Frequently Asked Questions

Ornn builds financial infrastructure for the GPU compute economy. We provide reference pricing, hedging instruments, and market data so institutions can transact compute the way they transact any other commodity.

Accessing compute is the bottleneck for frontier AI, and multi-year compute contracts can cost hundreds of millions of dollars. Without a benchmark price there is no way to mark a contract to market, hedge exposure, or write standard terms. Ornn supplies the benchmark and the market structure around it.

Three audiences. Institutional traders and operators hedging compute exposure, buyers and lenders that need a defensible reference price, and analysts pricing compute against the market rather than a single vendor’s rate card.

The ORNN Compute Price Index family (GPU rental prices), the ORNN Memory Price Index (DRAM and Flash parts), the Ornn Token Price Index (OTPI, the realized price of frontier-lab tokens), forward curves, analytics covering AI demand, compute buyers, and workload cost, a datacenter map, and a REST API. Free reference data is published at index.ornn.com and the full platform lives at data.ornn.com.

Yes. The Ornn white paper, the formal Index Methodology document, and a data dictionary are available, and the API is documented at data.ornn.com/docs. Institutional diligence teams can also request our standard due-diligence pack.

Ornn operates as a price reporting agency. Each index print is a volume-weighted average of executed transactions, typically hundreds of eligible data points per print, drawn from anonymized trades across a verified network of independent providers. Inputs pass wash-trade screens, minimum-notional thresholds, and counterparty verification, and outliers are clipped before the average is taken.

Because listed prices are not prices. A posted rate can sit far above where deals actually clear, and offers can be changed at will with no buyer on the other side. Executed, paid transactions are the strongest basis for a benchmark: the same reasoning that moved reference rates from submission-based to transaction-based regimes. Influencing a transaction-based index requires committing real capital at market prices.

From partnered cloud providers, datacenter operators, owners, and lessors. When compute is bought, the transaction is recorded at the price actually charged, so the inputs are verifiable, invoice-level records of executed trades rather than survey submissions. Aggregating this data is a function of Ornn’s business of tracking and transacting compute capacity globally.

On the order of 150 providers contribute to each index and roughly a thousand transactions are observed per day, with the majority of volume in the United States. Coverage grows as new providers are verified and integrated.

All eligible trades pool into one global, volume-weighted distribution, so a region’s influence on the print is its share of executed volume. Every trade stays tagged with its region, GPU type, and quantity, which means regional cuts such as US-only series can be produced, and regional basis shows up as dispersion around the global print.

If no eligible volume prints in a window, the index holds the last published value rather than interpolating. Because inputs are pooled globally across many providers, fully empty windows are rare and become rarer as coverage grows.

The methodology follows the IOSCO Principles for financial benchmarks: arm’s-length executed transactions only, defined eligibility criteria, outlier handling, and documented procedures. Moving the index would require committing capital in real transactions at market prices, and contributed data is screened for wash trades and self-dealing.

Settled prints are final by construction: each period is computed once, after its inputs have closed. If a confirmed data error is identified, the affected prints are corrected and the correction is noted, but history is never silently restated and methodology changes are not applied retroactively.

Earlier prints drew on fewer contributing sources, and smaller samples produce noisier averages. As the provider network has grown, print-to-print noise has fallen. The underlying market itself is also genuinely volatile: unsold compute hours expire worthless, which creates real price movement when demand shifts.

A family of transaction-based indices tracking clearing prices for rented GPU compute in USD per GPU-hour. It uses anonymized executed trades from a diversified set of independent operators. No offers or indicative quotes are included.

H100 SXM, H200, A100 SXM4, B200, RTX 5090, and RTX PRO 6000 WS. New GPU types are added as transaction volume on them crosses the eligibility threshold.

The index publishes hourly, and the headline number on the platform is a 24-hour rolling average of the hourly series. A daily reference print, defined as the volume-weighted average over the 24-hour window ending 4:00 PM ET, is available for products that settle on a single daily value.

H100 history runs from January 2025 at hourly granularity, with sparser data back to mid-2024. Newer SKUs begin when their volume crossed eligibility, for example B200 from November 2025. Earlier, sparser history can be shared on request for diligence and backtesting.

• Aggregate demand from new model architectures and enterprise AI workloads
• Effective supply tied to hardware delivery and power availability
• The perishable nature of compute: unsold hours are lost, which pressures prices down until a large buyer enters
• Geopolitics and export controls
• Efficiency gains that migrate workloads across hardware generations

A rate card is one provider’s posted price for its own capacity. The index is a market-wide settlement built from executed trades across many providers, regions, and contract tenors, and it is the same number every counterparty sees.

The index measures the executed on-demand market: the marginal clearing price of compute. Long-dated reserved contracts trade at negotiated levels around that benchmark, the same way term power or freight contracts trade around spot indices. That is precisely what makes the index useful as a mark for such contracts, and region- and tenor-tagged data supports basis analysis.

Yes. The H100 index is available to Bloomberg Terminal subscribers under the ORNNH100 ticker family.

The published indices are per-SKU, which keeps each series homogeneous. A FLOPS-weighted composite is in research; the open questions are how to weight new hardware vintages that carry high theoretical throughput but limited deployment. Per-vintage indices first, composite second.

It is a daily benchmark price for individual DRAM and Flash parts, reported as that trading session’s average. Pricing comes from partners based directly in the regions producing, manufacturing, and distributing memory.

SKU-level DRAM parts across DDR4 and DDR5 densities, plus Flash components, grouped into categories on the platform. Coverage follows the parts that trade with enough liquidity to produce a defensible daily print.

Memory settles once per trading session at the source: the market itself produces daily prints, not a continuous tape. An intraday memory series is under evaluation with our data partners, but we will not publish a cadence the underlying market cannot support.

The Ornn Token Price Index is a family of transaction-based indices that track the realized market price of frontier-lab language-model tokens in USD per million tokens, written $/MTok. For each lab, the index is the token-volume-weighted average price actually paid across that lab’s models over a daily UTC window. It is built from same-day transacted prices, provider weights pooled across the day’s intraday routing samples, and same-day token volumes, all drawn from executed on-demand inference traffic. Each day is settled once and never restated.

They price the two sides of the AI economy. The Compute Price Index measures the market’s input, the price of GPU-hours, and the Token Price Index measures its output, the price of tokens. One tells you what it costs to rent the compute; the other tells you what the market actually pays per token of model output once caching, routing, and model mix are counted.

The index is organized one index per lab and currently tracks token usage across OpenAI, Anthropic, and Google models. Each lab index covers that lab’s priced models as listed on the on-demand token exchanges, with new models entering as they list and begin transacting. Open-source models are being evaluated but remain outside the published index for now.

Free variants and zero-price models are excluded because they carry no transaction value. Open-weight models such as gpt-oss and Gemma are excluded because their tokens are sold by third-party hosts rather than by the lab, and the index measures tokens transacted on the labs’ own priced offerings. Volumes count paid variants only.

Both are already inside the transacted prices. Cache discounts are embedded in the input price, so fresh input, cache reads, and cache writes blend in the input term. Reasoning tokens bill within completion at the output price. Because the index is built from executed traffic, these cost classes register exactly as buyers incur them.

Each provider is weighted by its share of the day’s token throughput, built by summing hourly throughput samples across the full UTC day rather than reading a single instant. Pooling across the closed day removes the noise in instantaneous routing snapshots.

Daily, on the exchanges’ UTC day boundary. A day closes at 00:00 UTC and settles about 12 hours after close, once upstream counters have finalized. Only fully closed days are computed, and each day is computed once under a single same-day regime of prices, weights, and volumes.

New model launches and their adoption, the mix of demand across premium and budget tiers, cache-discount economics, the input and output composition of real workloads, reasoning intensity pushing volume toward higher-priced output tokens, provider routing shifts, and lab-level pricing changes as they appear in executed traffic.

Two design choices make it hard to move. Every input is executed, paid inference, so influencing the index means committing capital at market prices. And the volume weighting uses complete daily token totals, on the order of hundreds of billions to trillions of tokens across millions of requests, so it carries no sampling error. Settling only after counters finalize means published history is stable by construction.

It is published on the Ornn Data platform at data.ornn.com, with limited free access through index.ornn.com. For a licensed or higher-volume feed, contact the team.

The curves are sourced from active deals in the market and assembled by our team of analysts, then updated weekly. They express the price the market implies for compute delivered at future dates, plotted against tenor, and serve as the forward-looking companion to the spot indices.

Buyers go long at a forward tenor matched to their training or inference plan. Operators sell forward to lock in revenue, and lenders and insurers use the index as a mark-to-market reference. Compute exposure is tradeable today through our exchange partners, with listed futures expanding access.

Venues settle against the published series that fits their product: the hourly index, the daily reference print (the volume-weighted 24-hour window ending 4:00 PM ET), or period averages for monthly contracts. The observation window and settlement time are defined in each product’s specification, always against published, never-restated prints.

Yes, physically settled structures are in development alongside the financially settled products. If you have a delivery-based use case, we want to hear it.

Yes. The platform exposes a REST API returning JSON, documented at data.ornn.com/docs, and the platform itself supports CSV and Excel export from every chart.

Current prices for the tracked GPUs are public. The free tier covers five GPU series with three months of history and four token-index labs with one month of history, at 60 requests per minute per IP. Full history, memory data, analytics, and the remaining series require an API key.

Sign up at data.ornn.com and create a key from the API keys page. Commercial-scale feeds, higher rate limits, and additional endpoints are provisioned under a data agreement.

The compute index publishes hourly today, and higher-frequency delivery is available for partners whose products need it. Every subscriber receives each print at the same time: there is no early-access tier and no latency hierarchy.

Yes, the history endpoints accept date ranges, and deeper lookbacks are available on licensed tiers. For diligence and backtesting we also fulfill one-off historical extracts while platform backfill catches up.

Yes. We run a standard three-day trial with full platform access, usually after a short conversation about your use case. External distribution of the data is prohibited during a trial.

Yes. Ornn licenses its indices for derivatives, structured products, oracles, analytics, and redistribution. Licensing runs through a standard agreement scoped to your product.

A standard subscription covers internal use within your organization. Redistribution, republication, or embedding the data in products delivered to third parties requires a commercial license with attribution.

Yes. Licensed products reference the index by name, and licensing agreements set out the attribution requirements.

Ornn Data LLC, which holds the index intellectual property and is wholly owned by Ornn AI Inc. It is separate from Ornn’s regulated trading entities, which keeps the benchmark business independent of the trading business.

Yes. The standard pack includes a completed DDQ, the Index Methodology document, a data dictionary, and a curated data sample. It is routinely used by alternative-data procurement and exchange diligence teams.

Licenses are generally non-exclusive. Certain product categories on certain venues carry scoped exclusivity under specific agreements, which never restricts your ability to license the data for other uses.