Grid Standards v1.3

Every segment is tagged by what it can honestly say today. The tag is load-bearing, not cosmetic.

• [instrumented]— a direct daily series stands behind the state.
• [proxy]— an adjacent series stands in (an equity, a derivative, or a monthly-cadence release).
• [gap]— no live instrument; reported as a documented hole, never fabricated.

Packaging & Assembly is the single most binding segment in the chain and currently a [gap]. The Grid measures where pressure lands, not where it originates, and reports the hole honestly rather than inventing a reading. Networking & Interconnect is likewise structural-only: it carries no forced buyer verb until a supply-chain instrument earns one.

Each instrumented or proxied segment carries one supply-demand condition on a single axis, tightest to loosest:

These are commodity-market condition terms (Platts / Argus / LME register). Direction of movement is reported separately in the segment's delta, never folded into the state. A [gap]-covered segment carries nostate — the documented hole is shown instead of a fabricated condition. “Tight” and “Balanced” also name two of the GPI bands (§2.4); the overload is deliberate — the same condition read at segment scope and whole-market scope.

Segment states use buyer-side color inversion. Green is favorable for the compute buyer. Red is unfavorable. This is the opposite of equity market convention.

This color convention is permanent and will not change in future versions.

The GPI is composed of four weighted contributing factors, each independently normalised against its own 30-day history.

The GPU Rental Pricing Trend factor is intentionally a different series from GAP-H100 (Pillar 3, §3.1). The index factor tracks the daily median across all H100 variants (SXM, NVL, PCIe) and is retained unchanged for index-history continuity; GAP-H100 is the assessed price of the SXM class specifically, provider-pooled, defined 12 June 2026. A methodology reader should not conflate the index input with the assessed price.

Each factor is scored relative to its own 30-day mean and standard deviation (CNN Fear & Greed precedent). 50 = exactly at the 30-day mean. ±3σ maps to the 0–100 range, clamped to [5, 95]. When fewer than 2 observations exist: defaults to 50 (neutral). v1.0 used min/max scaling; the change to z-score is methodology v2.0.

Weighted sum, rounded to two decimal places, clamped to [0, 100].

The published read of whole-market state is the GPI band. The scale has five permanent bands — the names are fixed and do not change across versions:

The bands describe the composite; the per-segment states (Pillar 1, §1.3) describe individual segments. “Tight” and “Balanced” appear at both scopes — the same condition at different altitude.

The GPI follows the precedent of published composite indices: VIX (equity volatility), CNN Fear & Greed (market sentiment), PMI (manufacturing activity), BDI (shipping demand). No equivalent composite index exists for AI compute markets.

Published GPI history is immutable. Methodology changes are annotated and versioned; historical values are never silently rewritten.

One annotated exception precedes this rule: on 12 June 2026, before any external subscriber existed, the May 11 – June 10 history was recomputed once to correct two corrupt factor inputs — a misparsed DRAM series that had fed the Memory Pressure factor since launch, and unbounded GPU price rows that had contaminated the GPU factor's baseline windows. The pre-recompute series is archived. From the first external subscriber onward, the published series stands as printed.

Second annotation: on 29 June 2026 a v1.0-formula defect affecting May 11–17 was identified and corrected forward. Thirty-one days of window-drift this defect introduced are left in the published series as honest history — annotated here, not rewritten — per the immutability rule above.

Versions 1.0–1.1 published a daily market-condition word (the GMCC — clear / cloudy / rain / storm, a weather register). That vocabulary is retired from every published surfaceas of v1.2: it printed the same word for thirteen consecutive days while conditions underneath it moved, which is the opposite of signal. The whole-market read is the GPI band (§2.4); per-segment conditions use the single-axis states (§1.3). The internal classifier survives only as a computed input to the pipeline — it is never surfaced and never a headline. Historical editions that printed weather words stand as published.

GAP-H100is the assessed on-demand rental price of the NVIDIA H100 SXM class (80GB), expressed in per-GPU $/hr. The level is the median across contributing providers — one vote per provider: each provider's offers are reduced to a median first, so a marketplace listing several machine configurations cannot outweigh a provider quoting one price. Multi-GPU machine offers are normalised to per-GPU prices and tracked as distinct configuration series. Only on-demand offers contribute; spot and interruptible listings are excluded. One lineage note: contributing rows captured before 12 June 2026 predate reliable tier labelling at one source (Vast.ai recorded a fixed “spot” label while its stored prices were the on-demand rate); those rows are included as on-demand and the labels stand as recorded rather than being rewritten.

The contributing panel currently comprises two H100 SXM providers (RunPod, Vast.ai), with ThunderCompute contributing to other H100 configurations (PCIe). The panel size is disclosed because it is thin; it will widen as sources are added. H200 is not yet assessment-grade: its offer series must first demonstrate stability under the configuration-normalised methodology introduced 12 June 2026.

A published percentage change is only ever computed between observations of the same series: the same (provider, SKU configuration) pair for GPU rental, the same chip series for DRAM spot. Day-over-day and multi-day deltas are the median of per-series changes across series present at both endpoint dates (a matched-pair index) — never the difference between two daily medians of a changing offer basket. Level and change are therefore reported separately: the level describes today's panel; the change describes repriced like-for-like series. Where fewer than the minimum matched pairs exist, no change is published.

OCPI (Ornn Compute Price Index) is the exchange-traded settlement benchmark for compute, published by Ornn and distributed on Bloomberg Terminal. The Grid cites OCPI as the reference benchmark; it is never Grid-originated data.

Qualification rule: The Grid surfaces only API-sourced OCPI settlements. A newly observed settlement series is provisionaluntil seven consecutive days of API-sourced settlement exist. While provisional, the OCPI value is displayed with its observation count and the Grid Basis — the spread between OCPI settlement and GAP-H100 — is suppressed entirely: a spread against an unqualified benchmark is not informative, and The Grid does not publish it informally. The first API-sourced observation was 10 June 2026; earlier homepage-scraped values were retired from display and are retained only as collection history.

What the basis is — revised v1.3: the Grid Basis reads cleared settlement versus posted list: OCPI is executed transactions; GAP is posted rates (§3.1). A positive basis means transactions are clearing at a premium to posted rates; a negative basis, at a discount. Because the posted leg is sticky by construction, day-to-day basis movement is attributed to the leg that actually moved — ordinarily the cleared leg. The basis is not, by itself, a buyer's-market or seller's-market verdict.