GPU Cluster Fabric Inspection Report
Cluster ALPHA-1 (2,048 GPUs, operated by a fictional neocloud, "Example Cloud Co.") is financeable as a single training cluster once three conditions are met. Its fabric scores 82/100 today and a projected 96/100 after remediation.
The fabric matches its claimed design and is healthy enough to finance, but full-cluster bandwidth is 78% of design and the fabric-manager licence cannot pass to a new operator. Both are fixable before closing.
| Layer | Score (of 20) | Result | One-line reason |
|---|---|---|---|
| 1. Physical inventory | 19 | Pass | All 96 switches present; 3 optics missing from the collateral schedule |
| 2. Cabling vs. design | 15 | Pass with conditions | 14 miswired and 11 dark links out of 4,096 |
| 3. Link health | 16 | Pass with conditions | 31 links with high error rates; 6 running at reduced width |
| 4. Fabric configuration | 19 | Pass | Design claims confirmed; minor firmware drift on 2 switches |
| 5. Takeover readiness | 13 | Fail | Fabric-manager licence tied to the borrower's account |
| Total | 82 / 100 | Pass with conditions | Projected 96 / 100 after the three closing conditions |
The benchmark result (78% of design bandwidth at 2,048 GPUs) is not scored separately. It is the measured consequence of layers 2 and 3.
This report covers the compute (east-west) fabric of one cluster, inspected read-only over the management network. It says nothing about legal title, lien priority or customer contracts.
| Item | Detail |
|---|---|
| Cluster | ALPHA-1: 256 nodes × 8 H100-class GPUs = 2,048 GPUs, 8 scalable units (SU 1–8) of 32 nodes |
| Fabric claimed by borrower | InfiniBand 400G, rail-optimized, 2-tier fat tree, 1:1 (non-blocking) |
| Equipment in scope | 64 leaf + 32 spine switches, 4,096 links, 5,120 optical transceivers |
| Access used | Read-only fabric-manager API and switch access; neighbour discovery; optics diagnostics |
| Active tests | One cluster-wide all-reduce benchmark, run in an agreed 6-hour maintenance window |
| Design source | Borrower's topology file and cabling plan, version supplied at engagement start |
Not tested: GPU servers themselves (covered by the server-verification provider), front-end and storage networks, power and cooling, and customer workloads. No customer data or job contents were read. Findings reflect the cluster during the inspection window only.
Nine findings: three high, three medium, three low. All three high findings are closing conditions; the rest go on the remediation schedule.
| ID | Layer | Severity | Finding | Why it matters to the lender | Fix |
|---|---|---|---|---|---|
| F-01 | 5. Takeover | high | Fabric-manager licence and support contract are in the borrower's name and not assignable | A replacement operator could not legally run the fabric software after a default | Vendor consent letter making the licence assignable to the lender or its nominee |
| F-02 | 2. Cabling | high | 14 links in SU 6 miswired: GPU rails 3 and 4 swapped on 7 nodes | Forces traffic across rails; main cause of the bandwidth shortfall | Re-cable 14 links; re-verify |
| F-03 | 2. Cabling | high | 11 links dark: 9 in SU 8, 2 in SU 3 | SU 8 (256 GPUs) cannot join full-cluster jobs at full bandwidth | Replace optics or cables; re-verify |
| F-04 | 3. Link health | medium | 31 links above the bit-error threshold, 22 of them on one spine | Early sign of failing optics; risk of job interruptions | Replace 31 optics; watch the spine |
| F-05 | 3. Link health | medium | 6 links running at half width (2 of 4 lanes) | Each loses half its bandwidth without raising an alarm | Reseat or replace; re-verify |
| F-06 | 5. Takeover | medium | Operations runbooks last updated 11 months ago | Slows a handover to a new operator | Refresh and escrow runbooks |
| F-07 | 1. Inventory | low | 3 transceivers in service are not on the collateral schedule (replaced under warranty) | Schedule no longer matches the hardware | Update the schedule with the new serials |
| F-08 | 4. Configuration | low | 2 leaf switches one firmware release behind the rest | Minor behaviour differences; complicates support | Upgrade in the next maintenance window |
| F-09 | 3. Link health | low | 18 optics with receive power trending down over 30 days | Likely future failures | Add to watch list; replace at threshold |
The cluster delivers 78% of its design bandwidth at full scale (2,048 GPUs). It is near design within one scalable unit, so the shortfall comes from the cross-unit fabric defects in F-02 and F-03.
| Job size | Measured all-reduce (GB/s per GPU) | Design target (GB/s per GPU) | % of design |
|---|---|---|---|
| 64 GPUs (within one SU) | 370 | 380 | 97% |
| 256 GPUs (one SU) | 355 | 375 | 95% |
| 1,024 GPUs (4 SUs) | 310 | 365 | 85% |
| 2,048 GPUs (full cluster) | 282 | 360 | 78% |
Method: NCCL all-reduce, 8 GB messages, bus bandwidth, run on 2026-10-08. A re-run after remediation is a closing condition; the expected result is at least 95% of design at 2,048 GPUs.
After remediation, ALPHA-1 can be valued as one 2,048-GPU training cluster rather than as separate servers. Today it supports full-cluster contracts only at reduced bandwidth, and a default would leave the lender without the right to run the fabric software.
| Question | As inspected | After closing conditions |
|---|---|---|
| Largest job domain at 95%+ of design bandwidth | 256 GPUs (one SU) | 2,048 GPUs (whole cluster) |
| Can it serve a full-cluster training contract? | Yes, at 78% of design bandwidth | Yes, at 95%+ of design bandwidth |
| Network equipment on the collateral schedule | 96 switches, 5,117 of 5,120 optics | All |
| Could a new operator run it after a default? | No: licence not assignable (F-01) | Yes: licence assignable, configs and runbooks escrowed |
| Takeover package (topology, configs, firmware list, runbooks) | Present; runbooks 11 months old | Current, escrowed, refreshed every 90 days |
Rental rates and recovery values are the lender's assumptions. This report states only the technical condition they rest on.
After closing, six passive checks are made at each periodic re-inspection, from the network's own records; none needs access to customer workloads. A breach is reported to the lender and borrower together and starts the listed action.
| Metric | Baseline after remediation | Trigger | Action |
|---|---|---|---|
| Links matching design | 100% | Below 99.5% for 7 days | Borrower remediation plan within 14 days |
| Links dark or at reduced width | 0 | More than 0.5% of links for 7 days | Same as above |
| Links above bit-error threshold | 0 | More than 1% of links | Watch list; optics replacement report |
| Fabric switches and optics on the collateral schedule | 100% | Any unreported serial change | Schedule update within 30 days |
| Unexplained configuration drift | None | Any change outside an approved change window | Explanation within 5 business days |
| Takeover package age | 0 days | Older than 90 days | Refresh and re-escrow |
One active check: the full-cluster bandwidth benchmark is re-run every 6 months in a maintenance window, and whenever a passive trigger fires. A result below 85% of design starts a lender review.
Conditions before closing
Three conditions must be met and re-verified before funds are released; the other six findings can be fixed within 60 days of closing.
Closing conditions
- F-01: Vendor consent letter making the fabric-manager licence and support assignable to the lender or its nominee
- F-02 and F-03: Re-cable the 14 miswired links and restore the 11 dark links; inspector re-verifies cabling against design
- Re-run the full-cluster benchmark; result at least 95% of design at 2,048 GPUs
Within 60 days of closing
- F-04 and F-05: Replace the 31 high-error optics; fix the 6 half-width links
- F-06: Refresh runbooks and escrow the full takeover package
- F-07: Update the collateral schedule with the 3 replacement transceiver serials
- F-08: Bring the 2 lagging leaf switches to the fleet firmware release
- F-09: Put the 18 degrading optics on the watch list