SWCI Compute
V4 weighted · RT-1 channel reweight
Seven-component basket, weighted geometric mean rebased to the
trailing 168-day median. V3 (May 2026) added HBM stress and
frontier GPU OEM pricing so the basket reflects the real 2026 AI
buildout (B200 / GB200 / MI300X) instead of EOL A100/H100 gear alone.
| Component | Source | Weight |
| Frontier GPU OEM T1 | B200 / GB200 / MI300X channel pricing (Dell, HPE, Supermicro) | 35% |
| HBM stress T1/T3 | TrendForce ASP press deltas + Micron 10-Q HBM signal score | 25% |
| Hyperscaler GPU rental T1 | H100 + A100 per-GPU-hour median, AWS/Azure/Oracle/Hetzner | 20% |
| Legacy GPU eBay T4 | A100 40/80GB PCIe, H100 PCIe, RTX 4090, RTX 3090 | 6% |
| Server eBay T4 | Dell R750/R760, HPE DL360/DL380 Gen11 | 6% |
| Enterprise SSD eBay T4 | 3.84TB and 7.68TB U.2 NVMe median across brands | 4% |
| DDR5 RAM eBay T4 | 64GB and 96GB DDR5 RDIMM 4800 | 4% |
Rationale (V4, May 2026 — RT-1):
V3 still gave eBay (Tier 4 secondary market) 50% of SWCI. V4
caps eBay at 20% and redistributes the freed weight to Tier 1
sources — frontier GPU OEM (Dell/HPE/Supermicro direct: 35%),
HBM stress (TrendForce + Micron filings: 25%), and hyperscaler
public APIs (20%). The secondary market is still represented
but no longer dominates a supply-chain index.
Source mix: 67.5% Tier 1 · 12.5% Tier 3 · 20% Tier 4 (eBay)
YoY for SWCI is backfilled from public-source research prior to 2027-Q2. See
YoY backfill disclosure below.
Methodology consistency (RT-21, May 2026):
Both current-week and historical SWCI composites use the same V4
methodology: each component's basket value is rebased against its
own trailing window median = 100 BEFORE the weighted geometric mean.
HBM and frontier-GPU OEM components return None in historical
months (no public backfill series) — historical composites use 5 of 7
components honestly, with weight renormalization. Current composite
uses up to 7 components as the live data flows.
Six-component weighted geometric mean. InfiniBand carries the largest tied
weight (25%) because Dell’Oro tracks InfiniBand at roughly 80% of AI training
fabric share — SWII is intended to reflect AI-infrastructure interconnect
specifically. RT-15 adds direct-attach copper and active optical cables (DAC/AOC),
which often account for 30–40% of AI cluster interconnect spend. RT-16 adds
an NVSwitch/NVLink fabric component sourced from OEM rack-config channels (the
GB200 NVL72 bundles 18 NVSwitches per rack); standalone NVLink Switch System
pricing is largely RFQ-only, so this component remains thin until OEM channels
expose more pricing.
- InfiniBand FS.com new-list (30%, T2) — NDR 400G/800G, HDR 200G, XDR 800G — NVIDIA Quantum-class transceivers
- NVSwitch / NVLink OEM channel (20%, T1) — GB200 NVL72 racks, HGX H100/H200 baseboards. Weight redistributes when OEM pricing absent.
- FS.com Ethernet new-list (15%, T2) — 400G QSFP-DD DR4, 800G OSFP 2xDR4
- DAC/AOC FS.com new-list (15%, T2) — 100G/400G/800G direct-attach copper and active optical cables
- Networking eBay (15%, T4) — Cisco Catalyst 9300, 9500, Nexus 9300, Arista 7050
- Optical transceiver eBay (5%, T4) — 100G QSFP28, 400G QSFP-DD
RT-1 (May 2026):
eBay (Tier 4 secondary) reduced from 35% to 20%; weight
redistributed to channel partners and OEM direct.
Source mix: 20% Tier 1 (OEM direct) · 60% Tier 2 (channel partners) · 20% Tier 4 (eBay)
YoY for SWII is not displayed until 52 weeks of spot-desk snapshots accumulate.
SWPI Power
V2 diffusion · weights research-derived
SWPI is a five-component diffusion index. Each component is normalized to its
own trailing 24-month median (= 100) and then composited by the weights below.
Values above 100 mean the grid stack is more stressed than its trailing baseline.
| Component | Source | Weight |
| Queue depth | LBNL queue + EIA-860M | 30% |
| Backlog growth | EDGAR cohort (GEV, ETN, PWR, et al.) | 25% |
| Lead-time signal | EDGAR text-mining | 20% |
| PPI electrical | BLS 4-series blend | 15% |
| LMP | ERCOT + 6 RTOs (ISO-weighted) | 10% |
Rationale (V2, May 2026):
Queue depth, backlog growth, and lead-time signal together carry 75% of the
weight because Sightline Climate, JLL, and CBRE all identify the physical
electrical-supply layer — transformers, switchgear, batteries — as the binding
constraint on 2026 AI datacenter buildout (only 5 GW under construction against
16 GW announced). LMP is downweighted from V1's 25% to 10% because it is
volatile daily noise dominated by weather and gas, not a structural read on
AI stack stress. ISO-weighted LMP within the LMP component is ERCOT 30, PJM 25,
SPP 15, MISO 10, CAISO 10, NYISO 5, ISO-NE 5 (renormalized when an ISO is
unavailable).
GEV equipment vs RPO (RT-19, May 2026):
GE Vernova's headline backlog ($150-163B in earnings press
releases) is Remaining Performance Obligations (RPO) — includes
multi-year service contracts. SWPI's backlog component wants
equipment-only (transformers, switchgear, turbines in the
pipeline). The pipeline extracts Power + Electrification
Equipment RPO directly from GEV's 10-K segment disclosure
($55.2B for FY2025) and applies a documented 0.40 haircut to
press-release RPO figures as a proxy when only the headline is
available. Calibration: 10-K equipment ($55B) ÷ Power +
Electrification total RPO ($129B) ≈ 0.43, rounded down to 0.40
conservatively.
Full derivation, sensitivity tables, and citations:
docs/research/swpi_swfi_weights_methodology.md
.
SWFI Facility
V2 hybrid · weights research-derived
SWFI is a five-component hybrid index: two spot eBay baskets (T&M, PLC), two
BLS PPI series (process instruments, HVAC), and one EDGAR cohort diffusion
signal proxying cooling adoption from the otherwise RFQ-only liquid-cooling
market. Each component is normalized to its own trailing-window median.
| Component | Source | Weight |
| PPI process instruments T1 | BLS PCU334513334513 | 20% |
| PPI HVAC T1 | BLS PCU333415333415 | 20% |
| EDGAR cohort (cooling + AI-DC) T3 | VRT, TT, JCI, ROK, HON, DOV, EMR (sub-cohort weighted — RT-13) | 20% |
| Cooling (CDW channel) T2 | InRow, SRCOOL, Vertiv Liebert catalog | 20% |
| T&M eBay T4 | 12-week median basket | 10% |
| PLC eBay T4 | 12-week median basket | 10% |
Rationale (V3, May 2026 — RT-1):
eBay (Tier 4 secondary) reduced from 30% to 20%; PPI HVAC and
CDW cooling channel each lifted to 20%. The four Tier 1/2/3
disclosure-grade components now carry 80% of the index. JLL
projects AI fit-out costs exceeding $25M/MW, dominated by
cooling complexity — so cooling-channel pricing (Vertiv,
Liebert, InRow on CDW) and the BLS HVAC PPI carry the
cooling-stack signal alongside EDGAR sentiment.
Source mix: 40% Tier 1 (BLS) · 20% Tier 2 (CDW) · 20% Tier 3 (EDGAR) · 20% Tier 4 (eBay)
EDGAR sub-cohort split (RT-13, May 2026):
The 7-name cooling/automation cohort blends structurally
different businesses. To avoid smearing Vertiv’s direct
AI-datacenter cooling adoption signal with general industrial
noise, the EDGAR component is now a weighted blend across four
sub-cohorts:
| Sub-cohort | Constituents | Weight |
| Cooling pure-play | VRT — the AI-datacenter cooling signal | 40% |
| HVAC + building | TT, JCI — some DC cooling, mostly building HVAC | 30% |
| Industrial automation | ROK, HON — process automation, weak DC signal | 20% |
| Diversified industrial | DOV, EMR — “liquid cooling” mentions largely non-DC | 10% |
Full derivation, sensitivity tables, and citations:
docs/research/swpi_swfi_weights_methodology.md
.