Computing power capital market

Author:Translator:: Block unicorn

Google is one of the world's three largest cloud service providers, and it currently purchases $920 million worth of computing resources monthly from SpaceX (a rocket company).

This is the chaotic state of the GPU capacity market. There is no benchmark for pricing, and lenders cannot hedge the risks of their financed hardware; everything is based on blind capital allocation.

But this is about to change, as the Chicago Mercantile Exchange (CME) and the Intercontinental Exchange (ICE) have announced the launch of GPU computing power time futures contracts. Computing power is being integrated into a comprehensive capital market, much like electricity was in the 1990s. Today, I will delve into the transformative potential of this new liquidity forward curve, driven by stablecoin settlements, for the largest infrastructure construction since the railroads.

Making computing power tradable

I said that computing power is entering the capital market along the development path of electricity. I mean this very specifically, and understanding this will tell us exactly how this market is constructed.

I said that computing power is entering the capital market along the development path of electricity. I mean this very specifically, and understanding this will tell us exactly how this market is constructed.

I said that computing power is entering the capital market along the development path of electricity. I mean this very specifically, and understanding this will tell us exactly how this market is constructed.

In commodity markets, traders make a huge distinction between inventory goods and liquid goods. For example, oil is an inventory good because you can store it in tankers until you find a buyer. You can stockpile crude oil when prices are low and then sell it when prices soar. Computing power, on the other hand, is a liquid good because you rent GPUs for a period of time and pay for it. Any computing power not used during the rental period will be permanently lost. A GPU idle in a rack is not the same as "stored computing power," just as a disconnected power plant is not the same as "stored electricity," because in both cases, the valuable product is the flow—GPU hours or kilowatt-hours—not the physical machine that generates that flow. This is crucial for pricing because inventory goods have a built-in stabilizer, while liquid goods lack this stabilizer. Inventory can be released during periods of high price volatility to offset price increases. Liquid goods do not have this buffer; this is why the spot price of computing power often fluctuates wildly. In mid-2025, due to the release of Nvidia's next-generation Blackwell chip, a large influx of new supply into the market led to a decline in demand for H100 graphics cards, causing the spot price of computing power to plummet by 70% in 18 months. However, this year, due to the mass production of HBM chips, demand surged, and with no inventory to absorb, the price of H100 graphics cards soared again by 48% in just four days. For AI companies (whose training and operation costs can reach tens of millions of dollars) and lenders providing over $120 billion in data center credit for this hardware, this volatility without hedging tools is a matter of life and death. Furthermore, there is a second problem. A barrel of crude oil, no matter where it is in the world, is exactly the same as any other barrel of crude oil, which is why it can be traded on exchanges without physical verification. But the H100 in Virginia and the H100 in Iceland are completely different products because the chips, cluster configuration, and adjacent workloads all affect their actual performance. Benchmark data from global GPU vendors shows that even nominally identical hardware can exhibit performance differences as high as 38%. The power industry faced a similar problem in the 1990s: the Texas power grid differed drastically from that of the Mid-Atlantic region due to varying transmission and local demand at each node. The only solution at the time was to set different prices for each node and quote prices based on a benchmark price difference. This benchmark is precisely what the current computing power market lacks. SF Compute has built a real-time order book for GPU time, where buyers and sellers can trade time as easily as any commodity in the spot market. The logic is that once a highly liquid spot market is established, index prices can be derived from trading activity. These index prices can then be used to construct cash-settled futures contracts. Once a data center can sell futures contracts and lock in revenue months in advance, it can approach lenders, showing them that its revenue is hedged, thus securing lower interest rates and scaling up. This, in turn, reduces the overall computing cost for everyone. Another company, Silicon Data, has built a daily index called SDH100RT, which has been online on the Bloomberg terminal since last May. It currently aggregates 3.5 million data points from global vendors, forming a single benchmark, at the cost of just one hour of H100 GPU runtime. Newly announced futures contracts on the Chicago Mercantile Exchange (CME) will be settled using this index. Several other companies are currently racing to build similar indices because becoming a reference price means they can capture a small fraction of every trade in the market, as long as the market exists. The electricity market has gone through a similar phase: in 1993, Nord Pool launched the first electricity futures exchange, followed by the emergence of over 200 new electricity marketing companies. Industry insiders spent a decade debating whether electricity was a commodity in a legal sense, but today it has become a market worth $6 trillion annually. The computing power market is currently undergoing a similar process. Therefore, we now have what could be called the first spot market for computing power to adopt some form of price index, and exchanges have announced their intentions. However, between the index price on the Bloomberg terminal and the well-functioning capital market, there is a crucial link supporting all of this, and this link is quite different from traditional trading methods. The computing power futures market will operate very differently from stock exchanges, where standardized stock trading takes place between anonymous buyers. The computing power futures market will be dominated by dealers who act as a bridge between GPU owners (who want to lock in revenue) and AI companies (who want to lock in costs). For example, suppose a data center in the United States has a large number of H200 servers available starting in October. A startup needs 500 GPUs, but only cares about whether the interconnect is InfiniBand (a GPU communication medium), not the specific location of the servers. This is a very specific requirement, and someone needs to handle this custom order while hedging against the risks associated with standardization. This is nothing new; in the past, every commodity required such a step, allowing stakeholders to unravel the complex relationships within the physical product and transform it into a tradable, interchangeable unit on an exchange. The H100 contract on the shelf is merely a custom contract, whose price cannot be determined by others. It can only generate revenue for one party according to a private agreement, inaccessible to the rest of the financial system. But if it can be combined with index prices and a public clearing layer, it can become a live commodity that lending institutions can hedge. In 2023, CoreWeave borrowed $2.3 billion using Nvidia GPUs as collateral, marking the first time its H100 hardware had received a loan. Its most recent funding round received an investment-grade rating from Moody's, based on Meta's creditworthiness rather than CoreWeave's, because Meta signed a "take-or-pay" contract, requiring payment regardless of whether the computing resources were actually used. This is precisely where the cryptocurrency track plays a crucial role. Buyers and sellers of computing resources are located globally, but none can obtain approval from the U.S. Commodity Futures Trading Commission (CFTC) to open accounts on U.S. commodity exchanges. However, crypto wallets can settle stablecoin payments, and any wallet can hold tokenized computing resources. GPU export controls have revealed geopolitical stratification in access to computing power resources; for example, Nvidia is unable to export cutting-edge chips to China and dozens of other countries. A computing power futures market settled in stablecoins allows researchers and startups outside export-controlled areas to access computing power resource pricing and hedge costs through infrastructure that bypasses restrictions, much like stablecoins have already done in Argentina and Nigeria. Currently, building GPU clusters means borrowing millions of dollars against non-lockable revenue, as there are no suitable instruments in global financial markets. However, a highly liquid forward curve allows companies to borrow against hedged revenue at rates lower than those for unhedged positions. This translates to lower costs per computing hour. So, who will build the settlement layer for the forward curve? The only solution currently needed is to establish a settlement layer that allows anyone to verify collateral and treat the forward curve as a public good. Currently, we cannot verify the condition of the collateralized hardware, whether it has been double-collateralized, or its actual utilization rate. However, if GPUs and their yield streams are tokenized as on-chain assets, each lender can verify the collateral in real time, making the forward curve publicly visible instead of getting bogged down in bilateral negotiations. Furthermore, future generations of AI agents will purchase computing resources based on the number of inference calls, and they cannot open bank accounts. Cryptocurrency is the only payment gateway capable of completing microtransactions between the Tokyo agent and the Virginia GPU rack in less than a second. There are currently some strong checks and balances because GPU supply is highly concentrated. The top hyperscale data center operators control 78% of global IT computing power. Nvidia holds over 80% of the high-end AI chip market share, and its product release schedule can sway the entire market. Standardization is a bottleneck, but financializing an asset class during a construction boom could make it more contagious. Over $120 billion in AI infrastructure debt has been moved from balance sheets to Wall Street-funded special purpose vehicles (SPVs), much of it into corporate bond funds in target-date retirement products, unbeknownst to the individuals holding these bonds. I believe the financing models used to build this infrastructure likely contain assumptions about hardware residual value that are not adequately supported by available data. The electricity market doesn't stop at generators; it permeates the entire system, down to wall sockets, and affects the price of all electrical appliances. The computing power market also has much work to do.