TrackingtheAItradeinpredictionmarketsandperpetuals

CME Group and Silicon Data, a GPU pricing benchmark provider, plan to launch H100 and B200 Rental Index futures on 5 October, pending regulatory review, with each contract representing a month of rent for one GPU. The stated purpose is to let AI builders and hyperscalers hedge the cost of compute.

Valuations across the AI complex have risen steeply, with Nvidia's market value close to four times its end-2023 level. The harder question is the spending, which is large and committed years in advance. Goldman Sachs Research estimates global AI investment near $1 trillion in 2026 and about $1.8 trillion cumulatively since 2022. Much of it runs through circular arrangements, in which chipmakers and cloud providers invest in AI developers that then buy their products, so a demand shortfall hits the investing company twice. On 17 August, ECB staff published research arguing that a correction is likely, while noting that its timing cannot be known in advance and that boom-bust patterns are identifiable only with hindsight.

July gave a preview of what this could look like. The KOSPI, South Korea's benchmark equity index, fell a record 33% over the month, led by SK Hynix, the Korean maker of the high-bandwidth memory used in AI accelerators, and Samsung Electronics, which together are more than half that index. Situational Awareness, an AI fund running about 4x leverage, took a 67% drawdown and sold the bulk of its equity book to Citadel after margin calls. Jane Street, an investor in the fund, told employees it lost $15 billion in July and that its short-dated puts helped little, because the selloff was spread across the month. Its note put the largest memory and semiconductor stocks down around 50%.

If these patterns are only identifiable with hindsight in price, positioning is the other place to look. The same names trade onchain as perpetual futures, where funding, open interest and liquidations are observable continuously.

The exposure is not distributed the way the AI narrative is. Memory and storage carry roughly seven times Nvidia's open interest, and traded $7.16 billion in the week to 16 August against Nvidia's $219 million. That is also where the physical constraint sits: J.P. Morgan calls memory the latest bottleneck in AI expansion and estimates DRAM prices will have risen more than 400% between the start of 2024 and the end of 2026.

What to look for: the memory and storage band relative to everything else, and how far it has grown since December 2025.

Part of what trades here has no listed equivalent anywhere: the Chinese AI labs Zhipu and MiniMax, the robotics manufacturer Unitree, Kioxia, the privately held Japanese flash memory business, and ChangXin Memory Technologies (CXMT), the state-backed Chinese DRAM maker, before it listed in Shanghai in July. Anthropic and OpenAI were tradable the same way between late November 2025 and mid-June 2026.

Trading tends to follow the hours of the underlying market, though not evenly. Zhipu and MiniMax do 60% or more of their volume during Seoul and Shanghai business hours. SK Hynix and Samsung do 42%, which leaves most of their trading happening while the Korean market is closed. Micron, a memory maker in the same business, is the mirror image at 53% in US cash hours and 15% in Asian ones, and Nvidia sits at 61% and 10%. CXMT is the outlier: a quarter of its volume trades at the weekend, when no cash market anywhere is open.

What to look for: the Asian and US bands invert as you move down the list, and how much of the Korean and Chinese volume falls outside every cash session.

As Korean equities fell, Hyperliquid saw a wave of forced closures concentrated in the memory names. On 27 July, liquidations in SK Hynix alone were nine percent of the day's volume, at $86.25 million.

What to look for: two clusters, 13 to 15 July and 27 to 31 July, and the near-absence of any Nvidia bar.

Funding moved just as violently, and it shows how fast positioning turned. Positive funding means longs are paying shorts to hold the position open; negative means shorts are paying longs. SK Hynix longs were paying 297% annualized on 13 July. Two days later shorts were paying 63%. The same reversal ran again at month end, with open interest falling by a third in a single day.

CXMT moved furthest, with shorts paying through the last week of July. Its open interest had tripled to $93m over the previous fortnight while the price fell 20%, and the price then rose 30% in five days. That fits a crowded short being squeezed. It also fits the pricing reference changing when ChangXin listed in Shanghai the same week. The data does not separate the two.

What to look for: the sign reversals in SK Hynix and Samsung, CXMT's sustained excursion below −400%, and Nvidia holding a narrow band throughout.

The leverage in the AI trade was concentrated in memory. Nvidia's worst liquidation day was $971,000 and its open interest stayed inside a narrow band all month, so anyone watching Nvidia for signs of strain would have seen a quiet one.

Renting an H100 costs more now than it did in March. On Kalshi, which lists contracts on the hourly price, the implied rate has gone from about $1.70 to about $2.75, a rise of roughly 55%. The newer B200 and H200 are up nearer 80%. Both have come off their late-July highs: the Ornn B200 index fell 13% between 20 July and 17 August, while the H200 ended that window where it started.

The A100, H100, H200 and B200 are the Nvidia data-centre GPUs that most AI training and inference runs on, so their rental rate is the unit cost of the buildout. A rising rate means demand is still outrunning the supply of chips and data centres coming online. A falling rate would mean capacity is arriving faster than it is being used, which is the point at which the spending becomes hard to justify. Goldman treats GPU rental and memory prices as leading indicators for exactly that turn, and reports all of its indicators near the top of their post-2022 range.

What to look for: the direction since March across all four generations, and the pullback from the late-July high. Solid segments are consecutive weekly readings; dotted segments span weeks where the strike ladder supported no reading. Built only from strikes that traded.

Polymarket lists the same index as a set of year-end brackets, each paying out if the price finishes inside a given range, so the market prices a whole range of outcomes at once. Most of the weight sits below where prices are now, but participants are far from predicting a collapse. The Ornn index put the B200 at $6.17 an hour and the H200 at $4.99 on 17 August. The market gives the B200 a 65% chance of finishing the year below $6.00, the H200 a 60% chance of finishing below $5.00, and the B200 only a 5% chance of ending 2026 below $3.50, a fall of about 40%. That last probability has never exceeded 13% since the market opened in July.

What to look for: how low the below-$4.00 line stays throughout. Brackets grouped into $1 bands.

These markets give a continuous read on what it costs to rent the input the AI economy runs on. That is the reading that would have been useful in the late 1990s, when the fiber-optic buildout ended not with a change in the story about demand but with a collapse in the price of the capacity being built, which took the leveraged carriers that had financed it down with it.

There is real money on what these companies are worth and almost none on whether the spending produces revenue. Across both prediction market venues, contracts on company valuations have traded $230m. The contracts on Nvidia's data-centre revenue and gross margin and Broadcom's AI revenue, which are the ones that would settle whether the capex is working, have traded $124,000.

What to look for: the distance between the top row and the bottom two.

The same imbalance separates the perps from the prediction markets. Memory and storage is the largest position on Hyperliquid, at $7.16bn of volume in the week to 16 August. Kalshi's markets on the price of memory itself, which opened on 1 August, have traded $13,000. The exposure people take is to the security rather than to the economics underneath it.

Neither venue is deep enough to hedge institutional size in compute, which is the gap the CME contract is being listed into.

Anyone tracking AI concentration risk through the headline names would have seen almost nothing in July. Nvidia, Microsoft, Alphabet, Meta and Amazon each liquidated between $1.8m and $4.1m across the whole month, and their funding never left a band of roughly −4% to +28% annualized. SK Hynix liquidated $153m and its funding ran from −156% to +308%. Three venues offer a better way to track the AI trade and its risks.

On Hyperliquid, track funding and open interest in the memory and storage names. This is where the leverage sits, which is why July's unwind was legible here and invisible in Nvidia. Funding has averaged −9.6% annualized over the past thirty days, against +23% to +43% before July, so the long crowding has inverted. Funding turning positive again on rising open interest would mean leverage is rebuilding.

On Kalshi, track the GPU rental ladders. A sustained fall would be the first sign of the oversupply that ended the fiber buildout. Kalshi also lists CPI and Fed funds contracts running back to 2021, which price the rates that AI cash flows get discounted at.

On Polymarket, track the market-cap brackets and the data-centre revenue and margin contracts. Watch their liquidity rather than their price. If those contracts start trading at size, it means investors think the capex question is worth taking a position on.

Three dates are worth watching between now and the end of October. Nvidia reports on 26 August, and the Polymarket contract on that print resolves the same day. The CME and Silicon Data compute futures list on 5 October, settling on a different benchmark from the Ornn Index both prediction markets use, so there will be two competing reference prices for the same commodity. The hyperscalers report third-quarter results in late October, with capex guidance the main thing being priced.

Dune's prediction markets dataset covers Polymarket and Kalshi with granular, comprehensive and normalised trade data across both venues.