SpotWire

Client: Internal • Industry: Market Data / AI Infrastructure • Completed: June 11, 2026
Python Market Data Data Provenance AI Infrastructure Index Methodology Data Engineering

Challenge

The AI buildout is a physical-layer story — GPUs, transceivers, transformers, grid interconnection, cooling — but price discovery across that stack still runs on stale spreadsheets, broker gut feel, and OEM list prices that bear little relation to what equipment actually clears at. Hyperscalers know what they pay; everyone else quotes off numbers that aren’t transacted. There is no Platts for AI infrastructure, and the bottleneck keeps migrating — GPUs in 2023, transformers in 2024, now interconnection queues and 800G optics. The edge is seeing scarcity build across the stack before the crowd does, with a number you can trace back to a trustworthy source.

Solution

Built SpotWire — a local-first data engine that turns the noisy surplus, channel, and public-infrastructure market into clean, citable price signal and publishes it as four flagship weekly indices at spotwire.ai.

  • Four flagship indices — SWCI (compute: legacy/frontier GPUs, servers, SSD, RAM, HBM, hyperscaler rental), SWII (interconnect: networking, transceivers, InfiniBand, DAC/AOC, NVSwitch), SWPI (power: nationwide LMP, interconnection queue depth, OEM backlog growth, lead-time, BLS PPI), and SWFI (facility: T&M, PLC, process/HVAC PPI, EDGAR cohort sentiment, cooling channel) — each a weighted geometric mean of per-component baskets, with every component rebased to its own trailing-window median so the composite is consistent within and across time
  • Tiered source ledger (T1–T5) ranking every input from OEM-direct, hyperscaler API, and BLS PPI (T1) down to SEC EDGAR filing extracts (T3) and eBay completed transactions (T4, capped at ≤20% index weight), so confidence is explicit rather than implied
  • Price-primitive disclosure — every component is tagged with the kind of price it carries (transacted vs. OEM-list vs. channel-list vs. cloud-quoted vs. stress-proxy); a list price is not a transacted price, and the published payload says which is which
  • Automatic integrity guards — single-source and vendor-concentration warnings re-flag an index whenever one feed dies or one vendor exceeds a weight threshold, and indices carry an explicit indicative vs. live status (SWCI and SWFI are live; SWII and SWPI remain indicative)
  • Versioned, reproducible methodology — a dated changelog, published per-component liquidity counts, and red-team logs make every historical index value reproducible; methodology changes are versioned and prior values are never silently recomputed

Impact

  • Publishes a cross-stack scarcity view spanning compute, interconnect, power, and facility — a signal no single-vertical industry report assembles, showing where the bottleneck is migrating in real time
  • Most recent issue (No. 23, as of 2026-06-11) composited from ~50 sources, 36 verticals, and ~1,378 weekly completed comps across ~192 tracked SKUs over a 730-day window
  • Every published number is traceable to a source trust tier and a versioned methodology, holding the indices to an audit-ready standard rather than a black box
  • Clean export contracts (indices.json, methodology.json, per-index history CSVs, frozen monthly archive snapshots) ship downstream to the public site, released under CC-BY-4.0 with stable citation permalinks
  • Powers a weekly newsletter and the public index at spotwire.ai, turning fragmented infrastructure data into equal-access market signal
  • Methodology discipline is the product: when SWCI’s composite once read implausibly low, the fix went into the math layer (per-component rebasing) rather than a cosmetic patch — rigorous over shortcut, by default