Computing Power Capital Market

Author: Vaidik Mandloi, Compiled by: Block unicorn

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

This is how chaotic the GPU capacity market is. There is no pricing benchmark, and lenders cannot hedge the risk of their financed hardware; everything is based on blind capital allocation. But this is about to change, as the Chicago Mercantile Exchange (CME) and Intercontinental Exchange (ICE) have announced plans to launch futures contracts for GPU computing time.

Computing power is being integrated into a comprehensive capital market, much like electricity in the 1990s. Today, I will delve into how this new liquidity forward curve, driven by stablecoin settlement, can transform the largest infrastructure buildout since the railroads.

Making Computing Power Tradable

When I say computing power is following electricity's path into capital markets, I mean it very specifically, and understanding this reveals how the market is actually structured.

In commodity markets, traders draw a sharp distinction between storable and flow commodities. For example, oil is a storable commodity because you can store it in tankers until you find a buyer. You can stockpile crude when prices are low and sell it when prices surge. Computing power, however, is a flow commodity because you rent GPU time for a period and pay for it. Any computing capacity unused during the rental period is lost forever.

Idle GPUs sitting in racks are not "stored computing power," just as a disconnected power plant is not "stored electricity," because in both cases, the valuable product is the flow—GPU hours or kilowatt-hours—not the physical machines that generate them.

This is crucial for pricing. Storable commodities have the built-in stabilizer of inventory, which can be released during periods of sharp price volatility to offset price increases. Flow commodities lack this buffer; this is why spot prices for computing power often fluctuate wildly.

In mid-2025, due to the launch of NVIDIA's next-generation Blackwell chips, a large influx of new supply hit the market, causing demand for H100 GPUs to decline and spot computing prices to plummet 70% in 18 months. But this year, due to mass production of HBM chips, demand surged, and with no inventory to absorb it, H100 prices skyrocketed 48% in just four days. For AI companies (whose training runs cost tens of millions of dollars) and lenders who have provided over $120 billion in datacenter credit for this hardware, such unhedgeable volatility is a life-or-death issue.

Furthermore, there is a second problem. A barrel of crude oil is identical to any other barrel anywhere in the world, which is why it can be traded on exchanges without physical inspection. But an H100 in Virginia and an H100 in Iceland are vastly different products because the chip, cluster configuration, and adjacent workloads all affect actual performance.

Benchmark data from global GPU providers shows performance differences of up to 38% even for nominally identical hardware. The electricity industry faced the same problem in the 1990s: electricity from the Texas grid was very different from that in the Mid-Atlantic region because transmission and local demand created different conditions at every node of the grid. The only solution at the time was to assign different prices to each node and quote based on the price differential to a reference benchmark. That reference benchmark is exactly what the computing market lacks today.

SF Compute has built a real-time order book for GPU time, where buyers and sellers can trade time just like any commodity on a spot market. The logic is that once a liquid spot market exists, trading activity can be used to derive an index price. That index price can then be used to construct cash-settled futures contracts.

Once a datacenter can sell futures and lock in revenue months ahead, it can approach lenders, demonstrating that its revenue is hedged, thereby obtaining lower interest rates and scaling up. This, in turn, lowers the overall cost of computing for everyone.

Another company, Silicon Data, has built a daily index called SDH100RT, which has been live on the Bloomberg terminal since May last year, aggregating 3.5 million data points from global suppliers into a single benchmark, costing the equivalent of one hour of H100 GPU time. The newly announced CME futures contracts will settle against this index. Several other firms are racing to build similar indices, because becoming the reference price means capturing a fraction of every trade in the market as long as the market exists.

The electricity market went through a similar phase: in 1993, Nord Pool opened the first electricity futures exchange, followed by over 200 new electricity marketing companies. It took a decade for industry insiders to argue whether electricity was legally a commodity, but today it has become a $6 trillion market. The computing power market is currently undergoing a similar journey.

Intermediaries

So now we have what can be called the first spot computing markets to adopt some form of price index, and exchanges have announced intentions. However, between an index price on the Bloomberg terminal and a well-functioning capital market lies a critical link that makes everything work—and it differs significantly from traditional trading methods.

The computing futures market will not operate like a stock exchange, where standardized shares are traded between anonymous buyers. It will be dealer-driven, with traders acting as bridges between GPU owners (who want to lock in revenue) and AI companies (who want to lock in costs).

For example, suppose a US datacenter has a large batch of H200 servers available starting in October. A startup needs 500 GPUs but only cares about the interconnect being InfiniBand (a GPU communication medium) and doesn't care about the specific location. This is a very specific demand that requires someone to handle this custom order while hedging the risk against the standardized index.

This is nothing new; every commodity has historically needed such a link for parties to disentangle the complexities of the physical product and transform it into interchangeable units tradeable on exchanges. An H100 contract sitting on a shelf is just a custom contract; no one else can price it. It only generates returns for one party based on a private agreement, and the rest of the financial system can't even touch it. But if it can be combined with an index price and a public settlement layer, it becomes a liquid commodity that lenders can hedge.

In 2023, CoreWeave borrowed $2.3 billion solely backed by NVIDIA GPUs, marking the first time H100 hardware secured a loan. Its most recent financing received an investment-grade rating from Moody's, based on Meta's credit profile rather than CoreWeave's, because Meta signed a "take-or-pay" contract requiring payment regardless of actual usage.

This is precisely where the crypto rails play a load-bearing role. Buyers and sellers of computing power are spread across the globe, but none of them have CFTC approval to open US commodity exchange accounts. However, crypto wallets can settle stablecoin payments, and any wallet can hold tokenized computing power.

GPU export controls have already revealed the geopolitical stratification of computing access; for example, NVIDIA cannot export cutting-edge chips to China and dozens of other countries. A stablecoin-settled computing futures market could allow researchers and startups outside export-controlled zones to access computing pricing and hedge costs through infrastructure that bypasses restrictions, much like stablecoins already do in Argentina and Nigeria.

The Liquid Forward Curve

Currently, building a GPU cluster means borrowing millions of dollars against unhedgeable revenue because the global financial market lacks the necessary instruments. But a liquid forward curve allows companies to borrow against hedged revenue at lower interest rates than unhedged positions. This means lower cost per computing hour. So who will build the settlement layer for the forward curve?

The only solution needed now is to establish a settlement layer where anyone can verify collateral and make the forward curve a public good. Currently, we cannot verify the condition of the collateral hardware, whether it has been double-pledged, or its actual utilization rate. But if GPUs and their revenue streams are tokenized as on-chain assets, every lender can verify the collateral in real time, making the forward curve publicly visible rather than mired in bilateral negotiations.

Furthermore, the next generation of AI agents will purchase computing resources per inference call, and they cannot open bank accounts. Cryptocurrency is the only payment gateway capable of executing micro-transactions between an agent in Tokyo and a GPU rack in Virginia in under a second.

There are powerful countervailing forces now, because GPU supply is highly concentrated. The top hyperscaler operators control 78% of global IT computing power. NVIDIA commands over 80% of the high-end AI chip market, and its product launch schedule alone 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 moved from balance sheets to Wall Street-funded special purpose vehicles (SPVs), much of it ending up in corporate bond funds within target-date retirement products, held by individuals unaware of it. I believe the financing models used to build this infrastructure likely contain assumptions about hardware residual value that existing data cannot yet support.

The electricity market did not stop at the generator; it ran through the entire system to the wall socket and affected the price of all power-consuming devices. The computing market still has many lines to lay.