Economic Observer reporter Wang Yajie

The hourly rent for the Nvidia H100 graphics card is being turned into a tradable financial contract.

This may sound far from ordinary people, but for big model companies, cloud vendors, and computing power rental providers, every fluctuation in GPU (graphics processing unit) rent may swallow up a portion of profits.

Training large models requires prolonged and sustained use of computing power. The increase in computing power has led to a rise in training costs for AI (artificial intelligence) companies; With the decline in computing power rent, investors who invest heavily in purchasing cards and building their own computing power clusters will face dual pressures of shrinking revenue and hardware asset depreciation.

CME Group and Silicon Data had originally planned to launch futures contracts linked to the rental prices of Nvidia H100 and B200 graphics cards, which were scheduled to be listed on the NYMEX (New York Mercantile Exchange) on October 5th. However, the regulatory review of the products was postponed due to "novelty or complexity" issues.

As of October 19th, there are still over twenty days left until the CFTC (Commodity Futures Trading Commission) completes its extended review of CME Group's computing futures. At present, the listing time has not been determined.

The Economic Observer recently interviewed top domestic computing power leasing companies, private computing power service providers, small and medium-sized computing power owners, central enterprise computing power purchasers, large model companies, cloud vendors, and AIDC (Artificial Intelligence Data Center) park operators, and found that even though this computing power futures is still in the stage of supervision and delayed review, the fact that computing power can be priced as futures has already pushed a real problem to Chinese enterprises - when computing power rental prices fluctuate dramatically like oil prices, who will bear the cost risk for AI enterprises.

Pathfinding

In New York, the hourly rental price of a single GPU computing power (commonly referred to as graphics card rent in the industry) has been "stuffed" into the design of futures contracts.

Two computing power futures designed by CME Group and Silicon Data are linked to the leasing prices of Nvidia H100 and B200, respectively.

But before the futures contract was listed, the CFTC extended the review period to November 9th. November 9th is only the review deadline, not the listing date. CME Group will change the initial listing date of the futures contract to TBD in subsequent relevant documents.

A person close to overseas computing power trading platforms said that in the past, GPU leasing was more like a "familiar market". Whoever has a card, whoever wants a card, the transaction parties negotiate the price in private contracts. Now the exchange wants to move this quote onto the screen of the open market, turning it into an open curve. The difficulty also lies here. When the CFTC launched a public opinion consultation earlier, it pointed out that the GPU rental market is "fragmented and opaque", and prices are mostly determined by bilateral contracts. Regulators are concerned that top entities may affect settlement indices and manipulate futures prices.

Not only CME Group is rushing in.

ICE (Intercontinental Exchange) has partnered with Ornn (Computing Price Index Company) to launch GPU computing futures based on OCPI (Ornn Computing Price Index). On September 29th, ICE announced that starting from October 5th, it will provide Ornn Compute Index reference data and hypothetical daily settlement prices on the trading platform, with the first batch of contracts covering H100 and B200 graphics cards. They are still waiting for the completion of regulatory procedures.

The one that came out earlier is Kalshi (the US prediction market exchange). What it does is not NYMEX style standardized futures, but GPU computing power price prediction contracts and forward curves. In March, Kalshi launched the first batch of GPU computing hourly price prediction contracts. On July 14th, it launched a computing power forward curve covering B200, H200, and A100. Kalshi CEO Tarek Mansour previously publicly stated that computing power is the 'new oil'.

The price fluctuations have already been recorded on the books. According to publicly available data from Silicon Data and statistics from multiple industry platforms, the rental price of H100 graphics cards was at a high of about $8 per GPU per hour in the first half of 2024. After 2025, its rental price continued to decline, and some platform quotes have fallen below $2. According to Kalshi's relevant data, the hourly rent for B200 graphics cards reached $6.11 on May 30th and dropped to $4.22 on June 21st, a decrease of about 31% in less than a month.

This is exactly where financial institutions want to enter.

Architect (an encrypted derivatives platform) is also promoting GPU computing power "perpetual futures". It was founded by Brett Harrison, former president of FTX US, and its products track GPU daily rent and DRAM (Dynamic Random Access Memory) prices. Nodal Exchange, on the other hand, entered the electricity market and announced in September that it would collaborate with Compute Desk to launch GPU computing futures.

Overseas platforms are exploring, while Chinese companies are experimenting.

Zhang Lei, the business leader of a leading domestic computing power leasing company, told the Economic Observer that they have been paying attention to the trends of overseas GPU futures and derivatives, but have not directly copied them. He said, "There is no mature public computing power pricing index in China, and a large number of transactions in the industry are bilateral contracts negotiated privately, so the risk can only be absorbed between the supply and demand sides

Hangzhou private computing power service provider Zhou Bo has also seen Kalshi's GPU rental forward curve. He had previously locked in a batch of A800 graphics cards, but their quota was cut by 30% and the price increased by 15% before delivery. That loss can only be borne by oneself. Zhou Bo said that the overseas curve can be used as a market reference, but in order to truly fight against upstream quota cuts and temporary price increases, at this stage, we still rely on multi-channel stocking, safety stock, and business contracts.

Anxiety of buyers with high computing power

Zhang Lei's company used to rely on long-term contracts of 3 to 5 years, customer prepayment of security deposits, and closed lock up clauses (i.e. both parties settled at a fixed price agreed upon regardless of market fluctuations in computing power rent during the contract period) to lock future fluctuations in computing power rent into the contract in advance. Customers lock in costs in advance, and service providers also lock in a portion of sales gross profit in advance.

This method is effective for top enterprises, but not so easy for small and medium-sized players.

Zhou Bo said that medium-sized service providers have weak bargaining power, with upstream procurement costs rising as soon as possible, and downstream cloud service providers pressuring prices. He said, 'The impact on the cost side will be even more fatal.' He dare not transfer all the price increases, so the company's profits are often squeezed between the two ends.

The fluctuation of computing power prices is transmitted to the user end, becoming another type of account.

The head of computing infrastructure at a large modeling company told the Economic Observer that they are now using a hybrid approach of self purchased graphics cards and externally leased graphics cards: the core training cluster purchases a portion of the cards as long-term base disks; Model optimization, inference business, temporary experiments, and sudden expansion rely more on external leasing. The proportion of rental computing power fluctuates around 40%.

Over the past year, there has been a significant fluctuation in external computing power rent. The price of high-end training cards has fallen from its high in the past two years, but once there is a shortage of supply or a concentration of orders from major customers, the quotation will rebound. The supply of inference cards is more abundant, and the decline in rental prices is more pronounced.

The person in charge said that if there is a unilateral increase in rent, the cost change of leasing computing power may directly affect the project's gross profit by 3 to 5 percentage points. After the model is launched, it will continue to consume inference computing power. If graphics card rental suddenly increases and customer service prices cannot be raised synchronously, profits will soon be squeezed.

They are already using physical methods to avoid risks. The company will sign long-term agreements with top computing power operators for one to two years in advance, locking in the basic computing power scale and unit price. At the same time, set aside a portion of flexible quota to cope with temporary expansion. Financial derivatives have also been put on the internal conference table, but have not yet entered the execution stage. The company has had preliminary communication with futures institutions, and the attitude of finance and risk control is cautious and wait-and-see.

There is another layer of concern for the procurement side of central enterprises.

He Fang is the head of computing power procurement for a state-owned offshore oil and gas enterprise. The company rents computing power mainly for heavy load calculations such as seismic data processing and numerical simulation of oil and gas reservoirs. A task may run continuously for weeks or even months. Once the computing power is interrupted, it will not only affect the rent, but also delay the entire exploration project.

In the past year or two, he has seen significant fluctuations in the pricing of domestic intelligent computing parks. The rent of high-performance imported cards has fallen from a high level, but it will rebound during periods of tight supply. After the increase in supply of domestic computing power clusters, prices continue to decline. The price difference of computing power for the same specification between different parks can be as low as 20%.

But who is most afraid of not price increases. Some parks encounter power shortages, power restrictions, and temporary shutdowns after signing contracts. The project has a fixed schedule, and once the computing power is cut off, the waiting time of the on-site exploration team and the delay in data processing will all become chain costs. He believes more in physical long-term cooperation. The long-term agreement can clearly state the card type, scale, power supply guarantee, and downtime compensation. Financial contracts can lock in prices, but they cannot guarantee that computing power will be delivered on time.

Cloud vendors are also adjusting their product structure. A GPU cloud instance product manager from a cloud vendor told Economic Observer reporters that small teams prefer to use it flexibly by the hour. Customers are more willing to sign contracts for annual or multi-year reserved resources for large model training and long-term inference. Cloud manufacturers will provide fixed unit prices to key customers, but set resource quotas. Within the limit, the agreed price will apply, and any excess will be subject to market conditions.

From lessors to large model companies, from central enterprise purchasers to cloud vendors, everyone is using contracts, prepayments, reserved instances, and multi vendor procurement to provide a buffer against fluctuations in computing power prices. It is precisely because these methods are becoming more and more common that financial institutions have seen new entry points.

Supply Awakening

Nanhua Futures has taken a step forward.

In September, Henghua International, a subsidiary of Nanhua Futures, signed a strategic cooperation memorandum with Silicon Data. The cooperation revolves around GPU computing power pricing, forward curves, computing power swaps, and options. Silicon Data is the index partner of CME Group's two computing futures, providing H100 and B200 graphics card rental price indices.

This memo sends a signal that the price risk of computing power has begun to enter the product discussions of domestic futures institutions.

The industry is also waiting for such tools, but they are waiting very cautiously.

The business manager of AIDC, a smart computing center enterprise in Jiangsu Province, told Economic Observer that the company's risk structure has undergone significant changes after shifting from a light asset "second landlord" model to building its own 10000 card smart computing center. In the past, when subletting, the market downturn could reduce the number of cards and shrink the scale. Nowadays, servers, cabinets, and power support are all one-time large investments, and once hardware is put on the shelves, depreciation becomes rigid.

He is most concerned about two things. One is that the future computing power rent will decline, and the income of the already built clusters will not meet expectations. Secondly, after the implementation of the new generation of chips, the residual value of existing cards has rapidly declined. Long term and swap tools can hedge some of the downward risk of rent, but it is difficult to cover the depreciation of hardware itself.

The AIDC business manager mentioned above said that his company has both domestic and imported cards in the cluster, and the rental curves for different card types are not consistent. If derivatives are only anchored to a single model, it is difficult to match a mixed asset pool. Trading also requires the use of margin, which will squeeze the construction and operation funds of heavy asset projects.

But he does not exclude financial institutions from entering the market. He would prefer to see tools that can lock in future rental income first. The AIDC business leader mentioned above said, 'Once the Wanka cluster is built, fixed costs have already been invested.' If the market rent continues to decline in the future and the revenue after listing does not meet expectations, it will directly affect project returns. If financial institutions can design forward contracts that are in line with the real market conditions of domestic intelligent computing and match hybrid card pools, helping them lock in a portion of future income, their value will be higher.

The operator of the AI park, Li Zhe, has also received a price lock request from customers. Most of the large enterprises, central enterprises, and top AI company clients in his park sign long-term contracts ranging from 1 to 3 years, locking in computing power scale, basic electricity prices, and basic rent. Small and medium-sized AI teams tend to choose quarterly, semi annual, or monthly orders. In the past two years, more and more customers have proposed price reassessment clauses, hoping to re evaluate rent every six months based on market conditions.

If financial institutions launch forward contracts with computing power in the future, Li Zhe is willing to act as a bridge and introduce clients with lock up demand to professional institutions. But he won't make a deal on his own. He is concerned about the boundary of responsibility. Forward contracts are financial transactions, and if the index deviates from spot, liquidity is insufficient, or customers incur losses, the park side is easily involved in disputes.

Domestic small and medium-sized computing power owners also have expectations for local tools.

Ma Hongyuan, the owner of Zhongwei Small and Medium sized Computing Rental in Ningxia, mainly holds domestically produced AI computing chips such as Ascend 910 and Haiguang DCU. Unlike Nvidia GPUs, Ascend belongs to NPU and Haiguang DCU as deep computing units, not traditional graphics cards. He said that the overseas computing power futures linked to the Nvidia H100 and B200GPU US dollar leasing market are not the same market as them. The chip models, customer groups, pricing currencies, and transaction regions are all different. Even if one can participate, it has no practical hedging significance.

He hopes that there will be a public price benchmark in China that is linked to the domestic graphics card rental market. In the past two years, the domestic computing power market has shown significant differentiation. After the supply of domestically produced computing power for inference, the price of such graphics cards has been continuously decreasing. High performance domestic graphics cards that can train large models have relatively firm prices.

Ma Hongyuan calculated the accounts. If the utilization rate of computing power is below 70%, the computing power enterprise is basically losing money; Only when the profit margin of computing power reaches 80% or above, can enterprises have thin profits; Only when it reaches over 90% can the enterprise earn decent profits. For him, the price tool can only lock in the rental unit price and cannot guarantee that customers will definitely come to rent. Public price benchmarks have value, but cannot replace real orders.

Overseas exchanges are still awaiting regulatory review. Domestic enterprises have already made a round of hedging in contracts, products, and procurement systems. The cooperation between Nanhua Futures and Silicon Data has brought financial institutions to the doorstep. The demands provided by the industry are becoming increasingly specific. Top tier lessors want to manage their loose lease exposure, medium-sized service providers want to lock in procurement costs, heavy asset data centers want to smooth out future revenue, and large buyers want to stabilize their annual budgets.

But everyone is also reminding the same thing. The index should be close to real spot prices, contracts should match domestic card types and trading habits, liquidity should be sufficient, and compliance boundaries should be clear. Otherwise, if the computing power risk has not been hedged, new trading risks will be brought in first.

(At the request of the interviewee, He Fang, Zhang Lei, and Li Zhe are pseudonyms)