By A popular read of this week’s AI-chip tape is that the trade hinges on Nvidia’s next platform, or on a data centre in orbit. The document that actually moved was a regulator’s letter. Nvidia GPU futures, the contracts that would have turned an hour of H100 rental into a cleared price, missed their 5 October open after the U.S. Commodity Futures Trading Commission extended its review. The chips did not fail to ship. The hour failed to become a commodity.
Having covered the CME filing when 5 October was still the intended trade date, the split is now the story. Nvidia reported revenue of $96.2 billion for the quarter ended 26 July 2026, up 106% from a year earlier, of which $89.0 billion was data centre revenue. It guided the current quarter to $108.0 billion, plus or minus 2%, and said Vera Rubin is in full production. The binding constraint on financialising that buildout is a benchmark the CFTC is willing to let people settle against.
The ordering is the part no chip-roundup is writing down. Grayscale renamed its bitcoin-miners ETF the Grayscale AI Compute ETF, ticker GCPU, effective 22 September. The Information reported that a CFTC letter to CME, dated 21 September, extended the futures review by 45 days, to 9 November, because the contracts raise novel and complex issues. In the same week a brokerage account could buy companies that used to sell hashrate, and could not buy a hedge on the GPU-hour those companies are trying to sell. Crude got a futures market before the equity craze. Compute got the equity craze first.
Key facts
Launch filed for 5 October 2026, now on hold. The Information says the CFTC extended the review; reporting on its letter puts that extension at 45 days, to 9 November. Regulators call the rental market fragmented and opaque. — The Information, 25 September 2026.
Contract size 730 GPU-hours, cash-settled, one-cent tick worth $7.30. H100 is Globex GPU1; B200 is GPU2. — NYMEX Submission No. 26-370.
19 July neocloud curve: H100 about $2.72/hour spot versus $2.38 at 36 months. B200 about $5.62 versus $5.17. — Silicon Data, 17 July 2026.
Nvidia Q2 revenue $96.2 billion; data centre $89.0 billion. Q3 guide $108.0 billion, plus or minus 2%. — Nvidia, 26 August 2026.
GCPU about $11.7 million on 29 September, our arithmetic from the holdings file. CoreWeave 8.43%, Nebius 8.42%, Hut 8 8.06%, IREN 7.59%. — Grayscale.
AI compute profit about $1.5 million per megawatt a year, versus about $500,000 from bitcoin mining. — CoinShares, via FinanceFeeds, 16 September 2026.
Hyperscalers about 70% of capacity. Nine of 11 listed AI-compute specialists are current or former bitcoin miners. — Grayscale, 21 September 2026.
What a 730-hour Nvidia contract actually is
NYMEX asked the CFTC to approve two compute futures, on the Silicon Data H100 and B200 rental indexes. Each is 730 GPU-hours, one chip for a month, cash-settled, quoted in dollars per hour. The minimum move is one cent, worth $7.30. CME wanted Globex from Sunday 4 October for a Monday 5 October trade date, with monthlies out to 36 months.
A GPU-hour behaves like a megawatt-hour. The compute exists while the chip is powered and is gone when the job ends, so there is nothing to warehouse. The future is a view on the rental rate, plus a risk premium.
Silicon Data, backed by DRW, publishes the indexes CME proposed for settlement. Carmen Li, on 17 July, set out the design: an H100 hour, a B200 hour and an A100 hour are different products, and settlement is cash against a reference. Her neocloud curves that week were in backwardation across all three. Buyers would lock a lower three-year rate than today’s hour. She read that as more supply ahead. She also wrote the line the CFTC has now acted on: as of 20 July, no GPU future had final approval, and approval was the gate.
At about $2.72 an hour, one 730-hour contract references roughly $1,990 of rental, and one tick of $7.30 is about 37 basis points of that notional. The filing is a precision tool on a few thousand dollars of rent. The AI-chip view in size is being expressed in Nvidia equity and in the power owners, not in this contract.
Silicon Data put the one-year H100 rate at roughly $1.70 an hour in October 2025 and about $2.65 by March 2026, a 56% move on the tenor procurement signs, citing Blockworks. A listed curve would have given a treasurer a hedge and a lender a mark. It would not have repealed the move.
“For years, two companies buying the exact same GPU capacity could pay wildly different prices with no way to know who got the better deal. They will now have a benchmark to check that against. Compute futures give the market something it’s never had: a public, tradable reference price for the resource every AI system runs on.”
— Carmen Li, chief executive of Silicon Data, CME Group, 11 August 2026. The October open she was describing has since been held.
Who is moving, and who is waiting on the CFTC
CME has not withdrawn the product. The 11 August release still says 5 October, pending review, under NYMEX rules. The Information’s 25 September account is the update that matters: the launch is on hold, the review was extended, and regulators are examining manipulation risk in a rental market they call fragmented and opaque. Reporting on the letter adds a 45-day extension to 9 November, because the product is novel and complex. That date ends a review. It is a listing date only if CME sets one.
Li’s July essay remains Silicon Data’s last full public position: curves can exist before the contract, and bank desks generally cannot use an uncleared version. Nothing on the firm’s own pages answers the extension. The regulator’s position is that the negotiation record is not yet a price.
Nvidia is not an applicant. The hours are already sold under private contracts: $89.0 billion of data-centre revenue in one quarter. The 26 August release says financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR aim to mobilise more than $500 billion of third-party capital over time, subject to definitive agreements. The same release names SpaceXAI as a deployment partner and gives no chip count and no dollar figure, which is why orbital posts on X are ahead of any filing. Nvidia’s equity book already owns a slice of the customers, including a SpaceX line via xAI.
The product that did list is an equity wrapper. GCPU started under the new name on 22 September, on NYSE Arca, tracking the Indxx High Performance Computing Index. On 29 September, CoreWeave — a former ethereum miner, in Grayscale’s classification — was 8.43%. Nebius, the one name that research note says has no mining heritage among 11 listed AI-compute specialists, was 8.42%. DigitalOcean was 8.11%, Hut 8 8.06%, IREN 7.59%, with Galaxy, MARA, CleanSpark and Cipher still on the book.
CoreWeave’s line was worth $989,704 at an 8.43% weight, which puts the whole fund near $11.7 million. Nebius, $988,504 at 8.42%, gives the same answer. GCPU is a renamed miners wrapper the size of one neocloud invoice: proof of product sequencing. The money in size is in private credit and in megawatts IREN has been buying, including Nostrum in Spain. CoinShares expects IREN’s mining exit to finish in 2026.
Two other venues are not on NYMEX’s clock. FinanceFeeds reported in May that ICE and Ornn plan cash-settled GPU futures on a Compute Price Index of completed transactions, covering H100, H200, B200 and RTX 5090 hardware, with no date set. In July, Kalshi launched GPU forward curves. A tape of completed trades answers an opacity objection more directly than a composite of posted offers, and Kalshi is trying to publish a curve while CME waits.
“AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue. And demand is accelerating.”
— Jensen Huang, founder and chief executive of Nvidia, 26 August 2026. The revenue is real. The hedge on the hour that produces it is not listed.
Quick take. Nvidia is booking the hours. Grayscale listed a small equity basket of the firms that own the power. CME is the name standing in the CFTC’s queue.
What the numbers say when you put them in one place
X has run Vera Rubin, orbital compute and the futures hold in the same week. Only the hold has a contract specification. Shay Boloor’s 18 September map shows who supplies the chips, who rents them and who buys the hours.
Silicon Data’s July curve and Nvidia’s August guide do not describe the same scarcity. H100 out-month rent was about 13% below spot over 36 months, and the long end was already flat, roughly $2.38 to $2.44 beyond 15 months, while the company selling the silicon guided a quarter above $100 billion. Forward rental rates can fall while chip revenue rises once the bottleneck has moved onto power, shells and interconnection. Grayscale says hyperscalers hold about 70% of capacity. The filed H100 future hedges a neocloud sticker. It leaves the grid queue untouched.
CoinShares estimates AI and high-performance-computing profit at about $1.5 million per megawatt a year, against about $500,000 from bitcoin mining, and labels both as industry estimates rather than audited site returns. On that spread it sees at least 35 EH/s of listed mining capacity leaving, about 4.7% of a 750 EH/s network, and as much as 70% of listed-miner revenue coming from AI by the end of 2026, against roughly 30% earlier, on more than $70 billion of announced contracts. The option left in a mining stock is whether the interconnection can be resold to a GPU tenant.
CoreWeave shows why a GPU-hour future would not have duplicated the equity. On deals signed since 30 June it disclosed about $40 million per megawatt of annualised revenue, on terms of three to six months. That is the price of capacity available now, against a 12-to-18-month build, and it is not an H100 hourly print. The multiple and the $2.72 hour are different markets, and only the multiple is in a filing.
What you can mark this week
What is stuck at the CFTC
Nvidia equity, after a $96.2 billion quarter and a $108 billion guide
A cleared H100 or B200 rental future on NYMEX
GCPU, about $11.7 million, still full of miners and neoclouds
An approved settlement index for that future
Short-dated leases, including CoreWeave’s ~$40 million per megawatt
Daily margin on a GPU-hour for those leases
Silicon Data’s published curve, backwardated, not a contract
The 5 October open, now inside a review toward 9 November
Kalshi curves, and an ICE/Ornn transaction-based proposal with no date
One benchmark the street agrees is hard to lean on
“We are AI optimists but investing realists. The technology will transform the economy and society, in our view, and demand for compute — the raw material of machine intelligence — will be insatiable. But the supply side can only expand so fast, especially when large segments of the voting population are questioning the tradeoffs.”
— Zach Pandl, head of research at Grayscale, with Will Ogden Moore and Charlie Perkins, 21 September 2026.
Quick take. Chip revenue can rise while the forward rental curve slopes down. The equity is pricing an energised megawatt. The CFTC file is stuck on a rental print nobody can push around.
Why this looks like power markets, not like a new ticker
Pete Keavey, CME’s global head of energy and environmental products, made the oil comparison on purpose. He lists the commodities that already have benchmarks. The claim he made in August is the claim the Commission has not accepted.
“Compute has become the currency of the AI age, and this innovative market will bring transparency to the current and future costs that AI builders and hyperscalers need to hedge as they grow. Just as oil fueled the 20th century economy and evolved from spot trading into a global derivatives market, our futures contracts will now turn compute into a standardized, tradable commodity.”
— Pete Keavey, CME Group, 11 August 2026.
AI buyers and the banks lending against GPU fleets want a cleared hedge, so a default does not sit inside a bilateral contract. Li is right that clearing matters here: the names run from venture-backed labs to hyperscalers, and daily margin is how they face each other. A future is only as clean as its index. If a handful of lessors can move the posted offers that feed it, they can move settlement. The Information’s account — manipulation risk, a fragmented and opaque spot — is that objection, applied to a market with no hub.
Electricity is the closer precedent. Before grid operators published a locational price, wholesale power was a pile of bilateral contracts, and the futures that eventually listed settled to a hub rather than to an average of the morning’s offers. GPU rental looks like that earlier market. Prices differ by chip, interconnect, term and geography, and CME’s filing ties the indexes to neocloud offers rather than to the hyperscale contracts where Grayscale says about 70% of capacity sits. A future that misses most of the physical market can be called complex without any hostility to the technology.
CME bitcoin futures, listed in 2017, taught the industry that a future can legitimise an asset. This file uses the earlier lesson: do not list until the reference can survive a manipulation case. Oracle-priced GPU perps, which Li noted are already being sketched, copy the problem at a smaller size. ICE’s transaction-based design with Ornn answers the objection and has no date. Pandl’s other point is political. A daily AI price can be blamed like a commodity, and the contract adds no megawatts. Forty-five days is time to prefer silence to a weak index.
Three clocks from here
First, 9 November ends a review, and the base case is that nothing lists that morning. A 45-day extension does not create a hub. If settlement is still a composite of posted neocloud offers, the next step is another extension or a narrower contract. The firms posting the offers can move an offer-based index. A transaction index is harder to lean on and slower to build, because GPU rental has no central tape. Until one of those changes, GPU1 and GPU2 stay dark.
Second, if anything lists, the H100 fits the objection better than the B200. It has the longer history. On 19 July its long end was flat for two years inside a six-cent band, while the B200 forward turned up at the far end, a recovery Li said the curve could not explain. A review worried about a story-driven benchmark will take the duller chip first. Leading with the scarcer one is how a file earns another 45 days.
Third, through the fourth quarter the price that moves accounts stays the contracted megawatt. GCPU, IREN and CoreWeave can all reprice without GPU1 trading. The desks that lose the delay are the treasurers who wanted daily margin on the hour, so a short $40 million-per-megawatt lease could be hedged instead of held as unmarked basis. That basis is the opacity the CFTC refused to certify.
Vera Rubin shipping tells you Nvidia expects the hours to be consumed. It does not tell you the price of an hour in March 2027. Until a benchmark survives the review that now points at 9 November, the AI-chip trade is an equity and credit trade with a commodity story attached. The commodity itself is still a stack of invoices.
FAQ
What are Nvidia GPU futures?
Nvidia GPU futures are proposed cash-settled NYMEX contracts on the hourly rental price of specific Nvidia chips. CME and Silicon Data filed an H100 contract, Globex code GPU1, and a B200 contract, code GPU2. Each represents 730 GPU-hours, quoted in dollars per GPU-hour. No physical chip changes hands. Settlement uses a Silicon Data rental index.
Why did the 5 October launch stall?
The 5 October date was always conditional on CFTC approval. The Information reported on 25 September that the launch is on hold because the Commission extended its review, citing manipulation risk in a rental market it describes as fragmented and opaque. Reporting on the letter puts the extension at 45 days, to 9 November. CME has not announced a new trade date.
How much is one tick on the 730-hour contract?
The filing sets the minimum fluctuation at one cent per GPU-hour, worth $7.30, because 730 times one cent is $7.30. At Silicon Data’s 19 July H100 spot of about $2.72, one contract references roughly $1,990 of rental, so a tick is about 37 basis points of that notional. The B200 spot that day, near $5.62, implied a contract near $4,100.
How do bitcoin miners fit the AI-chip trade?
Miners already control permitted power and a grid interconnection, which the chip does not include. CoinShares estimates AI compute earns about three times the profit per megawatt of bitcoin mining. Grayscale says nine of 11 U.S.-listed AI compute specialists are current or former bitcoin miners. GCPU’s 29 September file still holds Hut 8, IREN, Galaxy, MARA, CleanSpark and Cipher.
What is the Grayscale AI Compute ETF?
GCPU is the former Grayscale Bitcoin Miners ETF, ticker MNRS, renamed effective 22 September 2026. It tracks the Indxx High Performance Computing Index on NYSE Arca. On the 29 September holdings file the fund is about $11.7 million. The rename offers the crossover book in a brokerage account. It does not hedge the GPU-hour.
When could H100 rental futures list?
The date in the public record is 9 November 2026, and it ends an extended CFTC review. It is not a confirmed open. A listing requires the Commission to finish and CME to set a trade date. If the index is still built from posted offers, a further extension follows from the objection The Information reported. ICE, Ornn and Kalshi are on separate clocks.