The Grid is an independent evidence and valuation layer for real compute commitments — the daily record of the compute market. It holds the connected map of how compute gets made and bought, records it legibly every day with the methodology public, and reads any quote against it through one question: what is this commitment worth, can it be delivered, and what risk is embedded in it. The buyer's lens is the organising frame — and it gives traders, analysts and lenders the demand-side telemetry the supply-side research houses structurally lack.
Twenty contributing sources are collected automatically every day at 04:30 AEST: GPU rental pricing, AI provider status pages, token pricing, semiconductor earnings, DRAM spot markets, power interconnection queues, supply-chain and tech press, and social-signal volume (Hacker News; X and Reddit collection was retired 24 July 2026).
Everything is organised into one fixed comprehension spine — seven supply-chain segments, the same order every day: Fabrication · Packaging & Assembly · Memory · Power & Cooling · Networking & Interconnect · Compute Supply · Demand & Deployment. The reader learns the seven once and never re-learns them. Each segment is tagged by what it can honestly say today: [instrumented] (a direct daily series), [proxy] (an adjacent stand-in), or [gap](no live instrument — reported as a documented hole, never faked). Packaging & Assembly is the most binding segment in the chain and currently a [gap]: that honesty is the point. The Grid measures where pressure lands, not where it originates — and says so.
From the daily observations the register derives its instruments: GAP-H100, the assessed price of an H100 GPU-hour from bookable asks, with GAP-H100-R1Y, its one-year committed counterpart; GTS-1Y, the premium for locking a year ahead of riding spot; Grid Basis, the exchange-cleared print read against the assessed price; GCCXand the divergence oscillator, a 29-ticker compute-equity composite held against the physical layer — the read that dated the July dislocation in real time; RWI, the daily cost of a fixed reference inference workload by its cheapest route, and the frontier-vs-open spread beside it; and The Ledger, dated third-party forward claims recorded when made and checked when due. The Grid does not forecast. It records who did, and checks.
The signal is the supply chain. Upstream structural conditions — HBM allocation, CoWoS and substrate lead times, wafer pricing, sold-out notes, export-control events — are the input. Equity and price moves are confirmation and consequence, never the lead: a move proves an upstream read landed. When a shift traces to a named business — ARM, Micron, SK Hynix, Ajinomoto — naming it is the product working correctly. Read first, name second. The Grid takes no investment positions and makes no price calls on equities.
The GPI is a 0–100 composite score measuring aggregate buyer-side pressure across four weighted contributing factors: GPU Rental Trend (30%), Memory Pressure (25%), Supply Chain Forward (30%), and Provider Incidents (15%). 50 = the 30-day mean; above 50, pressure is tightening.
Three factors are z-score normalised against their own 30-day history (Supply Chain Forward enters as a bounded 0–100 level composite), with severity-weighted incident scoring (critical ×4, major ×3, minor ×1, maintenance ×0.5). The full methodology — including the band scale (Abundant / Balanced / Tight / Constrained / Critical) — is publicly documented. Anyone with the same data pipeline can calculate the same GPI. The moat is the pipeline, not the formula.
Each issue is built for a reader deciding something. It opens with The Decision— what today's conditions mean, in plain language — then Buyer Posture, the standing verb for each segment; The Move, the day's figures; The Analyst's Note, the traced cause behind them; Signals, the dated arrivals ledger — only what newly printed, never narrative; and The Assessment, all seven supply-chain segments in one scan. One judgment per day, carried until the data changes it.
Every data claim carries inline source attribution and an observation tag: OBS (directly observed from data), AST (assessed from multiple signals), or EST (estimated with stated uncertainty). The editorial observes; it does not forecast. Forecast generation is suspended (identity decision of record, 24 July 2026): the remaining indicator series verify privately until a record exists, and the public accountability surface is the dated call archive.
Grid Standards defines the vocabulary The Grid uses — the GPI methodology and band scale, Grid Assessed Prices (GAP) by GPU class, the OCPI settlement benchmark (attributed to Ornn), the seven-segment supply-chain taxonomy, observation classifications, and temporal-horizon formats. Published under CC BY 4.0. Free, open, citable.
The market problem in one line: $2 trillion in cumulative AI compute expenditure by 2028, with no structured daily buyer-side intelligence about the market being purchased in.
The Grid is built by Jason Harvey. Manufacturing and supply chain operations background (COO, ModularWalls). The intuition behind The Grid comes from industrial commodity markets — LME warehouse warrants, stockyard yarding numbers, Baltic Dry Index — applied to AI compute.
Based in Sydney, Australia. The daily briefing publishes at 04:30 AEST — first read of the Asia-Pacific trading day.
Contact: briefing@thegridco.ai