Rapid technological change almost always invites rapid financial innovation. As new forms of production emerge, markets rush to name them, package them, and – inevitably – securitize them. Familiar structures are repurposed, analogies are drawn to established asset classes, and future revenues are projected forward with confidence. Securitization has followed this path many times before, migrating from mortgages and auto loans to aircraft leases, shipping containers, and data centers. Today, attention has turned to artificial intelligence, where some argue that the continuous computation required to run AI models, particularly inference models, can be transformed into a stable, income-producing investment category.
The narrative is compelling. Early AI demand was dominated by model training: large, episodic, capital-intensive bursts of activity. The next phase, we are told, is defined by inference: models running continuously inside applications, services, and platforms. Because inference is always on, demand appears steadier, utilization higher, and revenues more predictable. From there, it is a short step to the claim that inference computation resembles infrastructure and therefore lends itself to securitization with utility-like characteristics.
There is truth in this story. But there are also a great number of assumptions layered atop it.
What Is Actually Being Securitized?
Before evaluating the investment claims, it is essential to define terms precisely. By computation, I mean the execution of machine instruction – measured in operations, throughput, latency, and memory access – required to transform inputs into outputs. Inference computation is the continuous application of these operations to trained machine-learning models in real time. Economically, computation is not a durable asset. It is a service flow, produced moment by moment using capital equipment (hardware), energy, software, and network infrastructure.
This distinction matters. Hardware can be owned. Data centers can be financed. But computation itself cannot be stored, warehoused, or held on a balance sheet. What can be securitized, at best, are claims on future revenues derived from providing computation services. Those claims inherit all the risks associated with pricing, utilization, technological change, and asset depreciation.
Framing inference computation as an “asset class” therefore risks conflating the means of production with the service being sold.
From Training to Inference: A Change in Usage, Not in Economics
It is broadly accurate that inference computation is growing rapidly. As AI systems move from research environments into commercial deployment, models must operate continuously rather than intermittently. This has altered how computing capacity is provisioned, favoring low latency, high availability, and constant throughput.
Inference is indeed more continuous than training, which tends to be project-based and irregular. This continuity improves utilization rates and makes revenues appear more stable. But continuity of activity does not alter the fundamental economics of computation. What is being sold remains a flow, not a stock, and that flow is produced in fiercely competitive markets.
Moreover, training has not disappeared. Fine-tuning, retraining, and adaptive learning blur the line between training and inference, reinforcing the point that AI workloads are dynamic rather than settled. The shift of activity underway does not, by itself, confer durability or pricing power.
What Securitization Demands – and Why This Is Difficult
Successful securitization rests on several conditions: predictable cash flows, long economic asset lives, estimable residual values, and enforceable legal claims. These conditions are difficult to satisfy when the underlying revenue stream comes from selling computation.
Inference demand may be steady, but pricing power is uncertain. Markets for computational services are highly competitive, dominated by large buyers, and subject to relentless efficiency gains. Advances in software optimization, model compression, batching, and specialized accelerators continually reduce the amount of computation required per unit of output. Those gains place persistent downward pressure on prices, even as volumes rise.
At the same time, the capital equipment used to produce computation depreciates technologically rather than physically. New generations of chips can sharply reduce the economic value of existing hardware well before it is worn out. Because revenue depends on selling computation, not owning silicon, declining costs of production can quickly undermine cash-flow assumptions embedded in long-dated securities.
Residual value – the backbone of asset-backed structures – is therefore fragile. Secondary markets for used AI hardware are thin and cyclical, making liquidation values highly sensitive to timing and technology cycles. Any securitization structure that assumes stable residuals is implicitly underwriting obsolescence risk.
The Utility Analogy and Its Limits
Much of the enthusiasm for securitizing inference computation rests on analogy. Because inference workloads are always on, they are likened to base-load utilities such as electricity or water. But this comparison breaks down under scrutiny.
Utilities (electricity, gas) derive stability from slow technological change, regulated or quasi-regulated pricing, geographic constraints, and long-lived assets. Inference computation exhibits none of these features. Supply can expand rapidly. Efficiency improves continuously. Prices are set in competitive markets rather than by regulators. Margins compress rather than persist.
The history of computing itself makes this clear. Demand for computation has exploded for decades, yet the price per unit of computation has consistently fallen. Always-on usage has coexisted with aggressive deflation. There is little reason to believe inference computation will behave differently simply because it is embedded in AI applications.
Securitization cannot change this underlying reality. It can only repackage exposure to it.
Contracts Carry the Entire Burden
For revenues derived from inference computation to support securitization responsibly, contracts must do most of the work. Long-term agreements with creditworthy counterparties, minimum utilization guarantees, pricing protections, refresh provisions, and clear claims on specific hardware are essential.
Even then, risks remain. Counterparties may fail or renegotiate. Customers may internalize inference workloads as costs fall. Large buyers, particularly ‘hyperscalers,’ can exert sustained downward pressure on prices. Technological progress may outpace the assumptions embedded in contracts.
In practice, securitized inference computation looks far more like leased capital equipment than infrastructure. It may generate income under favorable conditions, but those conditions are narrow and sensitive to change.
Revenue Can Be Securitized…Computation Cannot
Inference computation has become more monetizable as AI deployment expands. That is an important development. But monetization is not durability, and continuity of demand does not guarantee stable returns.
Computation itself cannot be securitized. Only claims on future revenues derived from computation can be. Those claims inherit all the uncertainties of pricing power, technological deflation, and asset obsolescence that characterize modern computing markets.
In narrow, carefully structured cases, securitization may prove viable. Treating inference computation as a broadly stable, utility-like investment category, however, risks repeating a familiar mistake: mistaking financial packaging for economic certainty.
Peter C. Earle, Ph.D. is Senior Research Fellow at the American Institute for Economic Research.
