How much is the hourly rental price for a single Nvidia B200 GPU by the end of this year?

The prediction market breaks this question into a set of yes-or-no contracts. Traders think the B200 rental rate will be above a certain price, and then bundle contracts at different price levels and different dates together, forming a GPU rental curve shaped by market bets.

Polymarket has already launched GPU rental contracts before, but trading volume was low. This time, Kalshi co-founder Tarek Mansour announced that the platform has extracted a forward curve from prediction market prices for B200, H200, and A100.

Prediction markets no longer just answer election results, when rate cuts will happen, and corporate events—they are starting to build financial markets for an industry that has never had a publicly traded market.

This curve is still a long way from the forward curves in traditional commodities markets, and prediction markets also can’t allow buyers to receive a package of GPUs after expiration. But it captures what the GPU trading market lacks the most: a visible price benchmark line that everyone can see.

In recent years, capital has kept flowing into chips, data centers, and power, making compute capacity one of the biggest costs in the AI industry—yet procurement methods still rely on phone reservations, backchannel deals through acquaintances, and offline contracts.

The financialization of compute is underway in full force.

First comes contracts, then comes markets

Before large models emerged, businesses could obtain compute capacity mainly in two ways: buying servers themselves, or paying a relatively stable monthly fee to cloud providers. AI changes this procurement logic. Training and inference consume huge numbers of GPUs; prices begin to diverge across different chips, regions, and contract tenors, and cloud providers’ quotes also change rapidly with supply and demand.

Today’s compute market is not without forward trading.

Large labs lock in future capacity in advance. Neocloud pre-purchases GPUs from cloud providers and brokers. Between hyperscalers, resources are also reserved for each other. Contracts can be as short as hourly rentals or cover several years. They perform a function similar to long-term oil supply agreements—only the price is hidden in private negotiations.

A large inference services provider once described purchasing compute capacity as “finding a middleman who knows the supply.” You tell them what chips you need, how many cards, and what region you’ll use them in. They then look for inventory for you through their network of relationships. Brokers profit from information asymmetry, large holders profit from allocating capacity through relationships, and the true transaction price rarely appears on public screens.

This market can enable delivery, but it cannot form continuous expectations. AI labs don’t know what inference costs will be in half a year. Data centers can’t lock in rental prices in advance. Lenders also lack real-time updated data to assess how quickly the pledged GPUs will depreciate.

The size of capital no longer allows this kind of pricing approach to remain unchanged.

The numbers Tarek provided are that hyperscalers will invest more than $700 billion in compute this year, and the market size could reach $7 trillion to $10 trillion by 2030. More cautious institutions predict similarly large figures. Morgan Stanley expects global data center capital expenditures of about $2.9 trillion by 2028, of which about $2.5 trillion will serve AI. McKinsey estimates data center capital expenditures before 2030 at $6.7 trillion. Goldman Sachs puts AI infrastructure investment from 2026 to 2031 at $7.6 trillion.

These numbers use different years and statistical methodologies. Some include data centers, and some count both compute and power. The commonality is that compute and hardware each account for 55% to 67% of each firm’s estimate—the largest portion of this round of infrastructure investment.

Chips are also a kind of asset with sharply changing prices. Market estimates for GPU useful life range from three years to seven. New product generations refresh performance every year, and although supply is tight, older chips still retain rental value. Data centers need to tie enormous capital up in a batch of devices whose depreciation rates have not yet reached a consensus.

The heavier the burden of capital, the more important forward prices become.

Exploring a GPU trading market

The first stage of a GPU trading market is “private matching,” which has been running for years.

Buyers reserve capacity in advance; sellers lock in future revenue. Brokers handle the work of finding supply and matching trades. Real demand and forward commitments already exist—it’s just that there is no unified contract and no publicly quoted prices.

These buyers and sellers also form the underlying layer of the compute financial market.

Hyperscalers, large data centers, and GPU holders have inventory and worry that future rental rates will decline. AI labs, inference platforms, application companies, and neocloud—already committed to capacity downstream—need continued procurement and worry that future prices will rise. One side wants to protect device revenue, the other wants to control compute costs. That’s how the initial demand for trading emerges.

The second stage is building standardized price indexes. Ornn’s Compute Price Index extracts prices from actual GPU rental transactions, covering multiple mainstream chip types. Silicon Data publishes daily on-demand rental indexes for H100, A100, and B200, then sends the data into the Bloomberg terminal. Compute Desk is also building similar products.

The indexes defined by index providers are not just a string of numbers. Which chips, regions, network configurations, and contract types will be included, how abnormal trades are handled, how old indexes exit after chip transitions—all of these will change what the market calls the “GPU price.” Exchanges provide the trading venue, while the index provider defines what people are actually trading.

Ornn recently received a $33 million investment from a16z. Whoever can organize fragmented trading data into a benchmark accepted by the market may get a chance to become the “price entry point” in the compute market.

The third stage is writing indexes into tradable contracts. CME chose Silicon Data as the data provider and plans to launch compute futures settled daily based on GPU rental benchmarks. ICE, the parent company of the New York Stock Exchange, chose Ornn and is preparing to launch another set of GPU futures. Both traditional exchanges position their products as risk management tools for AI labs, cloud providers, data centers, and financial institutions—but the products are still in the stage of awaiting regulatory review.

Prediction markets took a different route. They keep asking traders the same question: “Will a certain chip, on a certain date, be higher than a certain rental price?” By calculating the price difference between adjacent thresholds, you can roughly approximate market judgment about that price range. Then repeat the calculation along different dates, and a term structure appears.

Traditional commodity markets usually define deliverable contracts first, and then futures trading forms the curve. Prediction markets form public expectations using event contracts first, and then consider using this expectation to serve over-the-counter trading, futures, and perpetual contracts.

Traditional compute futures are still waiting for regulatory approval, while prediction markets have already delivered the term structure.

What a curve can solve

After spending effort building indexes, futures, and forward curves, what problem can this really solve for a typical AI company?

Suppose an inference platform has already agreed to provide service for customers six months later. It knows that it will need a batch of GPUs then, but it doesn’t know how high rental prices will rise. If rental prices suddenly increase, the customer prices already negotiated won’t change with it. It can only absorb the added costs itself. By buying a contract whose value rises along with GPU rental prices, the cloud bill becomes more expensive, but the contract’s gains can offset part of the difference. Data centers face the opposite problem: they already bought the equipment; if future rental rates fall, revenue shrinks. By selling forward exposure, they can lock in part of device revenue early.

This contract does not need to be exactly identical to every individual card a company actually purchases. A company might use H200 in the New York region, while what trades in the market is an H200 index covering multiple suppliers. As long as the two prices roughly move up and down together, the contract can still work. Analysts estimate that when the correlation coefficient between the two prices reaches 0.7, well-designed hedging can eliminate nearly half of the volatility. Airlines can’t buy a contract exactly matching every single jet fuel expense they have, yet they still use crude oil futures to control costs. The logic is similar.

Lenders also need this curve. When a data center borrows using GPUs, the bank must judge how much rental income those chips will generate two years later. In the past, lenders could only rely on manufacturer quotes, scattered transaction results, and their own assumptions. With a public curve, lenders can adjust collateral valuations as the market changes, and data centers can more easily prove future revenue.

Prices can even influence which chips companies choose. If Nvidia GPU rentals trade more, the index is more reliable, and hedging instruments are more active, banks will be more willing to accept it as collateral, and holders will find it easier to rent or sell when they need to. Device financing becomes easier, and more buyers will continue to choose it. The trading liquidity that clusters around Nvidia could become an advantage that competitors struggle to replicate.

A price curve therefore serves more than just traders. It helps users know costs earlier, helps holders lock in revenue earlier, and also makes it easier for lenders to price devices and data centers.

Bottlenecks & challenges

The first problem is the index.

Ornn emphasizes real executed trades. Silicon Data focuses on on-demand rental prices, and other approaches also standardize energy costs. Each method retains some information while discarding other parts. No index can simultaneously cover chip types, regions, tenors, network parameters, and counterparties.

At the same time, chips will evolve quickly.

Oil measurement can stay valid for a long time, but the GPU market upgrades from H100 to H200, B200, GB200, and Rubin. AMD, Google TPU, Amazon Trainium, and custom chips further divert compute demand. When old indexes retire and how new and old chips connect will continuously change the underlying asset the contracts reference.

The second problem is delivery.

With cash-settled contracts at expiration, only money is accounted for—no GPUs are delivered. Companies that want to control costs can offset rental increases using contract gains, but neocloud, which has already committed capacity to customers, still has to go find cards in the market on its own.

Another risk comes from trade volume.

The actual transaction volume publicly recorded each day in the GPU rental market may be very small, and supply is concentrated among a few sellers. This means a single trade could significantly move the index, and the party holding the supply can more easily influence the final settlement price.

This is also a problem with curves drawn using prediction markets.

Traditional forward curves rely on delivery or spot exchange mechanisms to pull futures prices back toward real supply and demand. Prediction markets’ binary contracts lack that channel. The curve represents participants’ expectations rather than yet becoming a capacity price that is deliverable and arbitrage-able.

The third problem is liquidity.

Sellers prefer long-term contracts because data centers want to lock in revenue early; buyers prefer short-term contracts because AI companies need flexibility to switch chips and suppliers. The demand on both sides regarding tenors is naturally misaligned. Brokers and large holders also profit from an opaque market, so they lack incentives to proactively move all trades into a public market.

Despite all these obstacles, the demand for public pricing in the compute market will not reverse. Perhaps not long from now, we’ll see reports about some “smart money” going “5x long compute” on-chain.

Click to learn more on律动BlockBeats for job openings

Welcome to join the official律动 BlockBeats community:

Telegram subscription group: https://t.me/theblockbeats

Telegram chat group: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia