Before the exchange: who certifies that compute is substitutable?

A buyer comparing two GPU rentals encounters a deceptively simple question: what would make either supplier an acceptable substitute for the other? A shared chip name helps, but it does not answer whether the job will finish on time, whether interruptions are permitted or what happens when measured performance disappoints. A market becomes easier to use when buyers can rely on a common description without renegotiating every detail.

Grain trading offers a useful historical comparison. Its contribution is not a prediction that computing will follow the same path. It is a way to distinguish three activities often compressed into “standardization”: defining a grade, measuring compliance and deciding whose judgment counts when parties disagree. A specification can describe an exchangeable product. Institutions make that description credible.

The historical sequence

The Commodity Futures Trading Commission’s retrospective chronology records the Chicago Board of Trade’s founding as a grain cash market in 1848 and standardized forward-contract terms in 1858. Its account of the Illinois charter of February 18, 1859 is particularly instructive: the charter standardized grades and provided for Board-appointed grain inspectors whose decisions bound members. Formal rules concerning margin and delivery followed on October 13, 1865. S1

These dates resist the tidy story in which someone invents a contract and a liquid market appears. The documented sequence includes commercial organization, common language, inspection authority and delivery rules. It does not identify the contribution of each reform to trading volume, nor establish that inspection alone caused the market to develop. What it does show is that the meaning of a grade and the authority to apply it developed alongside the contracts.

That distinction survives in modern grain institutions. USDA describes an official inspection system comprising federal offices and authorized state and private agencies, operating with common standards, procedures and equipment. Internal quality-control programs and independent laboratories can analyze grain, but official certificates come from authorized participants in that system. S2 Measurement capacity and recognized certification authority are separate things.

An appeal is part of the product

The current United States Code makes the dispute mechanism unusually explicit. Section 79 requires regulations covering reinspections, appeal inspections and cancellation of superseded certificates. It also gives uncanceled official certificates a specified evidentiary status in federal proceedings. S3 This is not an assertion that every disagreement disappears. It is evidence that the system anticipates disagreement and provides a way to replace an earlier judgment.

For compute trading, that institutional function matters as much as the benchmark score. Suppose a hypothetical seller passes an agreed test before offering a cluster. During the purchased window, the buyer experiences much slower processing. The initial result cannot by itself establish whether the cause was the seller’s service, a changed workload, the buyer’s configuration or a failure in the measurement process.

A useful service agreement would identify the test version, relevant configuration, observation window, retained evidence and remedy for a defined failure. It would also say who can contest the result and how an amended determination affects the bill. Those are proposed design requirements, not features this article has verified across today’s compute marketplaces. Their purpose is to avoid turning every failed job into an argument about an undefined promise.

Grades preserve differences

Grain standardization also cautions against treating every attribute as interchangeable. USDA’s May 2014 wheat standards distinguish eight classes and specify grade requirements rather than assigning all wheat one universal quality number. S4 Standardization can organize differences; it need not erase them.

The equivalent compute question is therefore narrower than “What is an H100 worth?” It might be: what service can be bought, during a specified window, from any qualifying seller without violating this workload’s requirements? Hardware configuration, usable memory, interconnection, interruption rights and software conditions may belong in that definition. Which attributes are essential will depend on the job.

Consider a hypothetical offline inference batch that can tolerate restarting, compared with an interactive service facing strict response-time requirements. A service acceptable to the first buyer might fail the second buyer’s requirements even when the hardware is identical. A market that retains this distinction could be more informative than one that produces a single normalized accelerator-hour price.

Existing benchmarks illustrate one component of the task. MLCommons defines inference benchmarks using datasets and quality targets; its Closed division requires the reference model to support controlled comparisons. S5 That is evidence that comparability can be deliberately specified. Extending a laboratory result into a commercial promise about a particular rented service is an additional step. The benchmark result alone does not establish available capacity, delivery rights or a contractual remedy.

The case against premature grading

There are serious objections to importing the grain model too enthusiastically. Compute service changes with software and workload. A test that represents one model or input mix can mislead buyers running another. An elaborate classification system can fragment trading into so many categories that few participants meet in any one of them.

Certification also costs money and time. For small purchases, a direct test followed by an ordinary refund policy might be cheaper than a formal inspection process. Large cloud providers can internalize matching, performance measurement and dispute resolution within their own services. A neutral third-party institution must demonstrate an advantage over those arrangements rather than assume neutrality is sufficient.

The analogy has a physical limit as well. Grain can be stored; an unused hour of computing capacity expires. A reservation concerns a future service window, not an existing pile of inventory. Nor does a common performance grade settle questions of security, data location, licensing or authorization. Buyers cannot substitute away a mandatory restriction merely because another seller is faster.

What would count as progress?

The empirical test is whether a limited set of service definitions reduces the cost of buying from unfamiliar suppliers. Evidence would include repeat purchases, successful substitutions, the frequency of contested results and the time and cost required to resolve them. Better outcomes would have to survive changes in workload and supplier, rather than appear only in a demonstration.

A sensible initial experiment would choose one narrow workload class and test whether several suppliers can meet the same measurable promise. It would record rejected substitutions as carefully as successful ones. If buyers repeatedly need bespoke exceptions, that is evidence to revise the class, not hide the exceptions inside an average.

Grain’s lesson is that exchangeability is partly an achievement of governance. Compute-market builders should ask who defines an acceptable substitute, who verifies it and what happens when that verification is challenged. Those questions will not guarantee liquidity. They can reveal whether a quoted price describes something buyers can reliably purchase.