On November 18, Bloomberg reported that Kalshi is working with Compute Desk to launch futures on the cost of AI computing power [2]. We welcome the move. This piece explains what we would change, and why.
Accounts of the report describe three steps, each subject to regulatory approval [2]:
- Event markets. Kalshi would bring back its prediction markets on GPU rental prices. It removed them in August amid scrutiny from the Commodity Futures Trading Commission (CFTC).
- Perpetual futures that track Ornn’s benchmark. A perpetual is a futures contract with no expiry date.
- Dated futures that track Compute Desk’s indexes.
Nothing is listed yet. The plan comes from people familiar with it, not from a formal announcement, and the CFTC is not expected to decide on compute contracts before early 2027 [2].
There is a lot to like. Kalshi already runs regulated perpetuals: its crypto contracts traded more than $1 billion in their first week [1]. Compute Desk’s indexes are calculated by an independent benchmark administrator [4]. And another serious venue beside CME, ICE and Nodal is good for everyone who needs to hedge compute.
We make three points:
- Kalshi’s white paper misreads parts of the power market. The errors are easy to fix. They matter because the paper builds a product plan on them.
- Ornn and Compute Desk measure different prices. Each one fits a different contract.
- A perpetual on Compute Desk’s indexes would let a fund hold compute without rolling futures. That is the bigger prize.
The power section needs a second pass
In October, Kalshi published Compute, Power and the Rise of AI Infrastructure Markets [1]. It argues that compute and power are one market with two legs. Four parts of that argument do not match the data.
The paper matters as a guide to Kalshi’s thinking. Its first author, Udesh Jha, is the Kalshi chief risk officer quoted on the futures plan [2]. It also gets several things right. It ranks traded prices above indicative quotes when building a benchmark. It says margin offsets should be set conservatively. And its plan for hourly power contracts that trade at all hours addresses a real cost: in earlier work we found that a data center’s true cost of power can run up to 9.4% above a flat-block hedge [14].
1. Power is a much smaller share of the price than the paper shows
The paper’s illustrative breakdown gives power and cooling about 16% of the hourly price of a GPU [1]. Our measurement is 3.5% to 6.5% [6].
Here is the arithmetic. An H100 uses about 1.66 kilowatt-hours for each GPU-hour, including cooling. Dominion Hub power averaged $55 to $72 per megawatt-hour in 2026, depending on the months counted. At those prices the electricity costs 9 to 12 cents. The same hour rents for about $2.60. Add capacity and delivery charges and the share reaches about 6.5% at most.
For power to be 16% of that hour, delivered electricity would have to cost about $250 per megawatt-hour. That is roughly four times the Dominion Hub price. Kalshi labels its figure as illustrative, and it may include cooling equipment as well as electricity. Even so, the gap is wide.
A cost share is also not a risk share. Even a correct cost figure would not tell us whether the two prices move together. They do not. Compute rents follow the supply of chips, not the power bill. Month-to-month changes in compute prices against Dominion Hub power show a correlation of −0.19 on eight observations, and +0.07 against ERCOT West [6]. Neither is a usable link. In our margin study we called the two prices orthogonal: they move independently of each other.
2. The capacity price is not the price Kalshi plans to list
The paper cites the PJM capacity auction, which cleared at its cap of $333.44 per megawatt-day, and lists “rising capacity costs” as the price behavior of the PJM market [1]. It then proposes power contracts on hourly and daily-average prices [1]. These are two different products.
| Energy | Capacity | |
|---|---|---|
| What it pays for | Electricity delivered in each hour | A promise that power plants will be available when demand peaks |
| How the price is set | Day-ahead and real-time markets, every hour, at each location | A main auction for each delivery year, run by the grid operator |
| The 2027/28 result | Not yet known. It will be set hour by hour. | $333.44 per megawatt-day for June 2027 to May 2028, set in December 2025 [7] |
| Was the price discovered? | Yes | No. It stopped at a cap agreed with regulators. One estimate puts the uncapped price near $530 [8]. |
| How a data center pays | By the megawatt-hours it uses | By its share of demand at system peaks, called its peak load contribution [9] |
| Can an hourly or daily contract hedge it? | Yes | No |
The $333.44 figure is a ceiling, not a traded price. It is fixed for a whole year, well in advance, by an auction that reflects the physical state of the grid. Nothing about it changes hour by hour. So an hourly Kalshi contract could not hedge it, and the auction result says little about where such a contract would settle.
The amount is also small for a GPU operator. Spread over every hour of the day, $333.44 per megawatt-day is $13.89 per megawatt-hour. That is about 2 cents on an H100 hour, or 0.9% of its price.
Capacity prices are good evidence that the grid is tight. They are not evidence for the product on offer.
3. The “Silicon Spark Spread” has one large leg and one small one
The paper defines a spread: the price of compute, minus the power needed to produce it. It compares this to the spark spread of a gas power plant and says the spread puts a market price on a data center’s operating margin [1]. It does not put a number on the power term in that formula.
With the number filled in, the spread is nearly all compute. The power leg is 9 to 12 cents against $2.60. More than 95% of the “spread” is the compute price. A site with 10,000 H100s draws 16.6 megawatts. Hedging both legs for a month means about $18.5 million of compute against $0.87 million of power: a ratio of 21 to 1 [6].
The comparison with a gas plant also breaks down on behavior. A gas plant’s fuel is most of its running cost, and the plant switches off when the spread turns negative. That choice is what a spark spread prices. A GPU operator does not switch off when power is expensive, because power is too small a share to matter. The formula is useful bookkeeping. It is not a trading relationship.
4. A margin offset between power and compute is diversification, not a hedge
The paper lists power against compute among positions that offset one another. To its credit, it adds that any such offset should be supported by data and set conservatively [1].
The data support only a small one. In our reconstruction of a clearing house margin model, a book holding both legs needed 13% to 24% less margin than the two legs margined separately. All of that comes from the legs moving independently. None comes from one leg hedging the other. The combined requirement landed within about 1% of the compute margin alone [6].
Kalshi wants a leading place in a market where CME, ICE and Nodal are already at work. Nodal clears one of the deepest books of local power futures in the United States. These firms and their customers know the difference between energy and capacity. Kalshi’s product plan needs the same footing.
None of this is hard to fix. And one more gap in the paper leads to the rest of this piece: it never separates on-demand rental prices from reserved ones, and it names only one benchmark, Ornn’s [1]. Which price a contract tracks is the most important design choice in the whole plan.
Two indices that measure two different prices
GPUs rent in two main ways. An on-demand rental is paid by the hour and can stop at any time. A reserved rental commits the buyer for a fixed term, often a year, at an agreed price. Ornn’s index tracks the first. Compute Desk’s indexes track the second.
| Ornn Compute Price Index | Compute Desk indexes | |
|---|---|---|
| The price it tracks | On-demand rental, by the hour | One-year reserved rental |
| What goes in | Completed on-demand rentals between verified parties, weighted by volume, with extreme prices capped [5] | Completed transactions, blended with list prices from specialist GPU clouds. List prices from the largest cloud providers are left out [4] |
| Smoothing | None from one hour to the next [5] | Blended and smoothed to limit the effect of any single observation [4] |
| Region | Global | United States |
| When it publishes | Every hour and every day, including weekends | Daily, on North American business days |
| Who calculates it | Ornn Data. Its methodology page says an oversight committee is being set up [5] | General Index, an independent benchmark administrator that works under United Kingdom benchmark rules [4] |
| Chips | H100, H200, B200, B300, A100 and others | H100, H200, B200, B300 |
| Where it is used or planned | Kalshi’s event contracts from March to August 2026. ICE has published futures specifications [1] [3] | Planned futures at Nodal, Architect and Liquid Compute. Kalshi’s dated futures, as reported [2] [3] |
View as table
| Measure | H100 | H200 | B200 |
|---|---|---|---|
| Ornn on-demand, August 20 | $2.88 | $5.41 | $6.94 |
| Ornn on-demand, November 18 | $2.93 | $7.11 | $9.21 |
| Ornn lowest and highest in the period | $2.46 – $3.53 | $4.23 – $7.75 | $6.00 – $9.94 |
| Compute Desk one-year, August 20 (implied) | $2.59 | $3.29 | $4.87 |
| Compute Desk one-year, November 18 | $2.90 | $3.52 | $6.06 |
| Measure | H100 | H200 | B200 |
|---|---|---|---|
| Ornn on-demand: typical one-day move (median) | 3.5% | 3.7% | 0.8% |
| Ornn on-demand: days with a move above 5% | 32 of 90 | 34 of 90 | 3 of 90 |
| Ornn on-demand: largest one-day move | 12.4% | 14.7% | 8.9% |
| Ornn on-demand: largest two-day move | 19.5% | 26.2% | 11.4% |
| Compute Desk one-year: one-day change on November 18 | 0.0% | +1.4% | +2.0% |
| Compute Desk one-year: daily history | Not public | ||
| On-demand price above the one-year price, November 18 | 1% | 102% | 52% |
| Ornn on-demand: change over 90 days | +1.7% | +31.4% | +32.7% |
| Compute Desk one-year: change over 90 days | +12.0% | +7.0% | +24.4% |
Two things stand out.
- The levels differ, sometimes by a lot. On November 18 a B200 rented on demand at $9.21 on Ornn’s index. Compute Desk’s one-year price was $6.06. For the H200 the on-demand price was about double the one-year price. For the H100 the two were almost the same.
- The on-demand price moves a lot from day to day. The H100 index moved more than 5% on 32 of 90 days. Its largest one-day move was 12.4%, and its largest two-day move was 19.5%. Much of that movement reverses. In our last review, only 11% of an hourly H100 move was still there a day later [14].
Stable over what horizon? A day or less
When we call an index stable, we mean over one day or a shorter period. That is the horizon that sets margin on futures and swaps.
- Initial margin on a future covers a one- or two-day loss. Nodal Clear, for example, uses one day for customer accounts and at least two days for house accounts [6].
- Gains and losses are settled in cash every day. A larger daily move means a larger daily cash call.
- Swaps that are not cleared use a longer horizon, ten days under the standard rules. Their margin still starts from the size of the daily move.
So the size of the worst daily move decides how much collateral every position ties up. In our September study, the average one-day loss in the worst 1% of cases was 10% to 13% on Ornn’s transaction-weighted record. An index with the lower volatility that Silicon Data discloses came out near 4% [6]. Margin scales with those figures. The noisier index needs about three times the collateral for the same position.
A change over 90 days matters far less for margin. An index can rise 24% in a quarter and still be cheap to margin if it gets there in small daily steps.
We expect Compute Desk’s indexes to be steadier day to day. One-year contracts reprice slowly, and the indexes are blended and smoothed [4]. The one-day changes on its website for November 18 ran from 0.0% to 2.0%. But its daily history is not public, so this is an argument from design and not yet a measurement. Publishing that history would settle the point.
Neither index is wrong. They answer different questions. Ornn answers: what does a GPU-hour cost right now? Compute Desk answers: at what price are firms committing for a year?
Match the index to the contract
Each of Kalshi’s three contract types asks something different of its index. Our pairing differs from the reported plan in one row.
| Contract | What it needs from an index | Reported plan | Our pairing | Why |
|---|---|---|---|---|
| Event market Days to weeks. Pays out on a yes-or-no question. | A fresh number every day, weekends included. Real day-to-day movement. A long public history. | Not stated. Ornn settled the earlier markets. | Ornn | Short-term views are views on the on-demand price. |
| Perpetual future No expiry. Regular funding payments tie it to the index. | Low noise, because noise turns into funding cost. A price that long-term holders want to own. Independent calculation. | Ornn | Compute Desk | A steadier, term-based index keeps funding fair and suits funds. |
| Dated future Settles on a set month. | The price that hedgers have actually contracted at. | Compute Desk | Compute Desk | Buyers, operators and lenders are exposed to term prices. |
Event markets: Ornn
An event market asks a short question, such as “will the H100 index settle above $3.00 on Friday?” It needs a number that publishes every day and that really moves. Ornn’s index does both. It is built from completed rentals, it publishes hourly, and its history goes back to 2023. Ornn also gives researchers the most complete access to its data, which is why the daily figures in this piece are Ornn’s.
The day-to-day movement that makes an on-demand index hard to hedge with is the reason people trade an event market. The on-demand price also carries real information. It shows how much spare capacity exists right now, and that is what a short-term trader has a view on.
Dated futures: Compute Desk
The reported plan already has this right. Most compute is sold on term contracts, and that is where lenders’ exposure sits [3]. A buyer, an operator or a lender needs a future that settles on the price of those contracts. Compute Desk made the same case to the CFTC: fixed-term contracts line up with delivery dates and with the loans behind many compute purchases [10].
Perpetuals: also Compute Desk
This is where we differ from the reported plan. We see five reasons.
- Funding is only as good as the index. A perpetual stays close to its index through funding payments. When the contract trades above the index, holders of long positions pay holders of short positions, and the reverse. If the index jumps 5% and falls back the next day, the contract will not follow it. The gap between them swings, and so does funding. Holders end up paying for index noise. An index that moved more than 5% on a third of days would make that a regular cost.
- Daily noise also raises margin. Margin is sized to the one- or two-day move. A noisier index ties up more collateral on both sides of every position, and it makes daily cash calls larger. That cost falls on hedgers, market makers and funds alike.
- One venue should have one reference price. If the perpetual tracks Ornn and the dated future tracks Compute Desk, the two contracts on the same chip settle to prices that were 1% to 102% apart on November 18. Nothing forces that gap to close. A market maker could not hedge one contract with the other, and the clearing house could not safely offset them. On a shared index, the two contracts become one connected market.
- The one-year price is the price worth holding. A perpetual on a one-year index gives constant exposure to the one-year price of compute. That is the price at which real commitments are made. We return to this in the next section.
- Independent calculation matters more when a contract never expires. A dated future tests its index once, at settlement. A perpetual tests it every day through funding. Compute Desk’s indexes are calculated and published by an outside administrator under United Kingdom benchmark rules [4].
Compute Desk itself is cautious about perpetuals for hedgers. In its letter to the CFTC it said perpetual compute futures would have limited appeal to natural buyers of compute, though it did not dismiss the format [10]. We agree. The natural user of a compute perpetual is an investor, not a hedger. That is the subject of the next section.
A business-day index under a contract that trades all week. Compute Desk publishes on business days. Kalshi’s plan stresses trading at all hours. The funding rule must say what happens over a weekend.
Smoothing makes an index partly predictable. Traders will see where a smoothed index is heading before it gets there. The contract price will lead the index, and the funding rule must not punish that. Smoothing also means small daily moves can run in one direction for several days. Margin for a longer horizon should be measured directly, not scaled up from the one-day figure.
Depth is not published. Compute Desk does not say how many transactions sit behind each daily value or how long its smoothing window is. A perpetual would put more weight on both.
A way for funds to hold compute without the roll
An exchange-traded fund (ETF) that tracks a commodity usually holds dated futures. Those futures expire, so the fund must keep selling the contract that is about to expire and buying a later one. This is called the roll. A perpetual removes it.
What the roll costs
The roll is a trade the fund must make on a schedule that everyone knows. In a deep market, the cost is small. In a new, thin market, other traders can position ahead of the fund, and the cost can be large. Oil and volatility funds have shown how much damage a roll can do when prices for later months sit well away from today’s price.
Compute futures will start thin. And compute has a feature that makes dated contracts harder still for a fund: each chip’s rental price is expected to fall as newer chips arrive. Kalshi’s own paper shows a B200 curve that slopes from about $7.20 to about $5.50 through 2027 [1]. A fund in dated futures has to choose which months to hold and trade through that slope again and again.
What a perpetual changes
A fund that holds a perpetual holds one contract and never rolls. The cost of holding does not vanish. It moves into the funding payment. As one ETF analyst put it, “Roll economics are replaced by funding economics” [11].
That is still a real gain. Funding is paid in small amounts and is visible in one published number. There is no roll date to trade against. And if the market expects a chip’s price to fall, the perpetual should trade a little below the index, so the fund receives funding as payment for that expected fall.
The same analyst draws a useful line. Perpetuals suit an index that cannot be delivered, such as a volatility index. They add less for a physical commodity like oil, where expiry months serve producers who deliver [11]. Compute sits on the index side of that line. A GPU-hour cannot be stored, and these contracts settle in cash.
Why Compute Desk’s indexes suit a fund
- They include completed transactions. A fund’s share price is checked against its index every day. Auditors and fund boards want an index built on deals that happened.
- They are designed to be steady from day to day. Every jump in the index that later reverses is a tracking problem for a fund. A fund also posts margin on its position. Smaller daily moves mean less collateral set aside and smaller daily cash calls.
- The one-year reserved price reflects real hedging. A firm that reserves GPUs for a year has locked in its price. That reservation is a hedge. An index of those prices shows the level at which the market is actually committing money. An investor who wants exposure to “the price of compute” most likely means that price, not this hour’s spot rate.
- An outside administrator calculates them. Listing rules and fund sponsors look for this.
Chip indexes and a composite
Single-chip perpetuals are the right place to start. They are simple and they match the indexes that exist today. But a single chip is a wasting asset. An H100 fund would track a product that the market expects to lose value each year.
A broad fund needs a composite index: one number that blends several chips and brings in new ones by rule, the way a stock index adds and removes companies. Then the fund holds compute, not one chip. Compute Desk publishes four chip indexes and no composite yet. It already adjusts prices to a common performance basis [4], which is the building block a composite needs. Kalshi’s paper places any blended index after single-chip perpetuals [1]. We would move it up.
The tool below shows how a simple composite of the four published indexes would have behaved. It is our illustration, not a Compute Desk product.
Weights
With equal weights, the composite rose 14.3%. It sits between the steadiest chip and the fastest one, which is the point of a composite.
What has to happen first
- The CFTC must approve compute perpetuals. A decision on compute contracts is expected in early 2027 [2]. The agency adopted a policy statement on perpetuals in May 2026 [13], and CME has challenged in court whether perpetuals are futures at all [3].
- The securities regulator must accept a fund that holds them. The Securities and Exchange Commission opened a consultation on novel ETFs in June 2026 [12]. No fund holding perpetuals has been named in the sources we read.
- The funding rule must be published and tested against a business-day, smoothed index.
- The contract needs depth. A fund can only be as large as the market it trades in.
What we would change
- Keep Ornn for the event markets. It is the right index for short-dated questions about the on-demand price.
- Use Compute Desk for both the perpetuals and the dated futures. One venue, one reference price, one connected market.
- Ask for a composite index early, with a published rule for adding and retiring chips. That is what a broad fund needs.
- Publish the funding design before listing. Say how funding works over weekends and against a smoothed index.
- Rewrite the power section of the white paper. Separate energy from capacity. State how much power a GPU-hour uses. Treat power and compute as two products that share a customer, not as one spread.
- For Compute Desk: publish the daily history and more about depth. The daily values, the number of transactions behind each one and the smoothing window would let users and clearing houses measure day-to-day stability for themselves.
What this piece rests on
Sources
- Udesh Jha and Pranav Jha, Compute, Power and the Rise of AI Infrastructure Markets, Kalshi Research, October 2026. Cost breakdown (section 3, figure 4); PJM auction and market table (sections 4 and 6); power contract design (section 7); Silicon Spark Spread (section 8); margin offsets (section 9); product sequence (sections 5 and 10).
- Bloomberg, “Kalshi Turns to Compute Desk in Renewed Push for GPU Futures,” November 18, 2026 (Bloomberg’s post). We read the report through other outlets: Briefs, TokenPost and PANews.
- Asymmetrix Intelligence, “Everyone wants a piece of the future(s): Compute Desk raises $45m,” November 2026. Venue plans, the settlement index for Kalshi’s earlier contracts, and the CME court challenge.
- Compute Desk, Index Methodology and Indexes Overview, read November 19, 2026. Levels and changes shown for August 20 to November 18, 2026.
- Ornn Data, Ornn Compute Price Index methodology and daily index values for the H100, H200 and B200, August 20 to November 18, 2026, retrieved under a full research license on November 19, 2026. data.ornn.com
- Kinetic Alpha, Power, Compute, and the Margin Engine, September 7, 2026. Power use per GPU-hour, cost shares, correlations, notional ratio and margin results.
- PJM, “PJM Auction Procures 134,479 MW of Generation Resources,” December 17, 2025. The 2027/28 auction cleared at the cap approved by the federal energy regulator, $333.44 per megawatt-day.
- Modo Energy, PJM capacity auction 2027/28 research note, January 2, 2026. Estimate of the price without the cap.
- NRG, PJM auction results summary, December 2025. How capacity charges are calculated from a customer’s peak load contribution.
- CCIR, “Compute Derivatives at the CFTC: Letters and Statements,” and its summary of public comments. Compute Desk’s comment letter of October 20, 2026.
- Nicholas Phillips, “What If a Futures-Based ETF Never Had to Roll Again?” ETF Central, August 27, 2026.
- Willkie Farr & Gallagher, “SEC Requests Public Comment on Novel Exchange-Traded Funds,” July 6, 2026.
- CFTC, Policy Statement Concerning the Listing of Perpetual Contracts, adopted May 29, 2026.
- Kinetic Alpha, The Shape Is the Method, November 11, 2026 (how long hourly moves last), and Data Centers Are Not a Flat Load, July 28, 2026 (the cost of a flat-block power hedge).
Limits
- The plan is reported, not announced. We did not read the Bloomberg article itself. The pairing of Ornn with perpetuals and Compute Desk with dated futures comes from other outlets’ accounts of it. Nothing is listed, and the plan may change.
- Kalshi’s figures are quoted as published. The paper labels its cost breakdown as illustrative.
- Compute Desk’s daily history is not public. We used the levels and changes on its website. Our view that its indexes are steadier from day to day rests on their design, not on a measurement.
- The power arithmetic comes from our September study. It uses Dominion Hub prices through September 2026 and converts capacity at a flat load.
- The composite is our illustration. Compute Desk has not published one.