Crypto-style derivatives tied to artificial intelligence compute power are moving from concept to live trading, with offshore perpetual futures and event-based contracts already active before planned regulated futures launches in the United States.
The emerging market allows traders and companies to take positions on the future cost of GPU computing capacity, a critical input for training and running AI models. According to a Bernstein report, Bermuda-based Architect’s AX has listed compute perpetual futures designed in the style of cryptocurrency derivatives, while CME Group and Intercontinental Exchange are preparing cash-settled compute futures that are expected to launch after regulatory clearance, potentially later in 2026.
The products mark a notable shift in how AI infrastructure is financed and managed. Instead of treating computing capacity only as a physical or cloud-service expense, market participants are beginning to trade it as a financial exposure. The contracts are intended to help firms hedge future costs, manage production economics and speculate on whether GPU rental prices will rise or fall as demand for AI computing power changes.
Bernstein analysts Gautam Chhugani and Madison Rezaei said offshore platforms have moved first because they face more flexible oversight and because demand for GPU exposure has developed faster than the regulated market’s approval process. CME and ICE, by contrast, are waiting for clearance from the Commodity Futures Trading Commission before launching their own cash-settled products.
The early activity remains small and highly experimental, but it reflects a broader trend: AI compute is becoming a tradable commodity-like input, similar in some ways to electricity. As companies spend billions of dollars on chips, data centers and cloud capacity, the ability to lock in or transfer price risk is becoming more important.
Offshore platforms move first
Architect’s AX, based in Bermuda, has introduced compute perpetual futures that resemble the perpetual contracts widely used in digital-asset markets. These contracts do not expire on a fixed date. Instead, they use a funding-rate mechanism to keep trading prices close to an underlying index.
In a typical perpetual futures structure, traders holding long positions pay traders holding short positions when the contract trades above the index price. When the contract trades below the index, the payment usually moves in the opposite direction. The funding mechanism is designed to pull the contract back toward the reference price over time.
That structure is now being applied to GPU pricing. In practical terms, traders can take views on whether the cost of renting or using a specific type of AI chip will increase or decline. Companies that buy GPU capacity on demand may use the contracts to reduce exposure to future price spikes, while those with access to compute capacity may use them to manage revenue risk.
A consistently positive funding rate can indicate that long demand is stronger than short demand, suggesting optimism about spot GPU pricing. This is similar to patterns seen in cryptocurrency perpetual markets, though the underlying asset is very different. Here, the exposure is not to bitcoin or ethereum, but to the price of computing power used in AI workloads.
The arrival of these contracts also reflects the changing economics of AI infrastructure. GPU capacity is expensive, supply can be tight and rental prices can move sharply depending on chip availability, data-center constraints, model-training cycles and demand from large technology companies. For firms with large compute needs, unpredictable pricing can complicate budgeting and product planning.
Regulated futures remain pending
CME Group and Intercontinental Exchange are preparing their own compute futures, according to Bernstein, but those contracts are still awaiting regulatory approval. The planned products are expected to be cash-settled, meaning participants would settle gains and losses in money rather than deliver hardware or provide physical compute capacity.
That design is important because GPU hours are difficult to deliver in the same way as oil barrels, wheat bushels or metal inventories. The quality of compute capacity can vary by chip type, region, data-center provider, latency, software stack and contract terms. Cash settlement allows the market to reference a benchmark price without requiring traders to move actual hardware or allocate specific cloud resources.
If approved, CME and ICE contracts could give the market a more standardized and regulated way to trade AI compute exposure. That may appeal to companies that are unable or unwilling to trade on offshore venues. It may also help build benchmark prices for a market that is still fragmented across private contracts, cloud platforms, brokers and data-center operators.
However, regulated futures will need reliable indices. That is one of the central challenges in turning AI compute into a widely traded financial product. Unlike oil or gold, GPU capacity is often sold through negotiated agreements that are not publicly visible. Many large buyers secure supply through long-term cloud contracts or private data-center arrangements, while smaller users may pay spot rental rates on platforms that vary by geography and availability.
Prediction markets add forward pricing
A separate layer of trading has emerged through prediction markets. Kalshi, which is regulated by the CFTC, has listed event contracts tied to GPU rental costs. These contracts allow traders to take positions on specific outcomes, such as whether the rental price of Nvidia’s B200 chip will average above a given level at a future date.
One example cited by Bernstein is a contract asking whether Nvidia’s B200 chip would average more than $7 an hour by the end of 2026. If that contract trades at $0.40, the price implies a 40% probability that the outcome will occur, before considering fees and market frictions.
When several such event contracts are available across different price levels and dates, they can be combined to create an implied forward curve for GPU rental costs. This gives market participants a view of where traders believe compute prices may be headed.
Kalshi launched compute forward curves for Nvidia’s B200, H200 and A100 chips on July 14. According to Bernstein’s data, the B200 forward curve stood at $5.41, below a $7.39 historical benchmark. That gap suggests that traders on the platform were pricing in lower future rental costs than the benchmark level at that time.
The difference may reflect expectations for greater chip supply, expanding data-center capacity, improved efficiency or weaker-than-expected demand growth. It may also reflect limited liquidity and the early stage of the market. Thin trading can cause prices to move sharply and may not always represent broad commercial demand.
Why compute looks more like electricity
Bernstein compared compute pricing more closely to electricity than to oil. The reason is storage.
Oil can be stored in tanks, ships and underground facilities. That storage capacity affects forward pricing because traders can buy oil today, store it and sell it later if the economics make sense. GPU hours are different. An unused hour of compute capacity cannot be stored and used next month. Once the time passes, the capacity is gone.
That makes AI compute closer to electricity, where real-time availability and demand conditions matter heavily. A GPU sitting idle during a low-demand period may represent lost revenue for the owner. A shortage during a high-demand period can cause prices to rise quickly, especially if many AI developers need capacity at the same time.
Forward curves in this market therefore reflect expectations about scarcity, utilization, chip availability and future demand, rather than the storage costs that often shape commodity markets. If traders expect a wave of new data centers and chips to come online, forward prices may fall. If they expect supply bottlenecks, power constraints or a new surge in AI model development, forward prices may rise.
The comparison to electricity also highlights why regional pricing matters. Compute capacity in one region may not be equivalent to capacity in another. Power costs, data-center cooling, network connectivity, export controls, latency requirements and local regulations can all influence pricing.
Hedging demand is growing
The main commercial use case for these derivatives is risk transfer. Companies that need large amounts of GPU capacity may want to cap their future rental costs. Data-center operators, cloud providers or firms with contracted compute supply may want to protect revenue if prices decline.
For example, an AI developer planning a major training run several months ahead may face uncertainty over the cost of GPU hours. If rental prices rise sharply, the project may become more expensive than planned. A derivative contract can help offset that risk by increasing in value when the underlying compute price rises.
On the other side, a company that expects to sell compute capacity may be worried that prices will fall before its facilities are fully utilized. A short position in a compute futures contract could help reduce that downside.
This does not eliminate operating risks. Companies still must manage hardware availability, power supply, cooling systems, networking, software reliability and contractual obligations. But derivatives can separate some pricing risk from the physical business of running infrastructure.
That separation is a major step in the financial development of AI infrastructure. It allows compute pricing to become a standalone market exposure, distinct from the ownership of chips or data centers. Traders can take positions without buying machines, while corporate users can manage price swings without taking physical delivery of hardware.
Liquidity and benchmarks remain weak points
Despite the excitement around the new products, Bernstein said liquidity remains limited and current activity is largely speculative. That is not unusual for a new derivatives market, but it does raise questions about how reliable early prices are.
A liquid market needs a broad mix of participants, including commercial hedgers, market makers, proprietary trading firms and end users with real exposure to the underlying asset. At this stage, the market is still developing that base.
Benchmark construction is another major limitation. GPU capacity often trades through private agreements rather than transparent exchanges. That makes it difficult to determine a reliable market price. Two buyers may pay different rates for similar chips depending on volume, region, contract duration, service quality and relationship with the provider.
Silicon Data, which is working with CME, collects about 150,000 verified price records each day from 50 regions and as many as 100 trading venues, according to Bernstein. The firm aims to translate GPU usage hours, including for chips such as Nvidia’s H100, into standardized assets that can support futures contracts.
Ornn, which supplies indices for ICE and Kalshi, uses negotiated transaction prices to build its benchmarks. That approach may capture real market activity, but it still depends on the depth, consistency and representativeness of the underlying data.
The development of stronger indices will likely determine how quickly regulated compute futures can scale. Without trusted benchmarks, traders may be reluctant to use the products for serious hedging. With better benchmarks, the market could become more useful to businesses managing large AI infrastructure budgets.
Bitcoin miners shift toward AI infrastructure
The rise of compute derivatives is also linked to changes in the digital-asset mining sector. Some bitcoin mining firms have been repurposing power contracts, land and data-center expertise for AI and high-performance computing.
Mining companies already operate energy-intensive facilities and often have access to large power supplies. As mining economics fluctuate, some have looked to AI compute as a new revenue source. One large mining company has projected billions of dollars in annual revenue from computing facilities, underscoring how quickly the sector’s business model is changing.
This shift could increase the supply of AI compute over time, especially if more mining sites are converted into data centers capable of hosting advanced GPUs. However, the transition is not simple. AI data centers often require different networking, cooling, reliability standards and customer relationships than mining facilities.
Still, the overlap between power infrastructure and computation is becoming more important. AI workloads require enormous electricity consumption, and power availability is now one of the biggest constraints on data-center growth. That makes compute pricing sensitive not only to chip supply, but also to grid connections, energy contracts and construction timelines.
A market still in its early stage
The compute-derivatives market is not yet close to the scale of established commodity markets. Crude oil futures, for example, trade hundreds of millions of contracts each year globally. AI compute contracts are at a much earlier stage, with limited liquidity and evolving standards.
Even so, the direction of travel is clear. As AI becomes more central to corporate spending, the cost of computing power is becoming too important to leave unmanaged. Traders are beginning to treat GPU hours as a price risk that can be hedged, traded and benchmarked.
Mansour, cited in the source material, said the broader technology sector needs clearer derivatives markets to manage volatile hardware and compute costs. He said forward pricing models could help buyers calculate future daily expenses with greater confidence.
That view reflects the practical problem facing many AI companies: compute demand is rising, but prices can be unpredictable. If derivatives markets mature, firms may gain better tools to plan budgets, finance capacity and smooth the impact of price swings.
For now, the market remains experimental. Offshore perpetual futures, Kalshi event contracts and planned CME and ICE futures are early attempts to build financial rails around AI infrastructure. Their success will depend on regulation, liquidity, index quality and whether real commercial users adopt them for hedging rather than leaving activity mostly to speculative traders.
Bernstein’s disclosures state that the analysts maintain long positions across multiple digital assets.
Learn how prediction markets work with crypto-style hedging in this guide to Toobit event contracts and refine your GPU pricing strategies.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
