Kalshi Compute Markets Explained: How to Trade GPU Prices

The rapid expansion of artificial intelligence has created demand for an input that remains unfamiliar to many financial-market participants… compute. Or in other words the processing capacity required to train and operate artificial intelligence models.

Kalshi is now attempting to transform the price of that capacity into a tradable financial variable. The prediction market has introduced event contracts tied to the hourly rental prices of several Nvidia GPUs and has used activity in those contracts to construct forward curves showing how traders expect prices to develop over time.

The initiative represents a notable departure from the political, economic and sporting events commonly associated with prediction markets. It also reflects a broader effort to build financial infrastructure around the artificial intelligence economy.

Kalshi co-founder and CEO Tarek Mansour recently told The New York Times that compute markets were among the areas he found most exciting. His interest is consistent with the company’s increasingly ambitious description of compute as a commodity. When Kalshi formally announced its forward curves in July, Mansour argued that compute was becoming comparable to oil because businesses need a mechanism through which to price and transfer future risk.

Of course, a barrel of oil is a physical commodity with mature standards for grade, location and delivery. A GPU-hour is less uniform. Its economic value depends on the chip model, hardware configuration, data-center location, software environment, availability and contractual terms. Compute markets are therefore attempting to standardize a market that remains fragmented.

What Does Compute Mean?

Compute is a general term for the processing resources used to perform calculations. Within artificial intelligence, the term usually refers to access to GPUs and related data-center infrastructure used for two principal activities.

Training is the process through which a model learns patterns from large datasets. Inference is the process through which a trained model responds to a prompt or produces an output. Both can require substantial processing power, although their workloads and cost structures differ.

GPUs were originally developed to process computer graphics, but their parallel architecture also makes them particularly effective for artificial intelligence. Nvidia explains that GPUs can process many operations simultaneously and can scale across large systems, two features that are important for modern model training and inference.

Compute is often priced in GPU-hours. One GPU-hour represents access to one specified GPU for one hour. It is a unit of capacity, not a claim that all GPU-hours are economically equivalent. An hour on an Nvidia A100 does not provide the same performance or command the same market price as an hour on an H200 or B200.

Rental prices can change because of demand from AI laboratories, the delivery of new chips, electricity and data-center constraints, regional availability, export restrictions and the introduction of more efficient hardware. Compute is also perishable. An unused hour of GPU capacity cannot be stored and sold later, which can place pressure on providers to reduce prices when utilization falls.

How Kalshi’s Compute Markets Work

Kalshi currently lists contracts tied to Nvidia A100, H100, H200, B200 and RTX 5090 rental prices. The contracts can cover hourly observations, monthly averages and longer-dated outcomes.

They are binary event contracts. A trader does not purchase computing capacity, reserve a GPU or receive physical delivery. Instead, the trader buys a Yes or No contract on whether a specified price index will exceed a stated threshold at settlement.

Consider a hypothetical contract asking whether the hourly price of H200 compute will be above $4.50 at the end of August. A Yes contract settles at $1 if the official value is above $4.50 under the contract’s rules. It settles at $0 if the condition is not met. The No side has the opposite result.

The outcome is generally verified using the Ornn Compute Price Index, or OCPI. Ornn describes the benchmark as a family of transaction-based indices that measure the clearing prices of rented GPU compute in dollars per GPU-hour. It uses executed rental transactions rather than a single cloud provider’s advertised price. Kalshi’s individual market rules determine the precise index, observation window, calculation and settlement time.

This means a trader may see one price on a large cloud platform and another on the Ornn index because the two figures measure different segments of the market. The contract settles according to its written rules, not according to the price that seems most representative to an individual trader.

Kalshi’s AI Compute Markets

What Is a Compute Forward Curve?

In July, Kalshi introduced forward curves for the Nvidia B200, H200 and A100. A forward curve plots market-implied prices across future dates. In practical terms, it attempts to answer how much traders expect one hour of a specified GPU to cost next week, next month or later in the year.

Kalshi derives these curves from trading in its weekly and monthly chip-price contracts. The curve is not itself a tradable asset. It is a reference created from the probabilities embedded in the underlying binary markets.

The structure can be understood statistically. If multiple contracts estimate the probability that a future price will exceed several thresholds, those probabilities describe portions of the expected price distribution. Taken together, they can be used to estimate an implied future price.

A curve that rises with maturity suggests traders expect future prices to be higher than current prices. A downward-sloping curve suggests traders expect present scarcity to ease. Kalshi Research reported in July that its B200 curve was in backwardation, with implied prices declining over time, while several older-chip curves were comparatively flat. That pattern was consistent with an expectation that B200 supply would expand after an initially constrained launch.

The curve should not be treated as a guaranteed forecast. It is a summary of current market prices and is only as reliable as the liquidity, information and settlement design of the contracts from which it is derived.

MNX and the Emerging AI Derivatives Market

Kalshi is not the only company attempting to create tradable markets for artificial intelligence infrastructure. MNX, founded by Manifold Markets co-founders Stephen Grugett and Ian Philips, is developing what it describes as an AI-focused decentralized futures exchange.

The distinction between the two models is substantial. Kalshi’s compute products are regulated binary event contracts with fixed expirations and outcomes. MNX plans to offer compute perpetuals, which are leveraged contracts designed to track assets such as H100 rental rates without a conventional expiration date. MNX also intends to list products tied to private AI-company valuations, public AI equities and model-performance benchmarks.

Grugett has argued that the AI economy lacks a unified venue through which market participants can trade or hedge exposure across the full value chain. Compute is central to that thesis because it connects chip manufacturers, cloud providers, data centers, AI laboratories and end users.

The Central Challenge Is Liquidity

The economic case for compute derivatives is plausible. AI laboratories want predictable costs, infrastructure providers want predictable revenue and lenders need reference prices with which to value equipment and long-term contracts.

The more difficult question is whether enough buyers, sellers and speculators will participate to create reliable prices. Compute remains heterogeneous, and a global index may not match the cost faced by a particular company in a particular region. Thin order books can also produce wide spreads and unstable implied probabilities.

Competition is increasing. Kalshi and MNX are developing different market structures, while established derivatives exchanges and specialized index providers are also entering the sector. No venue has yet demonstrated that compute derivatives can achieve the depth associated with mature energy or financial futures.

The emergence of these markets is still significant. The AI economy is beginning to develop not only new chips and data centers, but also the financial instruments required to allocate their risks. For traders, Kalshi’s compute contracts provide an accessible way to express a view on GPU rental prices.

For the broader market, their greater importance may lie in whether they can produce a credible forward price for one of the digital economy’s most important inputs.