The conversation around building financial infrastructure for AI compute is often clouded by a few persistent myths. By looking at how established markets operate, we can see that the challenges facing compute are far from unique.

Misconception 1: Compute cannot be financialized because it is not fungible. 

Very few commodities are truly fungible. Crude oil is not. The United States is a major exporter of oil, yet it heavily relies on oil imports from the Middle East. This happens because the “light sweet” crude it produces domestically is not fungible with the “heavy sour” crude that major refineries are built to process. Furthermore, oil delivered to one pipeline is of no use to a refinery located at a different, unconnected hub. 

The situation with compute is nearly identical. Differences in data center locations and hardware specifications create natural barriers to interchangeability. However, fungibility is a gradient rather than a binary state. Markets have long since learned to trade non-fungible assets by creating standards and price differentials.

In addition, let’s also not confuse physical fungibility with economic similarity. Bitcoin and Ethereum are not fungible, yet they are highly correlated. Risk management and investment rely on economic values rather than physical substitution.

Misconception 2: Compute cannot be financialized because there is no standard unit of measurement. 

There is no shortage of units in this market, and each one is a standard in its context. We use TFLOPS (floating-point operations per second) to measure compute performance, hours to conduct rental transactions, and tokens for end-user applications. Any one of these could potentially become the base unit for financial contracts, or they might all coexist within different derivative structures.

The true concern for financialization is not a lack of units. The issue is the variability of the utility delivered by a single unit, which creates a challenge for transparent pricing. But this is an engineering challenge, not a structural show-stopper. There are already several ways to mitigate variability:

• Market segmentation: We can divide the underlying market to increase homogeneity in each subsection. For example, the current Silicon Data rental indices divide the GPU rental market into major categories: H100, B200 etc..

• Performance benchmarking: We can adopt standard performance benchmarking to avoid the apple-to-orange comparison. Performance-based pricing will account for hardware heterogeneity. 

These solutions are just the beginning. As the market matures, new methods for reducing unit variability will inevitably emerge.

Misconception 3: Compute cannot be financialized because of its short technological life cycle.

Rapid depreciation is often cited as a barrier, yet the financial world handles short-lived assets daily. For example, the underlying asset for a 2-year T-note Futures has a maximum remaining life of only two years, yet these contracts have traded actively since the mid-1990s.

In the compute technology, the life cycle is much longer than two years. The A100 GPU was introduced in May 2020. Even in 2026, it remains widely used with a stable price range. 

Even if product iterations accelerate, we already possess a mature and efficient financial infrastructure—from exchanges to clearing facilities—capable of supporting fast-moving assets and product iteration. 

A new asset class

To be clear, financialization of compute is not a solved problem. We are not suggesting that compute will trade like oil, treasury bond, or even electricity. No two commodities trade exactly alike. Compute will emerge as its own asset class with its own unique idiosyncrasies. 

It’s important to remember that many challenges we face today have been solved throughout 400 years of financial market history. We have the basic recipes. As the demand for compute liquidity grows, the market will do what it does best: adapt existing tools and invent new ones to meet the moment.