Skip to main content
Every dataset and every field a subscriber receives, tagged by plan. Delivered through data.ornn.com and the REST API at https://api.ornnai.com, with CSV/XLSX export from in-app charts and tables on Full only. (The public Index page offers its own CSV of the free-window chart — that series is already public; Full-tier exports cover every dataset and the complete history.) The Index window is served on the public Index pages and the public API routes; the signed-in /analytics workspace requires an approved account, and shows locked previews to everyone else. This is version 4.0, effective August 19, 2026. It supersedes the v3 dictionary (Public vs Full Data only) by adding Premium ($500/month, daily grain) as a first-class column. Compare plans at data.ornn.com/pricing. Grain and key rules are in Access tiers. When those rules change, this page must change with them — backend/tests/data-dictionary-gating.test.js fails the PR otherwise. Ornn Data LLC, 447 Broadway FL 4, New York, NY 10013.
access on a keyed response is premium or full. On the Index window it is public-3mo (GPU and OTPI alike).

1. Datasets

Premium adds the seven OTPI labs outside the Index four, unlimited history on both windowed series (including the hourly token price), five forward curves, and sections 3–11 and 13. Full adds GPU hourly history and real-time values, hourly power operations, CSV/XLSX export, the B300 / GB200 / GB300 forward curves, forward curve history, B300’s full index history, volatility and occupancy, and the full datacenter-site payload. Trial is Premium with a 72-hour clock.

2. Compute Price Index (OCPI)

Per-GPU price series calculated from anonymized executed transactions. The index reflects transactions, not list prices. Published hourly on the hour and aggregated to a daily average. One region is populated, global. Tracked GPUs (5). All five are on Index at a trailing 3 months, daily. B300 (trailing 3 months free; depth Full / Enterprise). Anyone can read B300’s current value (/api/gpu/B300), its latest daily values (/api/daily-index?gpuType=B300, /api/day-index, /api/daily-index/B300/previous, /api/daily-index/all), and its trailing 3 months on /api/gpu/B300/index-history, with or without a key; /api/daily-index-simple serves its latest value to any key. Its full history (index-history before the trailing 3 months, /api/gpu/B300/history-range), /volatility and /volume-metrics need a Full plan; a Premium plan gets the trailing 3 months on index-history and 403 on the others. Every new key is minted sk_live_. A key that already begins with sk_prem_ keeps that prefix and keeps working. Every other GPU route refuses B300. From 2026-11-10 through 2026-11-17 the index pools hourly rentals on the Ornn marketplace with the rentals it already read, weighted by rented GPUs, and from 2026-11-18 it reads none of them. The marketplace rentals count at 20 percent of their GPUs through 2026-11-13 (18 percent on 2026-11-10), then at 16, 12, 8 and 4 percent on the four days from 2026-11-14. For part of 2026-11-13 their weight sat a fifth of the way from that 20 percent toward a fixed 20 percent of the pooled GPUs, a move that was not continued. These dates are UTC days, so a daily value, which settles on New York days, takes its last five hours from the next step.

3. Compute analytics

Volatility and occupancy derived from the GPU index history, per GPU. No Index window. Occupancy is the share of GPUs in the tracked population that are currently rented, not the share that are powered on or computing. Changes are in percentage points, not percent change: both ends are already percentages.

4. Forward Curves

Published forward marks by tenor for tracked GPUs. The data array contains only curves with at least one persisted valid mark. The suppressed array contains fully unmarked curves. Each suppressed point has a null price and a machine-readable reason. The website catalog lists the tracked curve products. Premium and Trial see B300, GB200, and GB300 as locked rows. It shows a dash for each missing mark. A missing tenor remains a gap. The service never interpolates, carries forward, or fills a missing mark with a default. The Index tier has no marks API. Tracked curves (8). H100 SXM, H200, A100 SXM4, B200, RTX PRO 6000 WS, B300, GB200, GB300. Premium / Trial (5). H100 SXM, H200, A100 SXM4, B200, RTX PRO 6000 WS. Full / Enterprise only. B300, GB200, GB300. Tenors (5). 1M (1 month), 6M (6), 1Y (12), 3Y (36), 5Y (60). History (Full / Enterprise only). GET /api/forward/history returns past published curves, one row per publish, oldest first: name, capturedAt, and marks (each tenor’s label, months, price, and has_mark). Only published curves are served; model reconstructions are not. History starts when Ornn began recording each publish, with no earlier backfill. Premium and Trial keys receive 403. Published price values are term prices: the flat reserved-contract rate of that length. The implied forward is those term contracts layered — the leftover rate on the months only the longer deal covers, after subtracting the shorter contract already priced. Each mark prices a standard contract (256 GPUs, InfiniBand, nothing paid upfront, starting now), except that the 1M mark is also anchored to the last daily on-demand spot settle, taken as it is, where the chip has one; construction, and how quotes at other sizes, prepayments and interconnects are put on that basis, is on Forward Curves.

5. Memory Price Index

Daily spot prices for 14 memory types in three categories, refreshed each weekday. No Index window on prices; the type catalog is public. Tracked memory types

6. Token Price Index (OTPI)

A daily, volume-weighted blended price per million tokens for each lab, across its paid production models. Free tiers and open-weight self-hosting are excluded. GET /api/otpi/hourly returns the price of each hourly window (windowStart to capturedAt) as indexPerMtok, with the window’s tokens and coverage, on the same labs and the same window. Rows before the windowed series begins keep the earlier method, the day-so-far price at each hourly sample, with null windowStart, tokens, and coverage. The settled daily print stays on GET /api/otpi. Tracked labs (11). The first four are Index at a trailing 3 months; the rest are Premium/Full only.

7. Token Volume Index

Settled daily token volume OTPI has already counted. Daily granularity. No Index window. Open-weight models released by a closed lab and free variants are excluded before settlement. On the platform (data.ornn.com/analytics/models/volume): all 11 OTPI labs. Token Share and Volume Growth each have an independent selector (default: the closed frontier three). Share re-bases on the selection so the drawn lines always sum to 100%. Volume Growth indexes every plotted lab from one shared day — the first on which they all reported — and trims the window to start there. Frontier Token Volume stays the closed three combined, indexed to 100 at the first date. Open vs Closed Share splits the whole registry by model-weight class (three closed against eight open-weight) as two lines on a pinned 0–100 axis. See The Token Volume Index. On the REST API at GET /api/token-volume: the closed-lab series only (anthropic, openai, google). Each lab’s growthIndex is indexed to 100 at that lab’s own first settled day, and share is of the closed-lab daily total. startDate and endDate required. The growth index is fixed-base: overlapping windows return identical values for the same date. The platform’s shared baseline, selectable share, and open/closed split are recomputed from the same settled rows; they are not extra fields on this response.

8. Workload Cost Index

The OTPI companion that prices the request rather than the token, for the three closed labs. History from 2026-06-11. No Index window. Tracked providers and model classes. Classes are exact partitions of a provider’s daily rows, so class rows sum back to the aggregate. Wire values are IEEE 754 doubles after that rounding. Near-zero LMDI terms serialize as 0. Settlement storage is unchanged.

9. Model Frontier

Benchmark capability against price per task, for every model and thinking level. Cost is priced off the index: the lab’s OTPI dollars per million tokens multiplied by the model’s tokens per task. No Index window. On the REST API at GET /api/model-frontier, which also carries releaseDates — model display name to first-release date, for reconstructing past frontier states — and history, real month-by-month prices where they exist (currently GPQA Diamond only): fractional months plus each model’s cap and monthly cost series. Capability axes: Overall (Epoch ECI), Coding (SWE-bench Verified), Reasoning (GPQA Diamond), Math (AIME 2024/25). A model with no real score on an axis is dropped from that view, never estimated. Tracked benchmarks (11). WeirdML, Aider Polyglot, CursorBench, DeepSWE, ALE-Bench, ARC-AGI, ARC-AGI-2, ProofBench, Surface Evolver, CritPt, The Agent Company.

10. LLM Coding Analytics

Daily public-GitHub activity for the major AI coding tools. No Index window on values; the product catalog is public. Tracked tools (5). claude (Claude Code), codex, copilot (GitHub Copilot), cursor, devin (Cognition). PRs and merge rates for all five; commits for claude only. Products (13). A tool crossed with a metric. Each has an id, label, category, kind (count or pct), compare flag, and seriesLabels. Public-GitHub signals only. Private repos, self-hosted instances, and IDE-side telemetry are out of scope.

11. Power Markets

Daily wholesale power prices across every US market, with history from 2017-01-03. Six markets are ICE exchange trades carrying the volume behind each print; ERCOT is its own day-ahead clearing price, shown against grid load. No Index window. Power Markets is served on the REST API at GET /api/power/markets (daily prices, traded volume, and system load) and rendered on the platform charts at data.ornn.com/analytics/power. The hourly operations routes are GET /api/power/load (actual and forecast demand) and GET /api/power/fuel-mix (generation by fuel). Daily prices are Premium+; the hourly operations routes are Full-only. There is no free/Index tier. Traded volume and system load are separate fields and never share an axis: a clearing price has no traded volume, so ERCOT carries the energy the grid actually moved. Daily peak-block prices per hub, not hourly nodal. Tracked markets The operations routes use EIA regional data rather than ICE hub trades. Load is hourly actual demand and day-ahead forecast in MW. Fuel mix is hourly regional generation by fuel in MW. Both are operational context for the price series, not additional hub-level traded-price products, and both require Full access.

12. Datacenter Sites

Every tracked AI datacenter site worldwide, with the operator, power capacity, hardware, and build timeline behind each one. Full only for the payload. Index and Premium see coordinates on the map (points only) — not a locked overlay. Filters, site records, and the power overlay stay Full-only. Sites can be rolled up by country or by GPU type into site_count, total_power_mw, and total_gpu_qty.

13. Compute Buyers

AI-native startups that raised 10 million dollars or more in a single event, classified by the workload their compute is for. Built from SEC Form D filings and announced rounds, July 2024 to July 2026. No Index window. On the REST API at GET /api/compute-buyers.

Keys, payments, and grain

The account’s current plan decides what a key can see. Every new key is printed sk_live_. A key that already begins with sk_prem_ is left as it is and keeps working. Once the account’s plan changes, the prefix is not the gate. A key that has not been through a plan change since this rule shipped keeps the older rule: a sk_prem_ key, or one minted on Premium, is Premium; any other is Full. A plan change updates that key; it does not revoke it. Manual revoke and delete still do. See Access tiers. When API gating or Premium billing changes, update this page, Access tiers, and regenerate the internal Word file from frontend/app/(shell)/internal/data-dictionary/build-docx.py.