MicroLink × NVIDIA · Working Session Reference
← Back to Eight Zones
Prepared 4 May 2026 / Version 0.2 Draft / Confidential
Zone 02 · Build · Cluster A

Inter-pod andmulti-site fabric

For Karthik Mandakolathur and the Magnum IO team

200 distributed liquid-cooled cabinets across industrial host sites in the United States. The fabric question is how 200 physically isolated cabinets behave as one programmable fleet. NVIDIA already has the stack. We have the physical envelope. Together, this is the pattern.

Owner
Karthik Mandakolathur
Adjacent: Elad Blatt · sovereign
MicroLink lead
Shane Pather
CTO · Engineering
Programme
200-cabinet rollout
48-month deployment cadence
Site mix
Hospitality · WWTP · industrial
Diversified host channels
Working session
5 May 2026
30 min · Teams
02 The Thesis
200 distributed cabinets is not the edge in the small-data sense. It is a distributed substrate of liquid-cooled compute envelopes, each at an industrial host. The fabric question is how they become one fleet. NVIDIA's networking stack already answers it. Spectrum-XGS is the product that makes 200 isolated cabinets programmable.
01
Sites at maturity
200cabinets
50 kW liquid-cooled BYOS rack at each industrial host. One cabinet per site is the canonical unit.
Programme 48-mo cadence
02
Fabric layers
3
Cabinet · site · inter-site. Three layers, three NVIDIA products at scale.
Architectural Reference
03
Federation fabric
Spectrum
-XGS
The cross-site product that lets tenants federate compute across multiple physical cabinets.
Strategic Karthik · scope
04
Edge per site
Jetson
+ IGX Orin
At every cabinet, facility-side BMS, leak detection, attestation, and edge inference.
Per cabinet 200 nodes

The siteas the unit#

One 50 kW liquid-cooled cabinet at one industrial host site is the canonical unit of the programme. Everything else scales from this picture.

Every site is one cabinet. Every cabinet supports one tenant's GPU stack. Every host gives us a primary loop and 65 kW of three-phase power. From there, we run the substrate.

The cabinet is the investable unit, the operational unit, and the network unit. It is a 50 kW liquid-cooled rack inside an IP55 enclosure, sized to fit through a standard industrial doorway and commission inside the host facility within two days. The tenant brings standard 19-inch EIA-mounted equipment up to 50 kW IT load. MicroLink provides the envelope: power, cooling, network drop, physical security, BMS.

The thermal path is three loops, two of them ours. Server coolant runs sealed inside the cabinet through the CDU and rear-door HX. A facility loop runs from the cabinet's plate heat exchanger to the host's primary loop. The host's process loop runs to whatever they use the heat for: domestic hot water at hospitality sites, anaerobic digester preheating at WWTPs, brewery wort heating, food process water. Heat moves through the boundary at the PHE. No fluid crosses between MicroLink and host.

For Zone 02, the load-bearing fact is this: every cabinet exposes a defined network interface to the tenant and a defined control interface to MicroLink operations. Power drop, network drop, BMS API endpoint. Three things. The tenant runs their own switching above the rack. MicroLink runs the fleet control plane below.

Figure 01 · Cabinet anatomy · 50 kW BYOS unit
One cabinet.One tenant. Three loops. One Jetson.
The investable, operational, and network unit. Tenant zone above the PDU plane. MicroLink zone below it.
Confidence · high
MICROLINK ENVELOPE HOST DOMAIN PHE BOUNDARY IP55 OUTDOOR ENCLOSURE · 50 kW TENANT ZONE · BYOS · 50 kW IT Standard 19" EIA gear Tenant brings GPUs, switching, OS, workload TOR · BYO MICROLINK ZONE · ENVELOPE Mini CDU N REDUNDANT Plate HX 55 kW THERMAL RDHx RESIDUAL Jetson BMS · IGX Dual AC PDUs 63 A · 3-PHASE Network drop + LEAK DETECT Security CHEMISTRY LOOP 1 · SERVER COOLANT · SEALED IN CABINET · 35–65 °C HOST PROCESS LOOP · LOOP 3 Host's primary loop → Hospitality: domestic hot water → WWTP: digester preheating → Brewery: wort heating → Food: process water → District heating: network supply DRY COOLER FALLBACK 100% rejection capability When host demand drops or PHE offline LOOP 2 facility loop 45 °C glycol → TENANT NETWORK · WAN · 100 GbE TYP. CONTROL PLANE OUT · BMS · ATTESTATION · TELEMETRY to MicroLink fleet ops · k0rdent · Mission Control L1 SEALED · OURS L2 FACILITY · OURS L3 PROCESS · HOST TENANT NET CONTROL PLANE
Source · MicroLink Edge cabinet spec v1.0 Method · Cross-section schematic · loops, zones, interfaces

The 200-site fleetat scale#

200 cabinets distributed across the United States at industrial host sites, deployed over 48 months. Host categories diversified across five channels. The fabric ties them together.

The rollout reaches 200 operational cabinets by month 48. Site distribution targets diversification across five host categories, with no single channel exceeding 40% of the portfolio by year 3. Hospitality sites enter through an anchor MSA channel with a branded hotel operator providing access to industrial-scale domestic hot water demand. WWTP, brewery, food, and district heating sites enter through direct industrial outreach.

From the fabric perspective, the 200 sites are 200 leaf nodes of a national-scale substrate. Each is independently power-served, independently connected to the public internet via a tenant network drop, and independently managed by MicroLink at the BMS layer. The federation question is what NVIDIA's stack can do across the substrate, not within any one site.

Figure 02 · The 200-site fleet · stylised distribution
One anchor pod.199 distributed cabinets. One programmable fabric.
Indicative deployment density across continental US. San José anchored as the first node and the 11.2 MW reference site. Other 199 cabinets are 50 kW each.
Confidence · medium · indicative geography
San José ANCHOR · 11.2 MW 576-GPU pod HOST CATEGORIES · 200 SITES Hospitality MSA ~70 sites WWTP ~50 sites Brewery ~30 sites Food & beverage ~30 sites District heating ~20 sites ANCHOR POD San José · 11.2 MW · 576-GPU pod Reference site for the canonical pod design DEPLOYMENT CADENCE · MONTH-END Y1: 15 · Y2: 50 · Y3: 110 · Y4: 200
Source · MicroLink rollout plan v1.0 · indicative siting Method · Stylised geographic distribution · final site coordinates per host MSAs

Three layers of fabricthree NVIDIA products#

The fabric question across 200 distributed cabinets has a clean architectural answer. Three layers, each handled by a named NVIDIA product. None of this is hypothetical.

Cabinet, site, inter-site. Jetson, Spectrum-X, Spectrum-XGS. Plus Mission Control and k0rdent across the fleet. The substrate is ours. The stack is yours. The reference is co-developed.

At the cabinet layer, every site runs Jetson with IGX Orin handling facility-side control: BMS, leak detection, thermal control loops, hardware attestation, security cameras and ingress monitoring. This is the same Jetson hierarchy that Zone 03 expands into a full monitoring story. For Zone 02, the load-bearing fact is that each cabinet is itself an addressable node with NVIDIA silicon at the facility-control boundary, ready to be inventoried, attested, and orchestrated by the fleet control plane above it.

At the site layer, Spectrum-X handles tenant scale-out within and across the few sites where one host runs more than one cabinet. BlueField-3 provides the hardware-attested per-tenant isolation primitive. For most sites, which run one cabinet and one tenant, this layer collapses to the cabinet's own networking. For larger sites or shared-tenant deployments, it scales naturally into the standard Spectrum-X reference.

At the inter-site layer, Spectrum-XGS is the strategic centrepiece. Tenants who want to run a logical cluster across multiple physical cabinets at different host sites need a fabric that handles the cross-site bandwidth and latency without forcing the tenant to rewrite their workload. This is exactly what Spectrum-XGS is for. Sovereign federation across distributed industrial sites is not a future use case; it is the use case the product was designed to enable.

Figure 03 · Three layers of fabric · NVIDIA stack across the fleet
Cabinet, site, inter-site.Three layers, three products, one fleet.
Each band is a layer of the fabric. Each layer is anchored to a published NVIDIA product. The fleet control plane runs across all three.
Confidence · architectural · subject to engineering review
LAYER 03 · INTER-SITE FABRIC · ACROSS THE FLEET Spectrum-XGS Cross-site federation for tenant workloads spanning multiple physical cabinets FEDERATED MESH · 200 NODES · ONE LOGICAL POOL LAYER 02 · SITE FABRIC · WITHIN A HOST SITE Spectrum-X + BlueField-3 Tenant scale-out at the rare multi-cabinet sites · per-tenant isolation primitive ONE HOST SITE · MULTI-CABINET CASE CAB 1 CAB 2 CAB 3 SPECTRUM-X BLUEFIELD-3 · ATTESTED SINGLE-CABINET CASE COLLAPSES TO LAYER 01 LAYER 01 · CABINET FABRIC · PER SITE Jetson + IGX Orin Facility-side control · BMS, leak detect, thermal loops, attestation, edge inference SINGLE CABINET · WHAT JETSON HANDLES BMS FACILITY CONTROL power · cooling Leak / chem FLUID MONITORING PHE health Attestation CABINET IDENTITY to fleet ops Security INGRESS + AI INFER edge ML FLEET CONTROL k0rdent Mission Control DCGM · NVML
Source · NVIDIA published reference architectures · MicroLink fleet design v0.4 Method · Three-layer fabric stack · NVIDIA products mapped to MicroLink fleet topology
§
Spectrum-XGS finds its public-sector reference customer here
A federated cluster spanning 8 to 12 physical cabinets at different industrial host sites is exactly what cross-site fabric was designed to enable. Sovereign workloads with locality requirements, neoclouds with capacity diversification, public-sector tenants with data-residency boundaries — all of them want to treat distributed cabinets as one logical pool. The 200-site rollout is the substrate. Spectrum-XGS is the fabric. The reference customer is right here.

Three tenant patternsacross the fleet#

Tenants will show up at three levels of sophistication. The fabric has to handle all three from the first deployment, because the most demanding pattern can arrive on day one.

Single cabinet, multi-cabinet, federated cluster. The fabric is the same fabric. The tenant decides which pattern they need.

Pattern A Year 1+
Single-cabinet tenant
T1
Neocloud, GPU reseller, or regional AI startup rents 50 kW at one site. Anchor tier or free capacity tier. No federation needed. The fabric handles tenant-rack networking and the public internet drop. Most year-1 deployments.
Pattern B Year 2–3
Multi-cabinet single tenant
TENANT T2
Sovereign workload operator takes 4 cabinets across 4 host sites for redundancy and locality. Same tenant, different geography. Cross-site coordination via Spectrum-XGS. Hardware-attested isolation per cabinet.
Pattern C Year 3+
Federated cluster
Tenant runs a training job across 8 to 12 cabinets. Treats the fleet as one logical pool. Spectrum-XGS handles the cross-site fabric. k0rdent + Mission Control orchestrate the cluster. This is the use case Spectrum-XGS was designed for.

The askand what we bring#

The 200-site rollout is the substrate. The networking stack is the fabric. The reference is co-developed. Here is what we are asking the Magnum IO team for, and what we contribute in return.

Tier · Cross-site fabric reference architecture
200 cabinets, one fleetSpectrum-XGS as the federation product, deployed at scale

San José is one anchor pod. The 199 distributed cabinets are the substrate that makes Spectrum-XGS a deployed reference at hundreds of nodes, not a pilot.

From the Magnum IO team
  • Three-layer fabric architecture review
  • Spectrum-XGS deployment guidance for the federation pattern
  • BlueField-3 integration spec at the cabinet
  • Jetson reference for facility-side BMS at scale
  • Technical cadence through Q4 2026
From Karthik specifically
  • Internal advocacy for the rollout as a Spectrum-XGS reference
  • Co-authored federation pattern for distributed industrial sites
  • Public-sector demo case for cross-site fabric
  • Bridge to Elad Blatt for sovereign tenant scenarios
What MicroLink contributes
  • Live deployment at the substrate layer · 200 sites by Y4
  • Spectrum-XGS at hundreds of nodes
  • Diversified host channels · hospitality, WWTP, brewery, food, district heating
  • Anchor pod at San José as the canonical 11.2 MW reference
  • Federation telemetry for joint white paper