Define the unit before the BOM ·
Scalable units and fabric evidence
“One SU” is not a safe unit for an AI factory bid. Pin the architecture, source revision and physical quantities first. Then reconcile the switch, optics and cable BOM against a port-level connectivity ledger.
Which scalable unit does the bid mean?
| Basis | Meaning | Use in a bid |
|---|---|---|
| NVL72 Enterprise RA; network chapter, retrieved 8 Sep 2026 | One rack-based SU: 18 trays, 72 GPUs. Its eight-SU example has eight racks and 576 GPUs. | Record rack-based quantities and the selected fabric table. |
| Mission Control installation guide 2.2.0 | Compute plane: eight racks per superunit (SU). | Name the software/deployment grouping and version explicitly. |
| AI-cloud requirements; BM01 | 512 GPUs is an example benchmark unit. | Agree the actual acceptance lot and target. It does not redefine an NVL72 rack. |
Rack and fabric basis: NVIDIA NVL72 Enterprise RA network chapter.
Version-pinned software grouping: Mission Control 2.2.0 introduction. This is the version inspected for this definition; it is not a claim that 2.2.0 is the current supported deployment release.
Acceptance-lot example: NVIDIA AI-cloud requirements, BM01.
Check the arithmetic without inventing a case study
Take a hypothetical 288-rack NVL72 design. At 72 GPUs and 18 trays per rack, it has 20,736 GPUs and 5,184 trays. Grouping eight racks together yields 36 groups. The same physical design therefore has 288 rack-based SUs under the Enterprise RA terminology, or 36 eight-rack groups under the stated software grouping.
That arithmetic explains how the labels can differ. It does not verify an architecture or its software compatibility. Put racks, trays and GPUs alongside every SU count, then name the grouping in full. This hypothetical calculation is not a delivered customer result.
Choose a fabric on a named basis
Start with the platform and workload, then name the chosen reference design, fabric family, scale, rail layout and oversubscription. Keep alternatives in separate branches. A switch that fits an alternative architecture must not quietly enter the selected design’s bill of materials. Confirm supported device and software combinations against the deployment version before placing an order.
This is a concrete review need: Neon's supplied GUI run recorded the observed vendor-projection mismatch. GB300/Spectrum-4 options appeared for an HGX B300/Quantum-X800 scenario. The visible classification mechanism did not prove a correct projection.
Reconcile switches, optics and cables
For an NVL72 Spectrum-X design, start with the Enterprise RA’s fabric-specific BOM tables. Record the exact table and scale rather than lifting a count from a neighbouring example.
| Record | Check before release |
|---|---|
| Switch BOM | Selected model, quantity, role, port speed and software support evidence. Separate compute, storage, management and external fabrics. |
| Port ledger | Each physical endpoint has a stable ID. Record logical lanes, breakout mode and reserved capacity. No physical port may be allocated twice. |
| Optics BOM | Match both ends to media, speed, reach and form factor. Count modules separately from logical lanes: breakout does not make these counts interchangeable. |
| Cable BOM | Reconcile every ledger link with the chosen cable or media assembly. Add length and route evidence from the site design before installation. |
| Software support | Pin Mission Control release/SBOM and applicable plugins. Check Network Operator interoperability separately; it does not certify a physical fabric BOM. |
| Exceptions | List spares and unpopulated/reserved ports explicitly. Require an owner for every unexplained quantity difference. |
One observed trace, and its limit
In the 4 September source run, searching LINK-EW-N001-R01 opened a physical trace sheet with endpoint, switch, port and link identities. This is the observed path from the logical Clos view into the physical ledger. The recovered screenshot clips the far-right endpoint. We therefore publish the original capture with that limit and require the physical ledger export for the complete connection.
The run identity and qualification are recorded in the physical-connectivity FAQ.
For a review, ask for the actual trace export and its baseline ID. Follow each endpoint back to its BOM device, reconcile ports and optical assemblies, then attach installation and link-test evidence separately. The ledger proves the planned connection; a measured test is needed to establish the installed link’s behaviour.
For a proposed test and an explicit NOT RUN result, use the GB300 acceptance-evidence example.
Check the wider delivery requirements in the NCP requirements self audit.
The brief
How Neon fits your AI factory delivery pipeline
Nine pages for the people who sign: where Neon sits in an enterprise AI factory delivery pipeline, what it owns, what it never touches, and the three questions executives ask. Sent as a PDF.
The Neon briefings explain the design and delivery choices behind the package.
Email me the brief