# Rise of Compute Capital Markets: Pricing AI’s Most Important Commodity Source: https://x.com/jinglingcookies/status/2081726324516028463 Kiwi News: https://news.kiwistand.com/stories/Rise-of-Compute-Capital-Markets-Pricing-AIs-Most-Important-Commodity?index=0x6a68793cb0e11ec79297b49d02494723c2e8e4b9835c7a3c705779de156ed4aa1c68ee2b Submitted by rvolz.eth on 2026-07-28 UTC, 1 upvote, 0 comments ## Summary Compute needs standardized benchmarks and futures contracts so buyers and sellers can hedge price swings, as electricity and oil did after decades of opaque bilateral trading. H100 spot rental prices fell from about $8 an hour in late 2023 to between $2 and $4 two years later, according to SemiAnalysis. In May 2026, CME with Silicon Data and ICE with Ornn each announced plans for cash-settled GPU futures, pending regulatory approval. Four companies currently build compute indices: Silicon Data, Ornn, Compute Desk and Pluto. ## Article The need for compute financialization, and how it will be developed. Thank you to @0xmoonrunner, @0xMetaLight, for review and feedback! The state of compute In late 2023, at the height of the GPU shortage, renting an NVIDIA H100 spot cost about $8 an hour. Two years later the same chip clustered between $2 and $4 (SemiAnalysis). In oil or electricity, a two-thirds move in the cost would be routine: producers or consumers would simply hedge it. Hedging refers to the act of locking in a price today to neutralize the risk of price volatility, which affects a company’s cost or revenue. This can be done in the form of a futures or options, which are standardized financial contracts that allows buyers and sellers to lock in a specific price for electricity to be delivered at a future date. For example, jet fuel is one of the largest input costs for an airline company. When oil prices rise, costs climb and margins get compressed, reducing profitability. To protect against that, airlines can lock in the prices of fuel by buying oil futures. If prices do spike, the increase in oil futures value will cover the rising oil prices. In compute, now the most important input cost in the AI economy, the same form of hedging couldn't be achieved. There was no benchmark price, no forward curve, no contract to buy or sell. If compute prices dropped, compute providers suffer from declining revenue. If compute prices increased, compute buyers have to deal with higher costs. The market was at the mercy of compute volatility. That is rather inefficient for such a large industry. The five biggest US clouds will spend over $600 billion on infrastructure in 2026, roughly three-quarters of it AI-specific (IEEE ComSoc). AI labs are also seeing more than 50% of their budget allocated towards compute (Epoch AI). Goldman sees AI capex climbing toward $1.6 trillion by 2031 (Goldman Sachs). Yet until recently compute was priced mostly through bilateral deals. That is now changing fast. In May 2026, both major US derivatives exchanges has announced plans to make compute tradable. CME with Silicon Data and ICE with Ornn, each launching cash-settled GPU futures pending regulatory approval. With it we are also seeing the infrastructure stack mature, particularly spot marketplaces where compute trades and indices that measures compute prices. This report maps the market taking shape around compute: what it takes to build, who is building it, and who captures the value. It is anticipated that the compute index layer is shaping up to be winner-take-all, claimed not by whoever measures the market most accurately but by whoever gets their benchmark wired into the most contracts. The spot marketplaces, matching buyers to sellers, hold the thinnest moat of the three layers and risk competing on fees, which can potentially be a race to zero. And the exchanges listing compute futures face intense competition from the incumbents CME and ICE. And with that, let’s dive deep into compute capital markets. The commoditization of compute Before any asset can be considered to be a commodity and made tradable on an exchange, there needs to be a standardized method to measure the asset and define the grading dimensions around it to make it possible to price differentials. This makes it possible for the market to trade assets that have a certain level of heterogeneity to it. To understand this, let’s take a look at the history of electricity and oil commoditization. The history of commoditization Both electricity and oil were once opaque, bilateral, and relationship-driven trading assets. Electricity Before the 1990s, electricity was not a traded commodity. The market was dominated by vertically integrated monopolies that owned generation (power plants), transmission, and distribution all at once. Given the centralization of electricity capacity in the hands of a few, electricity was traded almost exclusively through long-term bilateral deals between these monopolies and electricity buyers. That changed with the Energy Policy Act of 1992 with the introduction of spot electricity markets. Independent System Operators (ISOs) were established to manage the grid and provide transparent, real-time pricing of electricity, with the trading unit being kilowatt hour (kWh). Once the spot market provide reliable with sufficient liquidity, a derivative ecosystem began emerging around electricity. New York Mercantile Exchange (NYMEX) launched electricity futures in 1996, and the Hourly Price Forward Curve (HPFC) was developed a few years later, which maps the price which electricity has to be delivered across the next few months. Electricity is standardized based on the following dimensions: Unit: Traded by the kWh Location: The grid location affects transmission Time: Whether electricity is delivered during peak or off-peak period Oil Through the mid-20th century, oil didn’t trade on the open market at all. Similar to that of electricity, there were vertically integrated monopolies that controlled the whole oil supply chain. As a result, oil was mostly traded opaquely through bilateral contracts. It will not till the 1970s did a spot market emerge. In 1973, an oil crisis introduced unprecedented volatility into the global energy market and there was a need for the industry to lock in prices for future production and supply of oil. Subsequently, oil futures were born in 1983 to manage the volatility of the market. Today, the oil market prices everything against established regional benchmarks, most notably Brent Crude (priced in Europe/Middle East) and West Texas Intermediate (WTI) (priced in North America). Oil is standardized on three dimensions Unit: Traded by the barrel Quality: Based on two intrinsic properties (a) density and (b) sulfur content Delivery location: Eurpe/Middle East for Brent, North America for WTI The difficulty in standardizing compute Compute is now at a similar stage to the early days of electricity and oil trading, where compute is mostly transacted through bilateral deals with opaque compute prices and compute prices are highly volatile. There has however, been difficulties in pricing compute. Despite there being a standardized unit GPU-hour (the cost of renting a GPU per hour), the output per GPU-hour varies significantly based on the plumbing around the GPU: (a) Form Factor / Network Interconnect: The form factor (PCIe vs SXM) determines how a GPU is assembled, which dictates the network interconnects (wiring that allows for data exchange between GPUs to run AI workloads) it can use. SXM variants are more performant due to the higher memory bandwidth and network interconnect, observed in the pricing difference (runpod). (b) SKU (Stock Keeping Unit): There are various types of GPUs in the market: H100, H200, A100, B200, B300, Google TPUs, AWS Trainium. Each of this SKU serves different workloads, with the newer series like Blackwell GPUs delivering higher performance than the older Hopper series (H100, H200). This difference in workload servicing can be observed in the following, with estimated costs for B200 being more than 2x of H100 (Silicon Data). (c) Contiguity / Duration: Contiguity refers to whether GPU-hours comes as a single, guaranteed, uninterrupted block or as scattered hours that can be reclaimed by the compute provider at any moment. A spot GPU-hour and an on-demand GPU hour aren’t substitutes, they are different grades with varying levels of compute delivery reliability, and are priced differently as shown below (Compute Exchange). (d) Provider: Compute provided by hyperscalers tend to command a premium given their reliability, offering enterprise-grade uptime guarantees. For companies that are running long training runs or build applications that might experience unexpected demand spikes, having reliable access to compute becomes non-negotiable, and is a feature that they are willing to pay a premium for. The graphs below showcase the difference in the H100 on-demand hourly rate between that provided by hyperscalers and the one provided by neoclouds (Silicon Data). (e) Location: Based on where a GPU is located, the compute price might differ due to the following factors: Energy costs: Power is one of the largest operating costs of a data center, required to power the GPUs to provide compute. Where electricity is cheap, compute can be provided at a cheaper rate. Similarly, if certain regions experience energy-price shock (as with the recent war), compute prices will be directly impacted and increase. Cooling costs: Cooling is crucial for a GPU to remain functional and run at high performance. Countries that have naturally cool climates can spend less on cooling infrastructure, and the cost savings can be reflected in the form of lower compute prices. Regional supply-and-demand imbalance: GPU availability is spread out across the world. Demand might be greater in certain regions e.g. if there is high density of AI startups, which leads to compute prices in that region to be higher. All of the above factors trickles down and causes compute to be priced differently based on the region they are based in, despite being the same GPU. The following charts depict the difference in H100 prices within US and globally (Silicon Data). Ultimately, two H100-hours can deliver different outputs based on the various factors mentioned above. This means the market clears at a range of prices for what is nominally the same unit. Index builders will have to decide which dimensions define the contract grade (and therefore belong in the benchmark) and which are treated as price differentials. The tighter the grade, the better the index hedges real exposure, but the thinner the underlying transaction set (since only a small number of transactions fit within the narrowly-scoped grade), leading to higher risk of manipulation. The looser the grade, the more liquid the data, but the index corresponds less to what the buyer is actually purchasing, increasing basis risk. State of compute standardization The crucial point, though, is that a commodity does not have to be entirely homogeneous to trade. The above examples of electricity and oil are testimony to that. The key is to standardize grading and transform quality differences into price differences, which can then be fed into building a compute index. There are currently 4 companies building compute indices: Silicon Data, Ornn, Compute Desk and Pluto. All of which have their own methodologies to price compute. We will take a deeper look into each of them in part II of this compute capital market series. The buyers and sellers of compute futures With the standardization of compute and creation of compute indices, compute futures contracts can be created by referencing them. These futures can then be listed on exchanges to be bought and sold by hedgers. This section explores the buy and sell side of compute futures. Compute supplier: Sellers of compute futures The natural shorts of compute futures are companies that loses revenue when compute rental prices fall. By selling compute futures, these companies are able to steady their revenue line as the gains from selling futures offset the decline in rental income when compute prices fall. Compute buyers: Buyers of compute futures The natural longs are everyone who loses when prices rise: labs committing to enormous fixed compute budgets before they know their revenue, enterprises watching inference bills climb, companies selling AI features at fixed prices. Other participants Speculators are an essential source of liquidity in the markets. This includes funds as well as retail traders that are looking for AI-cycle exposure Market-makers serve as a counterparty to buyers and sellers, and are key to maintaining a liquid market. Current compute capital market landscape With a detailed understanding of the compute market, it’s time to explore the burgeoning compute market landscape, with 3 main layers in the compute financialization stack: Indices There are a couple of methodologies utilized in building the index. The first is to crawl listed prices is comprehensive but measures asks and not fills. This might be misleading as providers tend to inflate list prices to negotiate more favorable bilateral terms. The other is to aggregate real compute transactions that occur. This is highly accurate but thin volume of data available might lead to price skews. Existing companies building indices include @Silicon_Data, @OrnnExchange and @ComputeDesk. Spot Marketplaces Spot marketplaces match compute buyers with compute suppliers. Beyond matchmaking, they collect and normalize raw pricing data for index providers (which in some case could be themselves), creating price transparency. Existing companies building spot marketplaces include @sfcompute, @computeexchange, @OrnnExchange, @vast_ai and more. Derivatives Exchanges There are 3 exchanges exploring building compute derivatives at this moment Intercontinental Exchange (ICE) Chicago Mercantile Exchange (CME) American Innovation Exchange by Architect Both ICE and CME have announced partnerships with Ornn and Silicon Data respectively, and it is anticipated that cash-settled compute futures will be made available on the exchanges by the end of 2026. Architect by far is the furthest ahead, and has already started offering compute futures on American Innovation Exchange, a CFTC-regulated Designated Contract Market (DCM) launched by Architect. In addition, the team is actively working towards building physically-settled compute futures, which will be the first of it’s kind. From the above, it’s evident that compute capital markets are increasingly gaining more traction, with both incumbents and new startups vying to get a piece of the market. This brings us to the next section which hypothesizes on who will emerge as the winner / dominant player in each layer. Who wins in the compute capital market stack Indices There are 2 key elements that contributes to an index’s ability to capture the market. The first being it’s accuracy. Compute futures pays out against the index. If the index drifts substantially from what compute buyers are actually paying, a compute buyer selling futures won’t really be hedged, and is left holding a basis risk. And the second being distribution. On top of having futures settle using the index as a reference, it also includes use cases where hedge funds and research firms utilize the index for market monitoring and inclusion into algorithmic trading strategies. Spot Marketplaces A marketplace that only matches buyers with sellers is a broker, and competition might push the market towards zero take-rate. There are a few ways to stand out from the competition: Move into delivery: Own hardware, guarantee reliability and charge for certainty. However, is it worth noting that owning hardware means taking on the risk of GPU depreciation and fluctuating compute prices Move into data: Turn order flow into a proprietary index others settle against, moving towards becoming a high-margin benchmark layer High reliability: Buyers that require high uptime e.g. AI application, are likely willing to pay for a steady compute source Embedded Financing: An all-in-one stop for compute buyers, who can post their compute as collateral to finance their debt which can be used for the likes of more hardware purchases, R&D etc. Derivatives Exchanges The incumbents, CME and ICE, bring with them a robust suite of trading infrastructure. In addition, they already have significant distribution with institutions already hedging on their exchanges, which is likely to contribute to compute futures liquidity upon listing in the future. It may seems like a tough market to break through. However, it is worth noting that Architect has been extremely quick to go-to-market, and are the sole exchange currently building physically-settled compute futures which is preferred by buyers of compute futures i.e. AI startups. What comes next, and what could go wrong 2026 has seen the rise of compute capital markets, with both funding flowing into the sector and startups gaining traction. Nonetheless, the sector is nascent, and there are 4 things that could derail the market and stump its progress: Index ManipulationThis is a huge risk, and is especially prominent if the benchmark is built on a small set of transaction data (very specific region / interconnect / provider etc.) where a single outlier transaction will cause a huge shift in the compute index price.For index methodologies that utilizes listed prices, it is prone to manipulation through spoofed listing prices by providers which are not intended to be filled, but just to spike the compute price index. Lack of hyperscaler participationHyperscalers primarily operate through bespoke bilateral contracts. They hence have a preference for markets to stay opaque to keep keep pricing power in their hands. However, their participation in spot marketplaces or being transparent with their contract data is crucial in building more accurate indices.It is likely that regulators will have to step in to mandate hyperscalers contribute their data towards index builders, similar to the process of how electricity and oil were commoditized in the past. Insufficient market liquidityHedgers won’t trade without depth. Market makers won’t provide liquidity without flow. This is the classic chicken-and-egg problem when it comes to creation of new markets. There is a need to incentivize these stakeholders to participate. Drastic GPU depreciationShould efficiency gains compound, or the AI industry sees a shift in focus from training to inference, we might see a decline in compute demand over time. If supply drastically dwarfs demand for compute, the cost of GPU will become deflationary, leading to less demand for compute futures buyers, since they are almost certain they will be able to purchase compute for a lower price eventually. The bottom line 2026 is the year of compute financialization. With indices laying the fundamentals for building derivatives for hedging, the sector is starting to look more robust. Compute price has remained volatile, and with the government advocating for compute commoditization, it will be worth looking out for efforts in bolstering the liquidity of compute futures, which will in turn attract more hedgers and market makers, creating a flywheel effect. This is part I of a series on compute capital markets. The subsequent pieces will cover the landscape in detail, dive deeper into index manipulation and explore blockchain's role in compute capital markets.