Skip to content
A server rack with rows of drive bays and status lights
AnalysisFinance

Can “Ghost Compute” Become the New Oil Futures?

Analysis | 15 November 2026 | 9 min read

By Palash Jakhotia · Edited by Shreya Narula

A server rack with rows of drive bays and status lights. Photograph: Domaintechnik / Unsplash.

Big Tech is spending about $6 on AI hardware for every dollar AI earns. Idle GPU hours, traded and hedged like barrels of crude, could help close the gap.

The question

Can “Ghost Compute” become the new oil futures? How idle GPU hours are shaping corporate balance sheets.

As approved by the editor before writing began.

The United States has more data centers than any other country, about 5,400 of them, more than Germany, the United Kingdom, and China combined ("Pausing Progress?"). Thousands more are planned as Big Tech races to build the computing power that artificial intelligence runs on. Alphabet, Amazon, Microsoft, and Meta together plan to spend roughly $725 billion on capital expenditures in 2026 (Udland). Most of that money goes into chips, servers, buildings, and the power systems that keep data centers running. But Big Tech companies are facing a big question: "Will this pay off?"

One possible answer is what this article calls "Ghost Compute": the GPU hours that companies have already paid for but leave sitting idle, which could be sold and hedged on a futures market the same way oil is.

AI revenue is growing fast, but it is nowhere near keeping up. Anthropic's annualized revenue passed $65 billion by the end of July 2026 (Mascarenhas et al.), and OpenAI's was running at roughly $50 billion by October (Torrence and Ghaffary). But together, the two largest AI companies bring in around $115 billion a year at their current pace, while the four biggest tech companies are spending $725 billion on infrastructure in a single year. That is about $6 in hardware for every dollar of revenue.

Figure 1Big Tech Investment vs. AI Revenue, 2026About $6 of infrastructure investment for every $1 of revenue at the top AI companies
Source: Chart: Palash Jakhotia. Udland (capex plans); Mascarenhas et al. and Torrence and Ghaffary (annualized revenue).

The comparison isn't perfect, since part of what the labs earn is paid back to Big Tech as cloud rent. But the size of the gap is the point. A JPMorgan analysis estimated that the industry would need about $650 billion in annual AI revenue to earn even a 10% return on the AI infrastructure being built (Swartz).

The gap matters because AI hardware loses value quickly. The median price of an Nvidia H100 chip fell from the low $40,000s in mid-2024 to the mid-$20,000s by late 2025 (Silicon Data). When a company buys billions of dollars of chips, each one starts losing value the day it is installed, whether it is busy or not. Every hour a GPU sits without earning money is an hour of depreciation, electricity, and cooling with nothing to show for it.

Figure 2The median Nvidia H100 price fell about 40% in 18 monthsMonthly median market price of one Nvidia H100 GPU (approximate)
Source: Chart: Palash Jakhotia. Silicon Data.

That pressure shows up in cash flow. Amazon's free cash flow for the twelve months ending in March 2026 fell to $1.2 billion, down from $25.9 billion a year earlier, as its spending surged (Amazon.com). For investors, that means a lower return on invested capital and a growing worry that companies are building faster than customers are buying.

The pressure is also coming from outside the companies. Across the country, communities are pushing back against new data centers. One tracker counts at least 321 moratoriums and restrictions across 32 states, with New Jersey leading at 72 and Georgia close behind at 28 ("Data Center Moratoriums"). At least 14 states are considering statewide moratoriums ("At Least 14"), and New York has already paused new projects ("Pausing Progress?"). Some analysts think the headlines overstate the impact: only about 2.3 gigawatts of planned capacity is actually delayed, despite more than 300 local moratoriums (Barkin et al.). Even so, building new data centers is getting slower and more expensive.

The reasons for the backlash are easy to understand. A large AI data center uses huge amounts of electricity and, in many cases, water for cooling. When a utility has to build new transmission lines to serve one, ordinary households often pay part of the cost through higher monthly bills. Once construction ends, data centers employ relatively few people.

The hardest limit, though, is the power grid itself. Northern Virginia, the largest data center market in the United States (CBRE), hit that wall in 2022. Dominion Energy told data center companies in eastern Loudoun County that power for some new facilities could be delayed for years, because local transmission lines could not carry any more load (Miller). Ireland went further. Data centers there used 23% of the country's metered electricity in 2025, up from 5% in 2015 (Central Statistics Office; Curran), and the state grid operator, EirGrid, has effectively stopped approving new data center connections in the Dublin area until 2028 ("Irish Data Centers"). In both places, the brake was not public opinion but physics: the grid simply ran out of room.

Figure 3Ireland's data centers went from 5% to 23% of its electricityData centers' share of Ireland's total metered electricity use, 2015–2025
0%5%10%15%20%25%5%2015’16’17’18’1911%2020’21’22’23’2423%2025
Source: Chart: Palash Jakhotia. Central Statistics Office (Ireland), 2015–2024; Curran, citing CSO data, 2025.

Put these problems together, and Big Tech faces a squeeze. It can't simply build its way out of the revenue gap, because building is getting harder, and the hardware it already owns loses value every month. The obvious move is to get more out of the machines it already has. If moratoriums and grid limits restrict how many new data centers get built, existing computing power becomes scarcer and more valuable.

This is where the history of oil becomes useful. In the 20th century, crude oil was expensive and risky to extract, ship, and store, and its price could swing sharply. A refinery did not want to find out that its main input had doubled in price, and a driller did not want to find out that its product had suddenly lost half its value. Oil futures solved this problem. An oil future is a standardized contract to buy or sell a set amount of crude oil at a price agreed on today, for delivery on a specific date in the future.

With futures, a producer could lock in its selling price, and a buyer such as an airline could lock in its fuel cost. Traders who had no use for the oil itself joined in too, taking on price risk in exchange for possible profit, which made the market easier for everyone else to use. Over time, benchmarks like West Texas Intermediate and Brent became reference prices for the entire world.

GPU hours could follow the same path. Tech companies that own billions of dollars in Nvidia chips often leave part of that capacity unused during breaks between training runs, while datasets are being prepared, or during off-peak hours. That idle capacity is Ghost Compute, and it is not free to keep around. Running just one megawatt of computing equipment, including cooling, costs roughly $950,000 a year in electricity at average U.S. rates (Axis Intelligence). Instead of letting those hours go to waste, a company could sell them, either right away on a spot market or months ahead through a futures contract. Meanwhile, smaller companies without their own data centers, which now pay high prices to rent computing power on demand, would get cheaper and more flexible access to high-end chips.

Table 1. How Ghost Compute could follow the path of oil
20th-century oil futures21st-century “Ghost Compute”
Unit tradedBarrel of crude oilStandardized GPU-hour
Main bottleneckTankers and storageGrid power capacity
Main riskPrice spikesIdle hardware and its fixed costs
Hedging toolWTI and Brent futuresSpot and forward compute contracts
Typical sellerOil producersBig Tech and cloud providers with spare GPUs
Typical buyerRefiners and airlinesAI startups, research labs, smaller firms

Source: Created by Palash Jakhotia.

A futures market only works if both sides have a reason to show up, and compute has both. On one side are sellers: tech companies and cloud providers holding spare capacity they want to turn into steady income. On the other are buyers: AI startups, university labs, and companies that know they will need a lot of computing power later and worry that its price will spike. An AI developer expecting to rent GPUs next year could take a futures position that gains value if rental prices rise, offsetting its higher bill. Financial firms can trade as well, even without owning a single GPU, which adds liquidity and makes it easier to find a fair price ("Explainer").

There is an important limit, though. Futures do not create demand. If chips sit idle because customers simply don't want more AI computing, then dumping those hours onto a market will push prices down, and sellers may earn very little. What a futures market can do is reveal that falling price early and let a seller lock in today's rate before it drops further. A public price for compute would tell investors, lenders, and boards much sooner whether the AI build-out is running ahead of real demand, instead of making them wait for quarterly earnings.

This idea is no longer just theoretical. In August 2026, CME Group, the world's largest derivatives marketplace, announced it would list two compute futures contracts on the New York Mercantile Exchange, the same exchange where energy futures have traded for decades. The contracts track rental prices for Nvidia's H100 and B200 chips, settle in cash, and were scheduled to begin trading on October 5, 2026, pending regulatory review. CME even compared the move to the way oil grew from a spot business into a global derivatives market (CME Group). Intercontinental Exchange announced its own GPU futures with the data company Ornn, and Nodal Exchange, a power futures exchange, followed with a partnership of its own ("Explainer").

If this market grows, it could change how Big Tech thinks about its hardware. Servers have traditionally been treated as capital expenditures that slowly depreciate over several years. With a working compute market, a company's GPU fleet starts to look more like a commodity reserve. During slow periods, those GPU hours could be sold, and the income could help pay for power contracts and the debt used to buy the hardware.

Plenty could still go wrong. The biggest problem is standardization. A barrel of crude can be graded by quality, but a GPU hour is harder to pin down. The same hour can mean very different things depending on the chip, the memory, the networking between machines, and where the data center is located. Exchanges are trying to solve this by settling contracts against a price index instead of delivering identical rentals. The goal is a trusted reference price, not a perfectly uniform product ("Explainer"). That only works if the index is built from enough real trades. In a young market with few participants, a price index can be thin and easier to manipulate, so regulators and transparency matter.

Security is another concern, but it belongs to the physical market, not the futures market. In the spot market, where a company actually rents its idle GPUs to an outside customer, different clients may run their work on the same hardware. Without strict isolation between workloads, that raises the risk of intellectual property leaks and cyberattacks. Cash-settled futures carry no such risk. Like the CME contracts, they settle in money against a price index, so no workloads or chips ever change hands between traders. Their risks are purely financial: losses when prices move against a position, and the chance that a thin index gets pushed around by a few large players. Exchanges manage the first by requiring traders to post margin through a clearinghouse; the second depends on how many real trades feed the index.

Power is a third problem. Electricity prices change by the hour, and running chips during peak pricing to serve a secondary customer can sometimes cost more in energy than the rental pays. Companies would need software that decides in real time when selling compute is profitable. It is probably no accident that Nodal, one of the exchanges entering this market, already holds most of North America's power futures trading (Nodal Exchange). One day, a data center might hedge its electricity costs and its compute income on the same exchange.

The connection between oil futures and GPU computing power is a strong one. Big Tech is spending far more on AI infrastructure than AI currently earns; its hardware loses value quickly, and community opposition and grid limits are making new data centers harder to build. Turning idle GPU capacity into a tradable commodity will not close the revenue gap by itself, and it cannot invent demand that isn't there. What it can do is help companies earn money from machines that would otherwise sit idle, give smaller firms affordable access to powerful chips, and give everyone a public price for the resource the AI economy runs on. Oil took decades to go from a risky physical business to one of the world's most important financial markets. With the first compute futures arriving on the same exchange where oil trades, Ghost Compute may be starting down that same path.

This piece is for general information and discussion. It is not investment advice and does not recommend buying or selling anything.

About this piece

Correspondent

Palash Jakhotia

Editor

Shreya Narula

References

  1. 1.Amazon.com, Inc. "Amazon.com Announces First Quarter Results." Exhibit 99.1 to Form 8-K, U.S. Securities and Exchange Commission, 2026. www.sec.gov/Archives/edgar/data/0001018724/000101872426000012/amzn-20260331xex991.htm
  2. 2."At Least 14 US States Weighing Data Centre Moratoriums." Capacity, 2026. capacityglobal.com/news/at-least-14-us-states-weighing-data-centre-moratoriums/
  3. 3.Axis Intelligence. "U.S. Data Center Electricity Cost by State 2026: All 50 States Ranked." Axis Intelligence, 30 Oct. 2026. axis-intelligence.com/data-center-electricity-cost-by-state/
  4. 4.Barkin, Maya, et al. "Everyone Says Datacenter Moratoriums Are Killing the US Buildout. We Disagree." SemiAnalysis, 15 Sept. 2026. newsletter.semianalysis.com/p/everyone-says-datacenter-moratoriums
  5. 5.CBRE. "Northern Virginia Extends Lead as Largest U.S. Data Center Market in 2025." CBRE, 18 Mar. 2026. www.cbre.com/press-releases/northern-virginia-extends-lead-as-largest-u-s-data-center-market-in-2025
  6. 6.Central Statistics Office. "Data Centres Metered Electricity Consumption 2024." CSO, 10 June 2025. www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2024/keyfindings/
  7. 7.CME Group. "CME Group and Silicon Data to Launch Compute Futures on October 5 to Unlock New Way to Hedge AI Risks." CME Group, 11 Aug. 2026. www.cmegroup.com/media-room/press-releases/2026/8/11/cme_group_and_silicondatatolaunchcomputefuturesonoctober5tounloc.html
  8. 8.Curran, Ian. "Data Centres Account for Almost a Quarter of Irish Electricity Usage in 2025." The Irish Times, 7 July 2026. www.irishtimes.com/business/2026/07/07/data-centres-account-for-almost-one-quarter-of-irish-electricity-usage-in-2025/
  9. 9."Data Center Moratoriums." Electric Choice, 2026. www.electricchoice.com/datacenters/moratoriums/
  10. 10."Explainer: What Are Compute Futures and How Do They Work?" FIA, 23 Sept. 2026. www.fia.org/fia/articles/explainer-what-are-compute-futures-and-how-do-they-work
  11. 11."Irish Data Centers Used More than 21 Percent of Nation's Total Metered Electricity." DatacenterDynamics. www.datacenterdynamics.com/en/news/irish-data-centers-used-more-than-21-percent-of-nations-total-metered-electricity/
  12. 12.Mascarenhas, Natasha, et al. "Anthropic's Annualized Revenue Tops $65 Billion Before IPO." Bloomberg News, 18 Aug. 2026. news.bgov.com/financial-accounting/anthropic-revenue-run-rate-surpasses-65-billion-ahead-of-ipo
  13. 13.Miller, Rich. "Ashburn Power Crunch May Cause Delays in Data Center Construction." Data Center Frontier, 29 July 2022. datacenterfrontier.com/ashburn-power-crunch-may-cause-delays-in-data-center-construction/
  14. 14.Nodal Exchange. "Nodal Exchange and Compute Desk to Launch Compute Futures Contracts." Nodal Exchange, 3 Sept. 2026. www.nodalexchange.com/wp-content/uploads/Nodal-Exchange-and-Compute-Desk-to-launch-Compute-Futures-Contracts-Sep-3-2026.pdf
  15. 15."Pausing Progress? The Emerging Debate Over Data Center Moratoriums." ACEC, 2026. engineeringinc.acec.org/blog/pausing-progress-the-emerging-debate-over-data-center-moratoriums/
  16. 16.Silicon Data. "The Illusion of Stability: Unpacking H100 GPU Market Value Trends." Silicon Data, 30 Jan. 2026. www.silicondata.com/use-cases/h100-gpu-market-value-trends
  17. 17.Swartz, Jon. "AI Infrastructure Demands $650 Billion in Annual Revenue for Viable Returns: JPMorgan Analysis." Techstrong.ai, 16 Nov. 2025. techstrong.ai/agentic-ai/ai-infrastructure-demands-650-billion-in-annual-revenue-for-viable-returns-jpmorgan-analysis/
  18. 18.Torrence, Rebecca, and Shirin Ghaffary. "OpenAI's Revenue Run Rate Nears $50 Billion, Less Than Reported." Bloomberg News, 8 Oct. 2026. news.bgov.com/artificial-intelligence/openais-annualized-revenue-nears-50-billion-ft-says
  19. 19.Udland, Myles. "'Magnificent 7' Earnings Rush Reveals AI Spending Surge, with Hyperscaler Capex Set to Reach $725 Billion in 2026." Yahoo Finance, 29 Apr. 2026. finance.yahoo.com/markets/article/magnificent-7-earnings-rush-reveals-ai-spending-surge-with-hyperscaler-capex-set-to-reach-725-billion-in-2026-224901707.html

Cite this piece

Palash Jakhotia. (2026, November 15). Can “Ghost Compute” Become the New Oil Futures?. The YEL Journal. https://yel-journal.lovable.app/article/can-ghost-compute-become-the-new-oil-futures

Corrections

No corrections have been made to this piece. We correct errors openly, with a dated note here. Corrections policy

The readers' list

New pieces from the Journal, by email. Unsubscribe any time.

Read next