GPUs have become a strategic resource. Goldman Sachs expects AI investment to reach about $1T globally in 2026, including $581B in the US, and counts GPU rental prices among its leading indicators for that spending. BCG expects the AI compute market to grow from $360B in 2025 to $2.3T by 2030. But GPU capacity is hard to price and hard to trade. Rentals differ by chip, location and contract length, and each price index handles those differences in its own way.
CME planned to list H100 and B200 rental futures, settling to Silicon Data's indices, on 5 October; the CFTC extended its review, so a launch is now expected in November at the earliest. Onchain markets have offered exposure for most of the past year: rental-rate contracts based on Ornn's index, yield-bearing tokens backed by GPU loans, and perpetual contracts on memory and chip stocks.
Figures to 30 September 2026 unless stated.
Since early 2026, Kalshi and Polymarket have priced where spot GPU rental rates are heading. In mid-July, with the B200's spot rental price above $7, Polymarket gave a 16% chance it would still be $7 or more at year-end, and the spot price did fall back, to $5.68 by August. On 8–9 September, before the price crossed $7 again, the odds rose from 22% to 58%, and they reached 78% by the end of the month. Kalshi's B200 contracts trade too little to give a clear read.
However, very little money sits behind these bets. Trading peaked at $3.2M in August and fell to $1M in September, and Polymarket's open interest was about $0.2M at the end of September. These markets have also not given strong signals until close to resolution. On 1 September, Polymarket gave about 10% odds that the B200 would end the month at $7.50 or more; it ended at $8.09, and the odds only rose above 50% after the price crossed $7.50 on 20 September.
USD.AI and GAIB each issue a dollar token that pays nothing and a staked token that earns interest from GPU financing (sUSDai and sAID). Year to date, sUSDai's value has grown 75% to $541M, while sAID's has fallen 14% to $16.2M.
sUSDai is a share in an onchain vault that lends to GPU operators. Since loan contracts launched in December 2025, the vault has lent $286M to 17 borrower addresses, with about $266M still outstanding. One borrower address has taken $132M, and the five largest loans make up 77% of the total.
sAID is backed by an offchain portfolio of GPU and robotics financing, and its value is a net asset value that GAIB publishes periodically.
Year to date, sUSDai has returned 7% annualized and sAID 9.7%, measured by the rise in each token's value. Both take about a month to redeem; holders who want out sooner sell on exchanges.
sUSDai is used widely in DeFi to earn more than the yield from its GPU loans. Four protocols (Fluid, Pendle, Jupiter Lend and Morpho) hold 61% of its supply and enable two main strategies.
1) Borrowing against it. As of 30 September, $159M was borrowed against sUSDai across Fluid, Jupiter Lend and Morpho, $123M of it on Fluid alone. The typical trade is a loop: post sUSDai, borrow dollars, buy more sUSDai, and repeat. At an 80% loan-to-value ratio, roughly the average across existing positions, an investor can borrow $0.80 for every $1 of sUSDai deposited. Repeating the loop means $1 of initial capital can support roughly $5 of total sUSDai exposure.
The extra return over simply holding sUSDai comes from the gap between sUSDai's yield and the borrow rate. On 30 September, sUSDai yielded 7.3% while USDC cost 6.5% to borrow on Fluid (Ethereum). At 80% LTV, that produced an annualized return of about 10% on the investor's original capital.
The trade carries two key risks: a higher borrow rate can eliminate the extra return, while a fall in sUSDai's market price can trigger liquidation.
2) Locking a fixed rate on Pendle. A holder can deposit sUSDai into Pendle and split it into two pieces: a principal token that can be redeemed at full value at maturity, and a yield token that represents the sUSDai yield until then. Selling the yield token and holding the principal token locks in a fixed return. For example, as of 30 September the principal token maturing on 25 February 2027 implied a 10.6% annualized yield.
If sUSDai's actual yield rises above 10.6%, the holder of the yield token captures the upside. If it falls below 10.6%, the fixed-rate holder benefits from having locked in the higher rate. The principal token can also be bought directly, so a buyer does not need to hold sUSDai first.
Onchain trading of AI-hardware stocks happens mostly through perpetual contracts. On Hyperliquid, perps on memory, chip, AI-cloud, server and AI-lab names traded $22B in September and $174.5B from January to September. Memory-chip stocks made up 76% of that, and SK hynix alone 30%: $52.9B across two contracts, SKHX on the Seoul share (000660) and SKHY on the US depositary share. That is more than all non-memory chip stocks in this set combined, NVIDIA (NVDA) included.
Traders now pay much less to stay long these stocks, even as the price of renting compute kept rising. Weighted by open interest, longs paid 27% a year in funding in June and 6% in September, and every segment fell, while open interest rose by two-thirds and the B200's rental price rose 83%. Memory longs paid 4.5% in September, down from 35% in June. The highest rate in September, 17%, was on perpetuals tracking the pre-IPO valuations of Anthropic and OpenAI, an exposure no traditional exchange offers. Server makers like Dell (DELL) and Super Micro (SMCI) were the only exception: more traders bet on their prices falling than rising, so short sellers paid 6% a year.
Until CME lists, onchain markets are the only venues where investors can take a view on compute or earn a yield from it.
The exposure with real size is GPU credit and stock perpetuals, but neither can be a real hedge of the rental cost. sUSDai finances the GPU operators, so a higher rental rate can make its loans safer and later loans pricier, but interest already agreed does not rise with it. Stock perps can rise when GPUs are scarce and offset some of a higher rental cost, but the gain will not perfectly match the rise in the hourly rate, and the long also has to pay funding. The position that pays out directly when a lab's GPU invoice rises is a rental-price contract, and onchain that market is still too small to hedge with.
If CME lists, it will be the first venue that could hedge the rental bill in size.
Looking to use Dune for your company?
