# SiliconX Lot Intelligence — Data Methodology

This document catalogs every data point in a Lot Intelligence Report, where it
comes from, and how confident we are in it today (May 2026).

## Confidence tiers

| Tier | Meaning | Example |
|---|---|---|
| **A — Sourceable today** | Public, scrapeable, or available via cheap API. Reliable enough to ship. | Refurb retail listings, OEM spec sheets |
| **B — Sourceable with friction** | Requires aggregating multiple noisy sources, scraping eBay sold listings, or buying a low-cost data feed. Ship with disclaimers. | Median secondary-market price by month |
| **C — Requires our own data** | Only available from SiliconX telemetry, partner ITADs, or our own transaction history. Today these are estimated/modeled. | Verified condition premium, buyer pipeline depth |
| **D — Customer-supplied** | The lot owner provides this at intake. | Original purchase price, deployment date, custody history |

---

## Lot-level fields

| Field | Source | Tier | Notes |
|---|---|---|---|
| `lot_id` | Internal | — | Generated at intake |
| `owner_org` | Customer | D | Intake form |
| `intake_date` | Customer | D | Intake form |
| `location` | Customer | D | Region matters for shipping comps |
| `original_purchase_price` | Customer | D | Optional but increases FMV confidence |
| `deployment_date` | Customer | D | Drives age-based depreciation |
| `units_count` | Customer | D | Asset-level rollup |

## Asset-level fields (per unit in lot)

| Field | Source | Tier | Notes |
|---|---|---|---|
| `sku` | OEM spec sheet | A | NVIDIA / Dell / Intel public catalogs |
| `serial_number` | Customer / scan | D | Triggers OEM warranty lookup |
| `manufacture_date` | OEM serial decode | A | Public for most NVIDIA, Dell, Intel SKUs |
| `age_months` | Computed | A | `today - deployment_date` |
| `condition_grade` | SiliconX inspection | C | "Like New / Normal Wear / High Wear" — today inferred from age + telemetry, eventually graded by inspection |
| `telemetry_hours` | Snapshot agent | C | Power-on hours; today modeled, eventually live |
| `thermal_score` | Snapshot agent | C | Stress-test result; modeled today |
| `unit_fmv` | FMV engine | A/C blend | See FMV engine below |

## Comparable sales (the core of the FMV engine)

| Field | Source | Tier | Notes |
|---|---|---|---|
| `comp_sku` | Public listing | A | Refurb sites, eBay sold listings |
| `comp_price` | Public listing | A | Refurb retail = upper bound; eBay sold = realized |
| `comp_condition` | Listing description | B | Heuristic mapping needed |
| `comp_date` | Listing date | A | |
| `comp_source_url` | Public listing | A | Citeable in reports |
| `comp_channel` | `refurb_retail` / `ebay_sold` / `private_b2b` / `auction` | A/C | First two are scrapeable today |

### Sources we can pull from today (Tier A/B):

- **Refurb retail (Tier A):** Alta Technologies, ServerMonkey, NewServerLife, Bytestock, ServerSupply, Network Outlet, Oplexa
- **Marketplace (Tier B):** eBay sold listings (need scraping or eBay Developer API), compute.exchange secondary market
- **Indices (Tier B):** Hashrate Index secondary GPU report (quarterly), TRG Datacenters guides
- **Cloud-equivalent rates (Tier A, indirect signal):** AWS / GCP / CoreWeave on-demand pricing — useful for floor-pricing logic

### Concrete published comp ranges (May 2026):

| SKU | New retail | Refurb retail | Realized secondary |
|---|---|---|---|
| H100 SXM5 80GB | $27K–$40K | $22K–$30K | $18K–$28K (1yr+), down to $7–12K at 2yr+ |
| A100 80GB SXM | $3.0K–$3.8K (current new) | $1.8K–$2.4K | $7.8K–$18.9K full SKU range |
| A100 80GB PCIe | $2.5K–$3.2K | $1.8K–$2.4K | $7.8K–$15K |
| Dell XE9680 (8× H100) | $157K (refurb config) | $150K–$180K used 8-GPU | comp via channel |
| Dell PowerEdge R760 | $25K+ new | from $18,748 | varies by config |
| Xeon Platinum 8480+ | ~$10K new MSRP | ~$2.7K–$4.1K | $2.5K–$4K |

## Market signals

| Field | Source | Tier | Notes |
|---|---|---|---|
| `demand_level` | Buyer activity in our marketplace | C | Today: estimated from web search trends + listing velocity |
| `supply_level` | Refurb listing inventory | A/B | Scrapeable from refurb sites |
| `comps_30d` | Aggregated comp count | A | From scraped sources |
| `avg_close_time` | Our transaction history | C | Modeled today |
| `next_gen_pressure` | Public roadmap signals | A | NVIDIA Blackwell ramp dates are public |
| `verified_premium_pct` | A/B test on our sales | C | Modeled today; needs ~50 sales to validate |
| `marketability_tier` | Computed | A/B | Derived from demand + supply + age |

## Depreciation curve

Built from public data:

- **Oplexa H100 resale by age:** <1yr $18-25K, 1-2yr $12-18K, 2+yr $7-12K
- **Hashrate Index used GPU pricing report** (quarterly)
- **Cross-validated** against refurb listings by manufacture year

This is **Tier A/B** — we can publish and defend it today.

## What requires SiliconX-specific data (Tier C)

These remain modeled until we have transaction history:

1. **Verified condition premium** — needs side-by-side sales of verified vs. unverified lots
2. **Buyer pipeline depth** ("12 active buyers") — needs our marketplace
3. **Average close time** — needs transaction logs
4. **Recommended split strategy ROI** — needs outcome data on prior splits
5. **Action plan timing** — derived from historical workflow data

For demos we can clearly mark these as **modeled estimates** based on industry
benchmarks (e.g., Hashrate Index notes "the ITAD ecosystem is mature" with
established close cycles).

## FMV engine logic (current heuristic, pre-ML)

```
unit_fmv = base_sku_price(sku, manufacture_year)
         × age_factor(age_months)
         × condition_factor(condition_grade)
         × market_factor(demand_level, supply_level)

base_sku_price = median(refurb_retail_comps_last_90d)
age_factor     = depreciation_curve(sku_family, age_months)  # A100/H100 family curves
condition_factor = {Like New: 1.10, Normal Wear: 1.00, High Wear: 0.78}
market_factor  = clamp(demand_index / supply_index, 0.85, 1.15)

lot_fmv = sum(unit_fmv) × bulk_discount(units_count)
itad_baseline = lot_fmv / (1 + verified_premium_pct)
```

The `condition_factor` and `verified_premium_pct` are the **two parameters that
will change most** as we gather real sales data.

## Citable sources (used to build the May 2026 dataset)

- Hashrate Index — [Used GPU Market: A100 & H100 Pricing, Depreciation](https://hashrateindex.com/blog/used-gpu-market-pricing-deprecation-secondary-ai/)
- IntuitionLabs — [NVIDIA AI GPU Prices guide](https://intuitionlabs.ai/articles/nvidia-ai-gpu-pricing-guide)
- Oplexa — [Nvidia H100 GPU Resale Price and Market Analysis](https://oplexa.com/product/nvidia-h100-gpu-resale/)
- Alta Technologies — [Used Dell PowerEdge XE9680 Servers](https://altatechnologies.com/collections/dell-poweredge-xe9680-server)
- NewServerLife — [Refurbished Dell PowerEdge R760](https://newserverlife.com/server-models/dell-poweredge-r760/)
- Bytestock — [Intel Xeon Platinum 8480+ refurb](https://shop.bytestock.com/intel-xeon-platinum-8480-cpu-processor-56-core-2-00ghz-105mb-l3-cache-350w-srm7h)
- ServerSupply — [Dell 338-CHOS Xeon 8480+](https://www.serversupply.com/PROCESSORS/Intel%20Xeon%2056-Core/2.0GHz%20-%2016GT%20UPI/DELL/338-CHOS_373162.htm)
- Network Outlet — [Dell PowerEdge XE9680 8× H100 SXM listing](https://networkoutlet.com/products/dell-poweredge-xe9680-ai-gpu-server-8-nvidia-hgx-h100-80gb-sxm)
- Introl — [Secondary GPU Markets guide 2025](https://introl.com/blog/secondary-gpu-markets-buying-selling-used-hardware-guide-2025)
- CompuCycle — [AI Hardware Refresh Crisis: ITAD Solutions](https://compucycle.com/ai-hardware-refresh-crisis-itad-data-centers/)

## Silicon Passport profile (OCP Tier 1 / Tier 2 / Tier 3)

The Silicon Passport is a per-asset data record that follows OCP's circularity
hierarchy: every retired asset gets the **minimum profile** (asset identity,
lifecycle, sanitization, policy flags) plus **one tier extension** — Tier 1
(whole-system upcycle), Tier 2 (component reuse), or Tier 3 (material recovery).
A given asset is in exactly one tier at a moment in time. Tier 1 failures fall
back to Tier 2; Tier 2 failures fall back to Tier 3.

This section lists every passport field, where it comes from, and our
confidence tier (A = directly observed, B = derived/blended, C = modeled).

### Minimum profile (every asset)

| Field | Source | Tier |
|---|---|---|
| `asset_id`, `serial`, `sku`, `asset_type` | Verification intake (serialized scan) | A |
| `asset_identity.oem_name` / `product_family` / `technology_generation` / `capacity_or_rating` | OEM database lookup keyed on serial | A |
| `asset_identity.manufacture_date` | OEM warranty registry | A |
| `lifecycle.primary_owner_id` | Customer onboarding record | A |
| `lifecycle.deployment_start_date` / `retirement_date` | Asset management system handoff | A |
| `lifecycle.location_region_at_retirement` / `last_host` / `prior_hosts_count` | Snapshot telemetry historical record | A |
| `sanitization.data_bearing` / `sanitization_required` | Asset class lookup (servers true; GPUs with HBM-only false) | A |
| `sanitization.status` / `standard` / `method` / `timestamp` / `facility_id` | Verification facility certificate | A |
| `policy_flags.export_control` (BIS 3A090 for H100) | US Department of Commerce export control rule | A |
| `policy_flags.compliance` (NIST 800-88, R2v3, …) | Standards mapping per asset class | A |

### Tier 1 — Whole-System Upcycle (servers only, not eligible in this lot)

| Field | Source | Tier |
|---|---|---|
| `tier1_evaluation.system_diagnostics_run` / `_result` | OEM diagnostic suite (Dell PowerEdge SupportAssist) | A |
| `tier1_evaluation.thermal_stress_test` / `burn_in_test` | SiliconX Verified Vault test bench | A |
| `tier1_evaluation.performance_validation.baseline_profile_id` / `result` | OCP Whole-System Recert profile (draft v1.0) | A |
| `tier1_evaluation.firmware_lock_status.oem_firmware_locks_present` | OEM firmware probe | A |
| `tier1_evaluation.firmware_lock_status.license_constraints` | OEM license registry (iDRAC Enterprise) | A |
| `tier1_evaluation.economic_qualification.recertification_cost_estimate_band` | OEM recert price sheet | B |
| `tier1_evaluation.economic_qualification.estimated_secondary_market_value_band` | Whole-system comp scrape (Bytestock, Network Outlet, ServerSupply) | B |
| `tier1_evaluation.economic_qualification.split_harvest_value_band` | Sum-of-parts model (component comps × bulk-discount) | B |
| `tier1_evaluation.economic_qualification.viability` | Decision rule: `not_viable_tier1` if recert cost ≥ 15% of resale OR whole-system value < split-harvest | B |
| `tier1_evaluation.failure_reasons` | Decision rule output | A |
| `tier1_evaluation.fallback_path` | Always `component_reuse` when `not_viable_tier1` | A |

### Tier 2 — Component Reuse (the 22 graded GPUs + 3 bare chassis)

| Field | Source | Tier |
|---|---|---|
| `tier2_evaluation.health_metrics.ecc_correctable_error_count` / `ecc_uncorrectable_error_count` | NVIDIA NVML / `nvidia-smi -q -d ECC` historical sum | A |
| `tier2_evaluation.health_metrics.thermal_throttle_incident_count` / `thermal_shutdown_incident_count` | Snapshot telemetry rollup (clock-gating + TDR events) | A |
| `tier2_evaluation.health_metrics.overclock_or_undervolt_flag` | NVML persistence flags | A |
| `tier2_evaluation.component_tests.electrical_test` / `functional_test` / `thermal_validation` | SiliconX Verified Vault automated test rig | A |
| `tier2_evaluation.component_tests.burn_in.duration_hours` / `result` | Standardized 8-hour burn-in (CUDA stress + memcheck) | A |
| `tier2_evaluation.grading.grade` (A/B/C/Reject) | Decision rule on health_metrics + component_tests vs. OCP-CGS-v0.9 thresholds | A |
| `tier2_evaluation.grading.grading_standard_id` | OCP Component Grading Standard v0.9 (draft) | A |
| `tier2_evaluation.grading.graded_by` / `grading_date` | Test bench audit log | A |
| `tier2_evaluation.compatibility.supported_platform_families` | NVIDIA platform compatibility matrix (H100 SXM5) | A |
| `tier2_evaluation.compatibility.required_firmware` | NVIDIA H100 GH100 VBIOS minimum version | A |
| `tier2_evaluation.compatibility.known_incompatibilities` | NVIDIA + OEM known-issues registry | A |
| `tier2_evaluation.platform_binding.secure_boot_dependency` / `bmc_auth_dependency` / `platform_whitelist_only` | OEM firmware probe (false for SXM5 modules) | A |
| `tier2_evaluation.inventory.status` / `current_location_region` / `preferred_use` / `economic_band` | Inventory management system | A |

### Tier 3 — Material Recovery (the 2 reject GPUs)

| Field | Source | Tier |
|---|---|---|
| `tier3_evaluation.material_profile.approximate_weight_kg` | Verified intake weight (calibrated scale) | A |
| `tier3_evaluation.material_profile.metals_estimate.copper_pct` / `aluminum_pct` | OEM bill-of-materials disclosure (heat spreader + chassis) | B |
| `tier3_evaluation.material_profile.metals_estimate.precious_metals_pct` | Industry yield model for GH100-class packages (Au + Pd + Pt content) | B |
| `tier3_evaluation.material_profile.metals_estimate.rare_earth_content` | HBM3 stack composition (industry-known trace yttrium / dysprosium) | B |
| `tier3_evaluation.material_profile.plastics_pct` / `other_notes` | OEM BOM disclosure + component teardown library | B |
| `tier3_evaluation.disassembly_guidance.priority` / `recommended_streams` | Recycler intake protocol (high-grade board smelter for GH100 packages) | A |
| `tier3_evaluation.disassembly_guidance.special_handling` | OSHA + OEM safety data sheets (TIM dust, HBM solder) | A |
| `tier3_evaluation.required_recycling_standards` | R2v3 + e-Stewards mapping for AI-grade silicon | A |
| `tier3_evaluation.final_data_destruction.all_data_bearing_components_processed` | Sanitization status check (false for non-data-bearing) | A |
| `tier3_evaluation.final_data_destruction.method_summary` | Always `N/A — volatile memory only` for SXM5 GPUs | A |

### Disposition (every asset)

| Field | Source | Tier |
|---|---|---|
| `disposition.target` | Decision tree: pass/fail on tier evaluations | A |
| `disposition.rationale` | Generated from failure reasons + grade | A |
| `disposition.primary_tier` | One of `T1`, `T1→T2`, `T2`, `T3`. Mutually exclusive at a point in time | A |

### Lot-level disposition rollup (`disposition_plan`)

The lot card aggregates per-asset dispositions:

- `tier_breakdown` counts assets in each tier (`T1`, `T1→T2`, `T2`, `T3`) plus
  the chassis-only Tier 2 frames.
- `value_recovery` compares **Tier 1 path** (sell 3 servers whole) vs.
  **Tier 2 path** (disassemble; sell 22 GPUs as components + 3 bare chassis +
  recycle 2 reject GPUs). The Tier 2 path is recommended because component
  market depth is significantly stronger than the whole-system buyer pool for
  retired hyperscaler hardware.
- `gating_status` tracks the four hard gates that must clear before any
  resale: server sanitization complete, GPU sanitization N/A, BIS 3A090 export
  screening committed, custody chain unbroken intake → verification.