AI Compute as Nascent Real Bills: A Clearing Instrument for the Machine Economy

Antal Fekete spent the last two decades of his life arguing that Adam Smith's Real Bills Doctrine — dismissed by Mises, ignored by most post-Mises Austrians, and largely absent from contemporary monetary economics — was the single most important organizing principle for understanding how commercial economies actually clear without recourse to central-bank credit. Shortly before his death, at the suggestion of his student Paul Bouvet, he renamed the concept from "real bills" to "gold bills," partly to emphasize the critical requirement that such bills must mature into gold coins, and partly because the adjective "real" had been used pejoratively by the Chicago school to dismiss the idea.

Under either name, the concept has a specific structure that makes it very easy to recognize wherever it appears. The goal of this essay is to argue that the concept is appearing right now, in a form Fekete could not have predicted, inside the commercial machinery of the AI industry. The AI economy is spontaneously developing a clearing instrument that has the exact structural features of a Fekete gold bill, minus the gold backing.

This is not a small observation. If it is correct, it means the first genuine monetary-adjacent clearing instrument to emerge in the commercial world since the industrial revolution is forming inside the operational plumbing of the model providers, and almost no one — including the companies issuing the instruments — is aware of what they are.

What a real bill actually is

The distinction Fekete insisted on, and that most of his Austrian contemporaries refused to accept, is this: a real bill is not a loan. It is a clearing instrument.

When a producer of semi-finished goods — Fekete's standard example was a Lancashire cloth merchant in the late 18th century — delivered those goods to the next stage in the supply chain, payment in gold coin at the moment of delivery was impractical. The receiving party did not yet have the gold. The gold would come, with high probability, when the finished product reached the ultimate consumer three to six weeks later. Rather than demand immediate settlement in coin, the receiving party would accept a bill — a short-duration claim on the future gold receipts, typically maturing in 91 days, the approximate time required for the good to move through the remaining stages and reach the final consumer.

The bill could be endorsed and passed along. A merchant receiving a bill could use it to pay his own supplier. The supplier could discount it at a bank, or hold it to maturity. When the final consumer paid gold for the finished good, the gold flowed back up the chain and settled all the outstanding bills along the way.

Critically — and this is the point that Mises missed and that Fekete died trying to restore — no party in this chain was lending anything. Each party was simply accepting a short-duration claim on gold that was known, with very high probability, to be arriving shortly. The bill did not finance consumption. It did not create new demand. It did not expand the money supply. It cleared a sequence of transactions that had already occurred in the real economy in a way that permitted trade to proceed without requiring each party to hold enough gold on hand to settle every stage in cash.

Fekete's summary: "Gold bills are not evidence of lending and indebtedness. They are evidence of clearing and cooperation."

The characteristics that made a bill "real" in Adam Smith's sense — and "gold" in Fekete's — were five.

Self-liquidating: the bill matured when the underlying good reached the consumer and was paid for in gold. No rollover was required. No discretionary decision by any party was needed to retire the bill.
Short duration: 91 days was canonical; longer maturities were treated as loans, not bills, and were handled differently.
Claim on a specific productive good in transit: the bill was never an abstract claim. It was tied to an identifiable good that was verifiably moving through the supply chain toward a known final consumer.
Arising spontaneously from commerce: bills were drawn by the parties to a commercial transaction, not issued by a bank or government. The bill market grew bottom-up.
Discounted at a rate determined by the market: the discount rate on a bill reflected the market's assessment of the probability that the underlying good would reach the consumer and be paid for in time. It was a market price, not a policy rate.

A real bill with these five features, Fekete argued, was the highest-quality earning asset that could exist in a commercial bank's portfolio. It was more saleable than gold bonds. It financed no new credit. It was simply the technical mechanism by which a commercial economy cleared production runs without tying up enormous quantities of gold at every intermediate stage.

What AI compute is, and why it matches

Consider the modern AI economy's operational layer. A developer needs compute — GPU hours, API credits, inference tokens — to serve end customers. The developer's end customers will pay for the service in cash, roughly continuously, as the service is consumed. The compute consumed between the developer and the end customer is a short-duration input, tied to a specific productive output (the service delivered), flowing through a value chain from hyperscaler to developer to end user.

The contracts that govern this flow have a very specific structure.

AWS Savings Plans, Azure Reserved Capacity, and Google Cloud Commitments are annual or multi-annual commitments to a certain volume of compute consumption, priced at a discount to on-demand. These are too long to fit the gold-bill structural definition, and are closer to forward contracts than to bills.

But the shorter-duration instruments matching the structure are proliferating rapidly. API credit packages — purchased in advance, consumed on use, expiring in weeks to months — behave as self-liquidating claims on specific productive output. GPU marketplaces like CoreWeave's spot market, Runpod, Lambda Labs, and a growing set of brokers like Together.ai, Parasail, and Nebius offer variable-duration compute contracts that are often in the 30-to-90-day range. Token bundles at the major model providers — most visibly Anthropic, OpenAI, and Google — are issued in specific denominations, consumed on use, and track a specific productive throughput of generation tokens.

Every one of these instruments exhibits the five features of a gold bill, with a single substitution: the underlying commodity is not gold but compute, and the denomination is not gold coin but U.S. dollars.

Self-liquidating: an API credit extinguishes when the inference is performed. A reserved GPU hour extinguishes when the hour passes. There is no rollover mechanism. The instruments retire themselves.

Short duration: the typical lifecycle from issuance to consumption is weeks. The longer-dated commitments are outliers; the working population of these instruments turns over on a 30-to-90-day cycle.

Claim on a specific productive good: the underlying is compute delivered by a specific infrastructure provider, measurable in well-defined units (FLOPs, tokens, GPU-hours, query volume).

Arising spontaneously from commerce: the instruments were not created by regulatory fiat. They emerged because the commercial reality of AI service provision required a clearing mechanism between the providers' capacity-planning horizons and the consumers' demand cycles. The providers did not consult monetary economists before issuing them. They simply began issuing them because it was the obvious way to structure the commercial relationship.

Discounted at a market-determined rate: the discount between spot and forward compute pricing is substantial (often 40 to 70% for multi-year commitments), is continuously renegotiated, and is an active commercial variable. The discount is a market price, not a policy rate.

The analogy is not approximate. It is exact, on every structural dimension that Fekete cared about.

What is missing

The gap between a compute bill and a gold bill is that the gold bill matured into gold, and the compute bill matures into dollars. This is not a small difference. The gold bill's ultimate integrity rested on the gold coin at the end of the chain; the compute bill's ultimate integrity rests on the stability of the dollar in which the underlying contracts are denominated.

If the dollar's saleability continues to erode along the trajectory described elsewhere in this series, the compute bill inherits that erosion. A short-duration claim on dollar-denominated compute is not structurally better than a short-duration claim on any other dollar-denominated output, at the level of monetary resilience.

But there is a potential bridge here, and it is the part of the analysis that most rewards further thought. The compute good underlying the bill — GPU cycles, inference tokens — is itself a physical, productively useful commodity in a way that many other service goods are not. It can in principle be priced and settled in any medium of exchange that both parties to the contract agree on. There is no inherent reason the compute bill must remain dollar-denominated.

If, for instance, a meaningful volume of compute contracts began to be written and settled in a cryptographically-settled, non-fiat medium — or in a hybrid arrangement where the dollar denomination was a pricing reference but the actual settlement was in some other instrument — the compute bill would become a monetary-adjacent clearing instrument that was independent of the fiat substrate. This is not a science-fiction scenario. The underlying primitives for it already exist. What is missing is a standard form and a critical mass of issuance.

The Menger angle

In Menger's marketability framework, the characteristics of a good that confer high saleability include homogeneity, divisibility, durability, and widespread demand. Compute, as a commodity, scores remarkably high on each.

Homogeneity is achieved through hardware standardization (an H100 GPU-hour is substantially fungible across providers) and benchmark equivalence (FLOPs or tokens or inference-queries-per-second are directly comparable units).

Divisibility is nearly perfect. A single inference can be billed in fractions of a cent. The smallest unit of compute is orders of magnitude smaller than the smallest unit of currency.

Durability is problematic in a specific way: compute itself is not durable — it is consumed on use — but claims on future compute are as durable as the issuer's solvency and the infrastructure's continued operation. This is a distinction from gold, which is physically durable independent of any issuer. But it is a similar structural position to that of a commercial bill, which is durable only through the maturity of the underlying good.

Widespread demand is the remarkable feature. In 2026, essentially every technology company, every financial institution, every research organization, and a growing share of consumer activity generates demand for inference compute. The universe of parties that will accept a well-structured compute credit in exchange for a good or service they have to offer is large and growing. This is the characteristic that Menger identified as most fundamental to saleability.

Compute is not money, in Menger's strict sense, and I am not arguing that it should or could be. But it is plausibly climbing the saleability spectrum faster than any new commodity has climbed it in modern history, and it is doing so in an instrument form — the short-duration, self-liquidating claim on productive output — that matches Fekete's preferred clearing structure precisely.

Why the Austrians missed it

The Mengerian and Fekete-ian traditions should have seen this coming first. They were handed the framework that fits the observation most exactly. Two factors have prevented the identification.

First, most contemporary Austrians inherited the Misesian dismissal of the Real Bills Doctrine. Mises held that real bills were inflationary because he treated them as a form of credit creation. Fekete's career-long project was to demonstrate that this was a misreading — that bills were clearing, not credit. The Austrian tradition as a whole did not accept Fekete's correction. As a result, most contemporary Austrians are not looking for real-bill-type structures in the modern economy, because they have been trained to regard any such structure as inflationary and therefore suspect.

Second, the schools that would be receptive to the structural observation — heterodox monetary theorists, MMT-adjacent thinkers, the Bitcoin and Ethereum developer communities — are not particularly literate in Fekete and Menger. The Bitcoin community, in particular, has developed a sophisticated understanding of the gold-bill-like properties of certain Lightning Network and Layer-2 instruments, but does so in a vocabulary that makes no contact with the classical monetary tradition. The two literatures are not in dialogue.

The New Austrian Economics, as distinct from the Misesian or Rothbardian lines, should be able to bridge this gap. Fekete's framework is exactly the one that makes the compute-bill observation legible. Applying it requires only that we take his corrections to Mises seriously, and that we look at the actual commercial structures that have emerged in the compute economy without assuming in advance that they must be some form of credit expansion.

What to watch for

If the compute-bill thesis is correct, three developments would confirm it.

Standardization of terms: the current compute-contract market is fragmented and bespoke. Each hyperscaler and broker issues instruments in their own format. Real bills in the 18th century went through a similar early phase before market practice converged on standardized maturities, standardized denominations, and standardized endorsement conventions. A similar convergence in the compute market would be a strong signal that the structural analogy is developing.

Secondary market depth: real bills became a major clearing instrument when a deep, continuous secondary market in them developed at the major commercial banks, through the discount window. A secondary market in compute bills — where a party holding future compute credits could sell them to a third party without recourse to the issuer — would be the single most important structural development for the monetary analog. Some early versions of this exist (compute broker resale, GPU spot market instruments), but the volume and standardization are not yet at the level of a mature bill market.

Emergence of specialized discount houses: in the Smith-Menger-Fekete framework, the commercial bank's core business was discounting real bills. A modern equivalent would be specialized financial firms whose primary business is the wholesale discounting of compute instruments — pricing, packaging, and distributing them across the economy. Several startups in the adjacent space are visible (compute-futures brokers, reserved-capacity aggregators), but none has yet assumed the full "discount house" role.

The first of these three milestones will probably emerge in the next 24 months. The second and third are further out but plausible within the decade.

The larger significance

The compute economy is quietly constructing, without intent or theoretical self-awareness, a clearing layer that has the structural features Fekete spent his career defending. This matters for two reasons.

First, it is evidence that the Smith-Menger-Fekete framework is descriptively correct: commercial economies spontaneously develop gold-bill-type structures when the demand for a productive input requires clearing faster than currency-based settlement can provide. Fekete insisted this would be true under any monetary regime in which genuine productive activity took place. The compute economy is confirming the claim.

Second, it means the New Austrian Economics has an empirical opening that the older Austrian traditions closed off for themselves. If the decline of dollar saleability continues along the trajectory of the last several years, and if the compute economy continues developing its own clearing mechanism, there is a plausible future in which the commercial world develops its operational plumbing around the fiat system rather than through it. Not by monetary revolution. By the same quiet, incremental, self-interested process that produced gold bills in Lancashire cloth mills 250 years ago.

The first step in making that future visible is identifying what is already happening. Fekete gave us the vocabulary. The compute economy is writing the next chapter, in real time, while the policy establishment and most of the economics profession continue to insist nothing of monetary significance is occurring.

Next in this series: the capstone essay on the emergence of frontier AI labs as de facto issuers of the cryptographic trust layer that secures every digital financial claim — a development that may render the Federal Reserve Act's architecture structurally obsolete in a way the Fed itself cannot address.