Glossary

The terms and methodology behind Ornn’s indices: what each number measures and how it’s built.

Compute

Ornn Compute Price Index: Transaction-weighted price of executed GPU rentals, in USD per GPU-hour.A benchmark rental price for a given GPU, built from anonymized records of trades that actually executed across independent providers. It is transaction-weighted, so larger trades move it more, and it excludes quotes and posted list prices, so it tracks the rate the market cleared at rather than the rate providers advertised. The inputs are verifiable hourly on-demand compute prices, collected as they are transacted across providers.View methodologySee alsoGPU-hour Daily Index
GPU-hour: One GPU rented for one hour, the unit every compute price is quoted in.The pricing unit for compute. A rate of $2.50 per GPU-hour means one GPU for one hour costs $2.50; eight GPUs for three hours costs 8 × 3 × $2.50 = $60. Quoting per GPU-hour makes rentals of different sizes and durations directly comparable.
Daily Index: Daily compute reference price in USD per GPU-hour, calculated over the 24-hour window ending 4:00 PM ET.The daily reference print is the volume-weighted average of eligible GPU rental trades over the 24-hour reference window ending 4:00 PM ET. The platform’s headline number and Price History chart open on this settled daily series at every range, and the cadence pills switch them to the hourly index.See alsoHourly Index Rolling 24-hour average Ornn Compute Price Index
Hourly Index: The compute index recomputed each hour from that hour’s trades.Each hour’s value is computed from the trades transacted within that hour, so intraday moves survive instead of settling into one print. Set the cadence pills to Hourly to plot it in place of the daily reference, or read it through the API.See alsoDaily Index Rolling 24-hour average Ornn Compute Price Index
Rolling 24-hour average: Average of the hourly index over the trailing 24 hours.An alternative compute-price series available through the API. At each point it averages the hourly index over the preceding 24 hours, so a single unusually thin or busy hour can’t swing the number, and it still moves continuously as new hours settle.See alsoDaily Index Hourly Index

Memory

Memory Price Index: Daily spot price for an individual DRAM, Flash, or module memory part, in USD per unit.A daily benchmark price for a specific memory component, such as a given DRAM, Flash, or module part (RDIMM or UDIMM), reported as that trading session’s average. Memory settles once per day, so unlike compute it has no intraday series. Pricing comes from partners based directly in the regions producing, manufacturing, and distributing memory.

Tokens

Ornn Token Price Index (OTPI): Token-volume-weighted price per million tokens paid across an AI lab’s models.A daily index of what a lab’s tokens actually cost, weighted by the volume of tokens transacted across that lab’s priced models over a UTC daily window. Free variants and open-weight models sold by third parties are excluded, so it reflects tokens bought on the labs’ own paid offerings. Because the index tracks a weighted average by lab, weighted by consumption, a shift in consumption toward heavier models pushes the index up.See alsoPer million tokens
Per million tokens: The pricing unit for token indices: the cost of one million tokens.Token prices are quoted per million tokens (MTok), the standard unit labs price against, so indices for models with very different absolute rates stay comparable.

Analytics

Volatility: Annualized volatility of the index over the selected rolling window.The annualized standard deviation of the index’s log returns over the chosen rolling window, scaled to a yearly figure (√365 for daily data, √8,760 for hourly). Higher volatility means the price has been moving more sharply; it’s a measure of risk, not of direction.See alsoRolling Window Ornn Compute Price Index
Rolling Window: The look-back length over which a rolling metric is measured.The number of trailing days a rolling calculation spans: 3, 7, 15 or 30 for volatility. A shorter window reacts quickly and reads noisier; a longer window is smoother and slower to turn. It sets how much history each point summarizes, not how much of the chart you see.
Utilization Ratio: Share of tracked GPU capacity currently rented.Rented units divided by rented-plus-available units across the capacity Ornn tracks: the fraction of supply in use right now. A rising ratio means the market is tightening; combined with price, it separates demand-driven moves from supply-driven ones.
Payback Period: Years of OCPI rent, along the forward curve, needed to recover the hardware book.Start at today's OCPI and walk the published forward curve: hourly rent times utilization, accrued until it recovers the hardware street book. That is years to pay back the chip, not remaining useful life and not an NPV residual. Utilization and the curve's backwardation are inputs — occupancy stretches the path, and a steeper curve pays back later. Hardware prices are curated from public street quotes because NVIDIA does not publish a list; each row names its sources.See alsoOrnn Compute Price Index Utilization Ratio
Forward Curve: What compute costs by contract length, on either of two bases.Two related figures, and the distinction matters. The term price is the flat rate paid across a contract of a given length: a three-year deal at $1.46/hr costs $1.46 every hour for three years. The implied forward is the marginal rate the ladder prices each stretch of the term at, backed out by differencing cumulative cost between tenors. Because a term price averages in every earlier and more expensive hour, the implied forward falls faster: a ladder quoting $2.10 at one year and $1.46 at three years prices years two and three at $1.14, not $1.46. Term price is what a contract carries; implied forward is what the market thinks an hour that far out is worth. Curves are sourced from active deals in the market, assembled by our team of analysts, and updated weekly.See alsoOrnn Compute Price Index Implied Forward
Implied Forward: The marginal rate a term ladder prices each stretch of a contract at.Take the cumulative cost of each quoted term, price times months, and difference it between neighbouring tenors: the result is the flat rate implied for the stretch between them. Only the shortest tenor agrees with its own quote; every longer one is lower on a falling curve, and higher on a rising one. The strip is piecewise constant by construction, which is why it is drawn as steps rather than a smooth line. A ladder whose cumulative cost fails to rise implies a free or negative stretch and is rejected rather than published.See alsoForward Curve
Public Pull Requests: Daily count of public GitHub pull requests attributed to each AI coding tool.Counts pull requests opened in public GitHub repositories that carry an AI coding tool’s signature, tallied per day for each tool. It is a demand proxy: more pull requests means more real work being pushed through the tool. The series trims its most recent two days, since those cohorts haven’t matured yet and the right edge would otherwise always dip. Only public repositories are counted, so private-repo activity is not captured.See alsoMerge Rate
Merge Rate: Share of a tool’s public pull requests that ended up merged.For each day’s cohort of public pull requests attributed to a coding tool, the fraction that merged. It reads as a quality signal beside the raw count: a tool can open many pull requests, but the merge rate shows how much of that work was accepted. The most recent two days are trimmed, since a pull request opened yesterday hasn’t had time to merge and would drag the right edge down.See alsoPublic Pull Requests
Model Frontier: Benchmark score of each model against its cost or time per task, with the frontier replayed through time.Each point is one model at one thinking level. The y-axis is its score on the selected benchmark; the x-axis is cost per task or, when the benchmark publishes solve times, time per task. The dotted line traces the frontier: the best score at each price or duration. A vertical date rail replays observed history from the oldest release at the top to today at the bottom, at today’s published prices. Open- and closed-weights filters rebuild the cloud. Selecting a point opens a detail column with today’s performance and cost standing, then intelligence for cost over time and cost per intelligence over time.See alsoCapability Over Time Ornn Token Price Index (OTPI)
Capability Over Time: Benchmark scores by model release date, with the rising capability frontier.Plots every dated model’s best score on the selected benchmark against its release date. The solid line is the capability frontier, the running best score as of each date, with its record-setting models labelled. The dashed continuation projects that frontier forward at the pace it has been climbing, inside a band calibrated by back-testing the projection against what actually happened.See alsoModel Frontier
Compute Buyers: AI-native startups that raised $10M+ in a single event, classified as likely compute buyers.A roster of startups that raised at least $10 million in a single event, drawn from SEC Form D filings and announced rounds, then classified as likely compute buyers. Tiers separate what the compute is for: training models, heavy inference and AI infrastructure, or GPU-adjacent simulation and HPC. Read it as forward demand: each raise is capital that tends to convert into compute spending.
Cost per Request: Average dollar cost of a single API call to a lab’s models, shown in cents.Total settled spend divided by the number of paid calls for the selected lab (and model class, when one is picked), shown in cents, or cents per image for image models. It is the real, blended answer to what one call actually costs right now across everything the lab served that day, so it moves with all three drivers at once: the per-token price, the tokens each request carries, and the mix of models being called. Covered labs are the closed providers where settled ground-truth dollars are observable.See alsoLike-for-Like Index Workload Intensity Ornn Token Price Index (OTPI)
Like-for-Like Index: Cost per request with model-mix shifts stripped out, indexed to base 100.Tracks only whether the same models got cheaper or more expensive to run, holding constant which models people used, and rebased to 100 at the start. This isolates true price and efficiency change from the cheaper-looking savings that come just from routing traffic to a smaller model like Haiku or Flash. Read it against the raw series: if cost per request drops while like-for-like stays flat, costs fell only because customers switched to cheaper models, not because any model actually got cheaper. The gap between the two is the mix effect.See alsoCost per Request Workload Intensity
Workload Intensity: Average tokens carried per request, a measure of how heavy a typical call is.The average number of tokens carried per request, in thousands of tokens. It is the quantity term in the cost identity: cost per request equals price per token times tokens per request. Rising intensity means workloads are getting chattier, with longer prompts, longer outputs, and more context carried per call, which pushes cost per request up even when per-token prices are falling.See alsoCost per Request Like-for-Like Index
Token Share: Each selected lab’s daily token volume as a share of the labs selected beside it, within one weight class.One lab’s settled volume divided by the combined volume of the labs ticked alongside it that day, which is what makes it a read on who is taking a larger slice rather than simply who is growing. A lab can be growing and still losing share if the others grow faster. The denominator is the selection, so the drawn lines always sum to 100%: ticking another lab re-bases the ones already on the chart rather than squeezing a new line in beside them. Closed and open-weight labs are never mixed into one denominator — the selector picks a weight class first, and the labs inside it second. That is deliberate: these volumes settle from OpenRouter, whose traffic leans heavily toward open-weight models, so a share taken across both classes would mostly describe the routing venue rather than the market. Within a class the same skew sits in every lab’s denominator and cancels. A lab with no settled volume on a given day is left out of that day’s denominator rather than counted as zero, so the shares always describe the labs that actually reported.See alsoToken Volume Index Frontier Token Volume Open vs Closed Share Ornn Token Price Index (OTPI)
Token Volume Index: Each lab drawn, indexed to 100 on the first day all of them reported.Where OTPI prices frontier tokens, this counts them. Every selected lab is divided by its own volume on one shared day — the earliest on which all of them reported — so each starts at 100 and labs an order of magnitude apart in absolute size are directly comparable. The shared base is what makes the lines readable against each other: indexed from its own first day instead, a lab that only started reporting last month would open at 100 beside labs that had been compounding since the spring, and the gap between them would read as a difference in growth when it is only a difference in start date. The base therefore MOVES with the selection — ticking a lab with a shorter history shortens the window they all share and re-bases the chart, and dropping it gives the window back — which is why the chart names its base day rather than leaving it implicit. The window starts at that day for the same reason. The selector gates on weight class first — closed source or open source — and offers the labs inside that class; the chart opens on the closed frontier three. A lab ticked but with no settled volume anywhere in the window is struck through in the selector and named beside the base date, because it cannot be drawn and so is not part of the shared base either — the base is the first day every lab ACTUALLY DRAWN reported, which is a narrower claim than the selection. Open-weight models released by a closed lab and free variants are excluded before settlement, so a lab’s totals sit below its published all-model figures.See alsoToken Share Frontier Token Volume Open vs Closed Share Ornn Token Price Index (OTPI)
Frontier Token Volume: Combined daily token volume across OpenAI, Anthropic, and Google, indexed to 100 at the first date.The closed frontier labs—OpenAI, Anthropic, and Google—are added together and indexed to 100 on the first date any of them reported. It answers whether the frontier as a whole is growing, which is a question neither of the other two series can: shares are forced to sum to 100% every day and so cannot all rise, and a per-lab index only says how one lab moved against its own past. The total is the sum of whichever labs reported that day, so a lab dropping out shows up here as a dip rather than being filled in.See alsoToken Volume Index Token Share Open vs Closed Share
Open vs Closed Share: Each day’s settled tokens split between proprietary and open-weight models.The three closed frontier labs—OpenAI, Anthropic, and Google—against the eight open-weight providers, as shares of the tokens that settled that day. The series begins on the first day an open-weight lab settled anything, not at the start of the window: the closed labs were settling well before the open-weight ones entered the registry, and the days in between would draw a band pinned at 100% closed that then appears to collapse — a picture of when we started recording rather than of the market. Unlike Token Share, whose denominator is whichever labs you tick within one weight class, this one spans the whole registry, so the two sides always sum to 100% and the read is which kind of model the volume went to rather than which lab took it. An open-weight model released by a closed lab does not move this line: gpt-oss and Gemma are excluded at settlement, so they count for neither side. A side with no lab reporting on a given day is a genuine zero rather than a gap, since both groups settle from the same job on the same schedule.See alsoToken Share Token Volume Index Frontier Token Volume Input vs Output Share
Input vs Output Share: One lab’s settled tokens split between the prompt side and the completion side.How much of a lab’s daily token volume is text sent TO the model versus text generated BY it, as shares of that day’s total. It is a read on workload shape rather than size: a lab whose output share is climbing is being asked to generate more per prompt, which is what a shift toward reasoning or long-form work looks like, while a rising input share points to longer context being fed in. One lab at a time, and deliberately so — this is a within-lab ratio, so two labs on one axis would be two unrelated proportions sharing a scale with no combined quantity for them to be shares of. The figures come from a different source than the rest of this page: the daily settlement adds prompt and completion into one total before storing, so the split survives only in the hourly provider samples. The proportion itself is measured — it is the prompt/completion ratio the upstream activity feed reports — though the per-provider token counts those samples carry are allocated from it rather than observed directly. A day with no tokens carries no ratio and is left out rather than drawn as a zero.See alsoToken Volume Index Open vs Closed Share Ornn Token Price Index (OTPI)