.

If the provider reports the inter-site service delivered while EX-LB1 still awaits physical completion, the missing object is Birch's local continuation. Its appointed performing party owes the agreed connection and handoff evidence, supported by the customer's material and access inputs. The customer package owner must resolve any outstanding input preventing that work. The provider's delivered status remains limited to its service boundary.

If both local teams return completed physical packages while transport delivery remains pending, the missing output is now the provider's delivery evidence for the inter-site span or service. The customer service owner pursues that dependency with the provider. Birch's resolved shelf question stays resolved, and the local records retain the scope they actually close. Repeating local installation or ordering EX-TB1 would not establish provider delivery.

If both local packages and provider delivery evidence are present, the network and application owners still assess the configured fabric and the intended distributed job against their agreed criteria. Local cable observations cannot establish congestion behavior, effective job bandwidth or an accepted training run. The remaining decision is workload acceptance with those owners, rather than another unspecified physical task.

Where physical measurements are requested, agree their boundary, method, instruments, personnel and acceptance criteria before including them in the installation scope. This keeps the buyer's closeout question precise: which connection, service output or system decision is still owed, and who can supply or accept it?

Bring the remaining physical work to the project inquiry

Leviathan's Complete Physical Infrastructure Deployment service is the relevant starting point for discussing receiving, rack integration, onsite installation, cabling and physical handoff records. In the Alder/Birch example, a useful inquiry would explain that the Birch shelf requirement has been resolved and identify EX-LB1 and the corresponding Alder scope as the local packages to discuss.

Describe the required physical stages at each location, the appointed equipment/material suppliers, confirmed provider demarcations and unresolved dependencies. Agree the actual tasks, checks, resources and availability for the engagement. Inter-site fiber leasing, carrier service delivery, network control-plane configuration and software commissioning remain with their appointed providers and system owners.

Use the service and project-inquiry links below with a nonconfidential outline of the two site scopes and the remaining boundary decisions. Detailed operational records can follow through an agreed authorized channel.

References: [L1]

Primary sources

  1. [N1] NVIDIA — Spectrum-XGS Ethernet announcement, August 22, 2025Checked 2026-10-21. August 22, 2025 architecture announcement and named platform components; project compatibility and performance require the selected design.
  2. [N2] NVIDIA — How to Connect Distributed Data Centers Into Large AI Factories with Scale-Across Networking, September 9, 2025Checked 2026-10-21. September 9, 2025 technical explanation of distance-aware routing and congestion control. Benchmark and ROI claims are outside this article's purchasing assumptions.
  3. [N3] NVIDIA — NVLink and NVLink SwitchChecked 2026-10-21. Current product description used for scale-up communication within servers and across a rack; generation-specific numerical specifications are outside this article's scope.
  4. [E1] Equinix — Metro Connect product documentationChecked 2026-10-21. Provider-specific distinction between optical Layer 1 and packet-based Layer 2 delivery, included transmission equipment, demarcations and excluded equipment patch leads.
  5. [E2] Equinix — Fiber Connect product documentationChecked 2026-10-21. Provider-specific leased-dark-fiber example within a metro area; the actual route, transmission design and service eligibility require their responsible owners.
  6. [L1] Leviathan Systems — Complete Physical Infrastructure DeploymentChecked 2026-10-21. Physical deployment stages and handoff scope; actual tasks, checks, resources and availability are agreed for each engagement.

Define the physical work at both ends

Share a nonconfidential outline of the site scopes, equipment, provider handoffs and unresolved dependencies. Our team will follow up to discuss requirements, scope and availability.

Discuss a ProjectEmail Project Scope