On November 23 Platts will start publishing a daily price for renting four Nvidia chips, on demand and interruptible. It joins Silicon Data, Ornn, Compute Desk, SemiAnalysis, Setara and CF Benchmarks. Three of those firms also publish forward curves, and for weeks the market has argued in public about what the shape of the curve means.
This piece asks five questions about that crowd of numbers. What do on-demand, interruptible and reserved prices each measure? What does the slope of a curve mean when no forward trades? Why is everyone indexing hourly data? What separates two indices that seem to price the same thing? And which one would you hedge with today?
Much of the evidence comes from Ornn Data. Ornn gave Kinetic Alpha a full research license, so this piece could look further into its data than into any other provider’s: the hourly index built from executed rentals, the reserved-contract marks out to five years, and the measure of utilization. That reflects the access Ornn gave. It does not mean other providers hold less. Each figure that uses Ornn’s data says so. The calculations on that data are Kinetic Alpha’s.
- 1. On-demand, interruptible and reserved
- They are three products with three sets of buyers. Reserved contracts carry most of the money and almost none of the published data. On-demand and interruptible prices cover what is left after reserved customers are served, and they carry almost all of the published data. No provider publishes traded volume by region.
- 2. Contango or backwardation
- The curves are built from reserved contracts of different lengths that all start today. Nothing trades for delivery in a future month. The two words describe a comparison between two points that each vendor picks. In early November the same H100 numbers supported both words.
- 3. The turn to hourly data
- Hourly on-demand and interruptible prices are the easiest data to collect, and perpetual futures and event contracts need a price that ticks. The hedge value is small today: about a tenth of an hourly H100 move lasts a day. The information value is real. The gap between on-demand and interruptible is the best public reading of spare capacity.
- 4. What separates the indices
- Five things: what counts as a price, what is left inside it, how chips are made comparable, where, and how often. In the first week of November, two one-year rates for the H100 sat about 10% apart, one day apart.
- 5. The best hedge
- Compute Desk’s one-year reserved benchmark, administered by GX Benchmarks, is the closest match to the exposure the market carries. Three things are not yet shown: how many deals stand behind a daily print, which series the Nodal contract will settle on, and when anything lists.
A daily assessment of what is on offer
Platts, the price reporting arm of S&P Global Energy, told subscribers on November 5 that it will launch US GPU Compute prices on November 23.[1] Platts wrote the details down, so they are worth reading closely.
- Four chips. The H100, H200, B200 and B300.
- Two products. On-demand and interruptible.
- One assessment a day. The price is an hourly rental rate in dollars per GPU-hour. The frequency is daily, based on “prices and availability at 1:30 pm CT.”
- Listed prices, averaged. Each price is “a simple average of the prices of available GPUs.” Prices more than two median absolute deviations from the median are dropped.
- A token view. The same prices are also quoted in dollars per million tokens of output, with a tokens-per-second figure for each chip.
- United States only. There is no reserved price.
Platts prices what is listed for rent. Most of the field starts from the same raw material. Platts differs in saying what comes next. It may adopt “a volume-weighted average approach once sufficient trade data has been collected,” and it asks market participants to send it trades. Platts built its energy assessments the same way: start from visible prices, then earn the trade reports. Those assessments are now written into physical supply contracts.
Its arrival means the business of pricing compute now has an incumbent in it. It does not change the gap in the data. The product that carries most of the money still has no Platts price.
On-demand and interruptible are what is left over
A GPU is rented in three ways. Each has its own buyers, its own way of setting a price and its own trail of data.
| What the buyer gets | Who uses it | How index providers source it | Where anyone can see it | What it tells you | |
|---|---|---|---|---|---|
| On-demand | A chip by the hour. Start and stop at will. The provider cannot take it back. | Small teams, experiments, fine-tuning runs, overflow inference. | Listed rates from provider pages and marketplace feeds (Silicon Data, Platts, Setara, CF Benchmarks). Executed rentals reported by contributing providers (Ornn). | Provider price pages. Marketplaces and aggregators such as Vast.ai, RunPod, Shadeform and Akash. SF Compute’s order book shows executed fills. | How scarce the leftover capacity is this hour. |
| Interruptible | The same chip, cheaper, until the provider wants it back. | Batch jobs and training runs that save their progress and can restart. | Listed spot or preemptible tiers. On open marketplaces, the lowest bid a host will accept. Platts will assess it daily. | The same pages. Fewer providers list it. | What an hour is worth when nobody reserved it and nobody needs it now. |
| Reserved | A fixed rate for six months to five years. Part is paid up front. The buyer pays whether it uses the chips or not. | Labs, enterprises and inference companies with steady load. | Completed deals and quotes on a deal platform (Compute Desk). Monthly surveys (SemiAnalysis). Observed contract rates and published rates (Silicon Data). Marks published at five lengths (Ornn). | Mostly nowhere. A few providers post a long-term rate. | What production workloads cost, and what lenders and operators depend on. |
The table explains a pattern in the data. The prices that are easy to collect belong to the smallest part of the market. Compute Desk wrote in May that “the short end of the rental market has effectively disappeared”: talks for premium chips no longer start below six months, and buyers prepay about a fifth of the contract.[12] SemiAnalysis, which surveys neoclouds and buyers every month, calls the contract market the one that “better reflects where durable AI workloads are actually going.”[17]
So read an on-demand index as a gauge of spare capacity. When reserved customers take nearly everything, the few chips left for hourly rent clear at a high price and the index jumps. Ornn’s B200 on-demand index rose from $5.03 on May 1 to $9.83 on October 28.[4] That measures how scarce the leftovers became. It is not what a lab pays for a thousand chips for a year.
View as table
| Type | Source and date | $ per GPU-hour |
|---|---|---|
| On-demand, listed | GPUPerHour median, as seen Nov 6 | 2.90 |
| On-demand, listed | AIMultiple median, November | 3.92 |
| On-demand, listed | Price of Compute, as seen Nov 6 | 3.99 |
| On-demand, listed | Marketplace rate cards, 45 offers from 9 providers, Oct 29 | 1.95 to 4.41, median 3.99 |
| On-demand, listed | Silicon Data H100 index, Nov 10 | 3.06 |
| On-demand, executed | Ornn Compute Price Index, Nov 10; range Oct 11 to Nov 10 | 3.10; 2.67 to 3.53 |
| Interruptible | Vast.ai bid floors, two H100 hosts, Oct 29 | 2.00 and 2.67 |
| Reserved, 1 month | Ornn mark, Nov 6 | 2.70 |
| Reserved, 6 months | Ornn mark, Nov 6; Silicon Data term rate, Nov 10 | 2.79; 2.95 |
| Reserved, 1 year | Ornn mark, Nov 6; Compute Desk as quoted by Forward Compute, Nov 5; SemiAnalysis midpoint of deals, Oct 1 | 2.48; 2.74; about 2.80 |
| Reserved, 3 and 5 years | Ornn marks, Nov 6 | 1.94; 1.71 |
Who shows the volume
Whether an index deserves trust depends on how much traded behind it, and where. Nobody publishes that.
- Ornn weights each rental by the number of GPUs in it and records the region of every trade. It publishes one global number per chip and does not name its contributors.[2] The volume figure it does publish is a utilization ratio: the share of tracked capacity that is rented. For the H100 that ratio ran between 83% and 97% from October 11 to November 10.[4]
- Silicon Data says it processes “over 10,000 GPU rental records” a day from more than 40 providers. Provider weights reflect “market relevance.” The document does not publish them.[6]
- Platts will use a simple average with no volume weights at launch.[1]
- Compute Desk builds from deals on its own platform and does not publish how many.[13]
- SemiAnalysis surveys more than 100 participants once a month.[17]
What an outsider can count is thin. On October 29 our own collector found 45 listed H100 offers from nine providers through one aggregator, priced from $1.95 to $4.41, and two H100 hosts on the largest open marketplace.[21]
Region is the weakest point. Platts is US only and Compute Desk describes its family indices as US. Silicon Data adjusts for country inside its model. Ornn’s site map counts about 470,000 GPUs at US neocloud sites, 56% of the chips it can count worldwide.[4] That shows where chips sit. No provider discloses where the trades behind a print took place.
Two names for one set of numbers
In oil, a forward curve is a list of prices at which the same barrel trades for delivery in different months. If later months cost more, the curve is in contango. If they cost less, it is in backwardation. Storage links the months. A trader can buy oil today, hold it and deliver it later, so the gap between months cannot drift far from the cost of holding.
Compute has neither part. Nothing trades for delivery in a future month, and an unused GPU-hour cannot be stored.[18] What exists is a set of reserved contracts of different lengths that all start today. A one-year rate is an average over the next twelve months. It also pays the buyer for committing, for prepaying, for taking size, and for the risk that it will not use the chips.
Three firms turn those contracts into a curve, in three ways.
| Firm | Starts from | How the forward rate is derived | The commitment discount | How often |
|---|---|---|---|---|
| Silicon Data | Reserved contract rates it observes, plus published rates, for terms of 1 to 36 months. | A no-arbitrage formula applied to the term rates. Its chief executive compares the method to the Treasury curve. Two academics who rebuilt it describe a continuous version of the subtraction used by Ornn.[8][18] | Stays in. The whole gap between short and long rates is read as what the market expects. | Every business day.[5] |
| Ornn | Reserved quotes at five lengths: 1 month, 6 months, 1, 3 and 5 years. Gaps are left as gaps. | Subtract the shorter contract’s total cost from the longer one’s. Divide by the extra months. | Stays in, and the documentation says so.[3] | When a mark is published. There is no fixed schedule. |
| Compute Desk | Reserved deals on its own platform. | An estimate with the commitment discount taken out first. The method is not published in detail.[11] | Removed. | Not stated. |
Figure 2 puts two of the three on one chart for the H100.
| Table 3. Same ladder, two readings | On-demand | 1 month | 6 months | 1 month to 6 months | On-demand to 6 months |
|---|---|---|---|---|---|
| Silicon Data, H100, Nov 10 | $3.06 | $2.84 | $2.95 | +3.9% rising | −3.6% falling |
| Ornn, H100, Nov 6 | $3.28 | $2.70 | $2.79 | +3.3% rising | −14.9% falling |
| Ornn, B200, Nov 6 | $9.53 | $5.20 | $5.02 | −3.5% falling | −47.3% falling |
Compare one month with six months and both vendors show a rising H100 curve. That is the comparison behind a widely shared post on November 9, which said Ornn’s H100 curve had moved into contango for the first time.[10] Compare today’s on-demand price with six months and both vendors show a falling curve. Both statements are true of the same numbers. The name depends on which point you call the front. Beyond six months Ornn’s marks still fall, to $2.48 at one year and $1.71 at five.
The front matters because the on-demand price and the reserved ladder come from different deals. On Ornn’s data for November 6, a B200 rented on demand at $9.53 an hour and reserved for one month at $5.20.[4] A one-month commitment does not halve the price of the same hour. My reading is that the two numbers come from different sellers, deal sizes and buyers. A slope drawn from one to the other measures the distance between two markets.
What the arithmetic does to a ladder
The usual way to get a forward rate is to subtract. Take the total cost of a three-year contract. Subtract the total cost of a one-year contract. Divide by the extra 24 months. Ornn’s documentation describes this sum and states its limit: the curve “does not take financing, prepayment, volume, or other commitment discounts out of the reserved quotes,” and the implied forward “inherits them.”[3]
View as table
| B200 | 1 month | 6 months | 1 year | 3 years | 5 years |
|---|---|---|---|---|---|
| Reserved mark, $ per GPU-hour | 5.20 | 5.02 | 4.49 | 4.15 | 3.29 |
| Implied rate for the added stretch | 5.20 | 4.98 | 3.96 | 3.98 | 2.00 |
The marks fall gently, from $5.20 to $3.29. The rate they imply for years four and five is $2.00, about 21% of the on-demand index. Compute Desk’s David Lopez Mateos reports the same effect on his firm’s B200 deals: the subtraction “prices year-3 delivery at 21% of today’s rate.” His conclusion is that “compute cannot be stored, so no arbitrage ties the forward to spot, and the curve has to be estimated with the discount separated out.” With the discount removed, his firm puts the B200 a year out “at or above spot.”[11]
So three careful firms can look at one chip and publish a curve that falls, a curve its publisher calls rising, and a curve that is flat. The difference is what each does with the commitment discount.
What a slope can still tell you
- Change over time inside one vendor’s series. In July Silicon Data put the H100 36-month rate about 13% below spot.[7] By late October its chief executive said buyers were “not getting enough discount for longer term contract.”[8] The method did not change between those dates. Sellers stopped paying buyers to commit. That is real information about who holds the bargaining power.
- The gap between on-demand and the shortest reserved rate. It widens when the hourly market runs short of chips.
- The age of the chip. A five-year rate on an H100 prices a 2022 design in 2031. Part of the fall is honest.
A slope cannot tell you whether the market is bullish or bearish. This week’s argument on that point borrowed its logic from stored commodities, where contango means there is surplus to put in tanks.[9] Compute has no tanks.
Ask four things of any claim about the curve. Which two points are being compared? Is the front point on-demand or reserved? Has the commitment discount been taken out? When was each point last marked? On Ornn’s November 6 curve the H100 and B200 marks are labelled as computed. The H200 marks are labelled as entered by hand, and two of them carry an October 29 date.[4]
Why everyone indexes the hour, and what it is worth
Four reasons explain the rush to hourly on-demand and interruptible data.
- Nobody has to give it to you. Listed prices sit on public pages and marketplace feeds. Reserved prices sit in contracts.
- Some products need a price that ticks. A perpetual future trades around the clock and needs an index that does too. Event contracts resolve at a set time. The Ornn future that ICE has specified settles on a monthly average built from hourly prints.[16]
- Frequency is how data is sold. Ornn’s daily series costs $500 a month. The hourly series needs its top tier.[3]
- It is where a newcomer can start. Platts begins with listed prices and asks for trades.
The test
Does the hourly data carry something to hedge? Ornn’s hourly index allows a direct test. If every hourly move were new information, the moves would add up, and the change over a day would be about 4.9 times the typical change over an hour. If hourly moves are mostly noise, they cancel.
View as table
| August 1 to October 28, 2026 | H100 | H200 | B200 | A100 |
|---|---|---|---|---|
| Typical hourly move | 6.0% | 3.1% | 0.9% | 1.7% |
| Share still there after a day | 11% | 20% | 44% | 27% |
| Share still there after a week | 2% | 7% | 52% | 9% |
| Typical gap between a day’s high and low | 18% | 11% | 3% | 6% |
For the H100 they mostly cancel. The typical hourly move from August through October was 6.0%. The typical move over a day was 9.7%, which is 1.6 times the hourly figure where 4.9 would mean pure information. In variance terms, 11% of an hourly move is still there a day later and 2% a week later. A rise in one hour is followed by a fall in the next more often than chance would give (a correlation of minus 0.33).
There is also no daily shape. Averaged by hour of the day, the H100 index stays within about 1.2% of its daily mean. The typical gap between a day’s high and low is 18%. Power has hourly products because demand peaks at the same time every day and someone has to serve it. The hourly price of compute moves, but not on a schedule.
My reading is that the hourly H100 index moves mainly because the mix of trades changes from hour to hour: a cheap host in one hour, a dearer provider in the next. The B200 behaves differently. Its hourly moves are small, 0.9%, and about half of each one lasts. The same method gives a noisy series for one chip and a steady one for another.
The two on-demand indices also disagree on how much the H100 moves at all. Silicon Data reports annual volatility of 26% for its H100 index over September 2024 to August 2026. Ornn’s own 30-day measure for its H100 index was 118% on November 9.[4][22] The windows differ, and so do the inputs: one smooths listed prices across providers, the other follows executed trades.
A contract that settles on a monthly average removes the hourly noise by design. That is the right design. It also means the hourly data adds little to the hedge.
| Price | Who carries the exposure in size | Does the index match their bill | Is the movement signal at the hedge horizon | Verdict |
|---|---|---|---|---|
| Hourly on-demand | Few. Marketplace hosts and pay-by-the-hour users. Most users can stop renting when the price rises. | Good for a marketplace host. Poor for anyone on a contract. | Weak for the H100: 11% of an hourly move lasts a day. | Useful information. Little to hedge. |
| Interruptible | Almost nobody has to buy it. Sellers earn it on idle chips. | It is the seller’s price for an hour nobody reserved. | A direct reading of spare capacity, which no one else publishes. | The best gauge of slack. Not a hedge. |
| One-year reserved | Neoclouds at renewal. Lenders sizing loans. Labs and enterprises signing next year’s capacity. | Close, if the index prices the same region and deal size. | Yes. It moves slowly and keeps going: up about 40% in five months to spring 2026, and from $1.70 to about $2.80 in the year to October.[12][17] | The exposure worth hedging. |
Is it a sign of where indices are going?
Partly. The interruptible price is the nearest thing compute has to a real-time balancing price. The gap between it and the on-demand price is the best public reading of spare capacity. Platts will publish both prices daily for four chips, so anyone can read the gap each day.
The gap will not be one number. On October 29 we compared the two prices machine by machine on one marketplace, Vast.ai. Each of 35 machines carried an on-demand price and a lower interruptible price. On the typical machine the interruptible price was 74% of the on-demand price. The widest gap was 17% and the narrowest 98%.[21] A tracker of posted rates across many providers finds a wider typical gap, with interruptible at about 54% of on-demand.[20]
Hourly prices become a hedge when someone has to buy or sell by the hour. That takes a dispatcher: software that moves work between chips, providers and contract types as prices change. Every Seat but the Dispatcher found that nobody holds that seat yet. Until someone does, hourly and interruptible indices are good for watching the market and thin for hedging it. The firms that publish them are competing for the benchmark seat with the data they can get.
Five things that separate two indices of the same price
| Provider | What it prices | Built from | Where | How often | Who runs the calculation | Where it settles |
|---|---|---|---|---|---|---|
| Silicon Data | On-demand rent by chip. Term rates and implied forwards to 36 months. | Over 10,000 rental records a day from 40+ providers, adjusted to a standard contract. | Global, adjusted by country. The CME contracts use US series. | Daily | Silicon Data | CME futures on the H100 and B200, under review to January 8, 2027. |
| Ornn | On-demand rent by chip, executed. Reserved marks at five lengths. | Executed rentals from contributing providers, weighted by GPU count, with extreme prices capped. | Global | Hourly and daily | Ornn Data. An oversight committee is being set up. | Event contracts, live. ICE futures specified, no date. |
| Compute Desk | One-year reserved rent for the Hopper and Blackwell families and the H100. | Completed deals and live quotes from its deal platform, converted to performance units. | United States | Daily, 4:30 pm New York | GX Benchmarks, a regulated UK administrator. | Nodal futures planned for this year. Architect’s exchange-for-physical network. |
| Platts | On-demand and interruptible rent, four chips. Also per million tokens. | A simple average of listed prices for available GPUs. | United States | Daily from November 23 | Platts | Nothing announced. |
| SemiAnalysis | One-year H100 contract rate, public. Other chips and terms for clients. | A monthly survey of 100+ neoclouds and buyers, checked against deals. | Not stated | Monthly | SemiAnalysis | Nothing announced. |
| Setara, CF Benchmarks | On-demand list prices. | Public price lists and provider feeds. The CF Benchmarks method is a draft. | Not stated | Daily (Setara). Hourly (CF Benchmarks draft). | Each firm | Nothing announced. |
When two series claim the same price, check five things.
- What counts as a price. A listed rate, an executed rental, a completed reserved deal or a survey answer.
- What is left inside it. Prepayment, deal size and credit terms stay in Ornn’s reserved marks. Compute Desk says deal size alone moves its B200 rates by 25% to 30%.[11]
- How chips are made comparable. Silicon Data adjusts each record to a standard contract by rental type, country, processor platform and chip variant.[6] Compute Desk converts chips to performance units within a family.[11] Ornn verifies the hardware and makes no price adjustment.[2]
- Where. One global number, or the United States.
- How often, and how fresh. Hourly, daily, monthly, or when marked.
Take the one-year H100 rate in the first week of November. Ornn marked it at $2.48 on November 6.[4] Compute Desk’s one-year H100 stood at $2.74 on November 5, as quoted in Forward Compute’s daily market note.[23] The two sit about 10% apart, one day apart. Silicon Data shows a six-month rate of $2.95 for November 10 and keeps the longer terms for subscribers.[5] SemiAnalysis put the midpoint of one-year deals at about $2.80 on October 1.[17] That was five weeks earlier in a moving market, so it is background and not a like-for-like comparison. Each of the five points is a design choice, and each choice moves the number.
The indices also agree more on old chips than on new ones. On November 10 the two on-demand indices for the H100 were four cents apart: $3.10 at Ornn and $3.06 at Silicon Data. For the B200 they were $2.93 apart: $9.90 and $6.97.[4][22] A newer chip has fewer sellers, so the method decides more of the level.
A contract on one chip covers little of a mixed fleet. From May 1 to October 28 the on-demand index rose 95% for the B200, 61% for the H200 and 30% for the H100, and slipped 4% for the A100. That is the case for a family index, which Compute Desk publishes for Hopper and Blackwell.
Compute Desk is the closest fit, on three conditions
A hedge should move with the bill it protects. Most of the money in this market is paid under reserved contracts of a year or more. The risk is the price at which the next contract gets signed. Against that exposure, Compute Desk’s benchmark is the closest fit, for four reasons.
- It prices the right contract. Compute Desk describes every print in its US family indices as “a 1-year reserved contract, performance-normalized.”[11] A third-party market note quotes its H100 series as a one-year rate.[23]
- It starts from deals. The inputs are completed transactions and live quotes from Compute Desk’s own deal platform.[11][13]
- Someone else does the sums. GX Benchmarks, a regulated UK benchmark administrator, calculates and publishes the series each US business day at 4:30 pm New York.[13]
- It has a route to clearing and to delivery. Nodal Exchange plans cash-settled futures on it, cleared beside its power contracts. Architect’s exchange-for-physical network is built on the same benchmarks.[14]
Three conditions stand between that design and a working hedge.
- Depth. The number of deals behind a day’s print is not published. One platform’s deal flow is a narrow base, however good the administrator.
- The settlement series. Nodal’s release says its futures “will financially settle against Compute Desk’s daily, blended GPU price indexes.” Compute Desk describes a one-year reserved rate. Bloomberg’s report of the launch described spot prices.[14] No public document defines “blended.” Until the contract terms are public, a hedge cannot be sized.
- A listing. Nodal said its contracts are “scheduled to be available this year, subject to regulatory compliance.” The regulator’s review of CME’s contracts now runs to January 8, 2027. NYMEX made its consent to that extension conditional on other registrants “not self-certifying, obtaining Commission approval or otherwise listing comparable contracts” in the meantime. If one does, the extension ends early.[15]
| Whose risk | The exposure | Closest index today | Why |
|---|---|---|---|
| A neocloud with contracts rolling off | The rate on its next one-year deal | Compute Desk | It prices that contract, in the United States. |
| A lender to a GPU fleet | Income after the first contract ends | Compute Desk | Loan cover depends on renewal rates. Resale value needs a separate series that nothing settles on yet. |
| A lab or enterprise signing next year’s capacity | The rate on its next reserved deal | Compute Desk | The same contract, seen from the other side. |
| A host selling by the hour on a marketplace | Hourly on-demand income | Ornn | It is built from executed hourly rentals. |
| A trader who needs a listed, cleared contract | Any liquid compute price | Whichever lists first | CME’s contracts on Silicon Data are the furthest along. |
| An event-contract trader | The short-dated on-demand price | Ornn | Live event contracts already resolve on it. |
Two limits apply to every row. A future on a one-year rate protects the next contract. It does nothing for contracts already signed. And no index covers whether the chips are rented at all. Utilization is still the largest risk in the business that nothing hedges.
Five things to watch
- Listed prices give way to trades. Platts has said how it means to get there. Watch its weekly report from December 4, and watch for the first provider that publishes the number of trades behind a print.
- The one-year rate gets a public reference. The marker is a reserved contract whose price resets against a named benchmark.
- Curves stay models until a strip trades. Until then, treat a curve as its publisher’s view and ask the four questions in section 03. The first listed strip will replace three methods with one price.
- The hour finds its user. If software starts to route work by price, hourly and interruptible series become the dispatch signal. They gain the hedge value they lack today.
- Rent connects to tokens. Platts will quote each chip in dollars per million tokens. That is the unit an enterprise is billed in, and a hedge for the end user has to be built there.
What this rests on
Ornn data. The Ornn Compute Price Index values, hourly series, forward curve marks, utilization figures and site counts in this piece come from Ornn Data (data.ornn.com) under a full research license that Ornn issued generously, and are published with Ornn’s permission. Ornn is the source of that data. The averages, percentages, implied rates and variance statistics built on it are Kinetic Alpha’s calculations and are not Ornn’s figures.
- S&P Global Energy, Platts, “Platts to launch US GPU Compute Prices, Nov. 23,” subscriber note, November 5, 2026. spglobal.com
- Ornn Data, “OCPI Methodology,” first published July 24, 2026. data.ornn.com
- Ornn Data documentation: “Forward Curves,” “The Price Index” and “Access tiers,” retrieved November 11, 2026. data.ornn.com
- Ornn Data: Ornn Compute Price Index hourly and daily values, forward curve marks with their provenance labels, utilization and neocloud site counts, retrieved under a full research license on October 29 and November 11, 2026. data.ornn.com
- Silicon Data, “GPU Term Rental Implied Forward Rates,” product page with H100 term and implied forward rates for November 10, 2026. silicondata.com
- Silicon Data, “GPU Rental Index Methodology Overview,” updated August 27, 2026. silicondata.com
- Carmen Li, “The GPU Forward Curve: Pricing the Curve, Not Just the Spot,” Silicon Data, May 2026; Silicon Data, “GPU Futures: How Compute Is Becoming a Tradable Commodity,” July 2026. silicondata.com
- Fireside Alpha, video clip of Carmen Li posted on X, October 25, 2026, as transcribed by FourWeekMBA. fourweekmba.com
- Luke Kawa, “Nvidia hits a honeypothole,” Sherwood News, November 9, 2026. sherwood.news
- “Nvidia H100 Rental Prices Shift to Contango for First Time,” HuggingNews, November 10, 2026, summarising a post on X of November 9 about Ornn’s forward curve. huggingnews.com
- David Lopez Mateos, Compute Desk, LinkedIn posts on forward-curve construction and on the GX Hopper US and GX Blackwell US indices, September and October 2026. linkedin.com
- David Lopez Mateos, “There is no GPU price hedge,” Compute Desk, May 5, 2026. computedesk.substack.com
- General Index, “Compute Desk’s indexes are now regulated,” dateline August 13, 2026. general-index.com
- Nodal Exchange and Compute Desk, “Nodal Exchange and Compute Desk to launch Compute Futures Contracts,” September 3, 2026; Bloomberg, “Nodal Exchange Planning AI Compute Futures to Rival CME, ICE,” September 3, 2026. compute-desk.com
- Commodity Futures Trading Commission filings on NYMEX’s Silicon Data H100 and B200 Rental Index Futures: staff letter of September 21, 2026 extending the review to November 9, and NYMEX letter of November 4, 2026 consenting to a further extension to January 8, 2027. cftc.gov
- ICE Futures US, Ornn Compute Price Index H100 Future, product page and exchange notice of September 29, 2026. ice.com
- SemiAnalysis, “The Great GPU Shortage: Rental Capacity,” April 2, 2026, and its thread on spot and contract prices; Dylan Patel at Modal Runtime, October 1, 2026, as reported by FourWeekMBA. newsletter.semianalysis.com
- Federico Bandi and Yinan Su, “(Early) AI Compute Asset Pricing,” 2026. arxiv.org
- Setara Financial, “Setara Announces the Release of Its Initial GPU Price Index,” November 3, 2026; CF Benchmarks, “CF GPU Rental Series Methodology Guide,” version 1.0, July 28, 2026, marked as a draft. prnewswire.com
- Public price trackers, as seen on November 6, 2026: AIMultiple Cloud GPU Rental Price Index (H100 median and spot discount); Price of Compute; GPUPerHour. aimultiple.com
- Kinetic Alpha market collector, snapshot of October 29, 2026, with hourly medians for October 27 and 28. Listed offers from the Vast.ai public feed and from Shadeform.
- Silicon Data, B200 Rental Price Index page, value for November 10, 2026; Yuhua Yu and Steve Hou, “Cash-Settled Compute Futures: The First Hedge for AI Compute,” Silicon Data, August 24, 2026, for index volatility. silicondata.com
- Forward Compute, “Compute Daily Tape: the AI trade unwinds, rents hold,” November 6, 2026, quoting Compute Desk’s one-year H100 rate. forwardcompute.substack.com
Assumptions and limits
- Ornn’s data gets closer study here than any other provider’s because of the access Ornn gave. What the piece finds in Ornn’s hourly series and reserved marks may hold just as much for data that other providers do not open up.
- Ornn’s index values for November 6, 9 and 10 are its published daily values. The daily values in Figure 6 and the statistics in section 04 are calculated from Ornn’s hourly series and can differ from its published daily values by a few cents.
- Ornn’s reserved marks are dated November 6 and Silicon Data’s term rates November 10. Figure 2 and Table 3 compare them across four days.
- The hourly statistics cover August 1 to October 28, 2026. They describe that window and may not hold in another.
- The collector snapshot is one hour on October 29, 2026. Interruptible prices on Vast.ai are bid floors set by renters and hosts, read from its public feed.
- Public tracker figures are as seen on the dates shown. Those pages change daily.
- The SemiAnalysis one-year figure is from a talk on October 1. Compute Desk’s statements are from its public posts. Its one-year H100 level is as quoted in Forward Compute’s daily note, not read from Compute Desk directly.
- The view that the on-demand index and the reserved marks come from different sellers and deal sizes is an inference from the two methods. Ornn does not publish its contributors.
- The firms named were not contacted. Where any of them publishes more than is found here, the omission is ours and will be corrected in place.