Compute Has No Kilowatt-Hour

Companion essay to the ASSAY-1 specification.

In 1882, Edison's Pearl Street station didn't sell electricity. It sold lamps — a monthly fee per light bulb connected, because there was no way to measure what a customer actually consumed. Westinghouse sold something different and incompatible. Every generator was its own little economy, every contract was bespoke, and there was no grid, no market, no arbitrage, and no way for a buyer to compare two sellers — because there was no unit.

The kilowatt-hour fixed all of it at once. Not a technology — a definition: a unit of delivered energy, independent of whether it came from coal, hydro, or gas, measurable at the point of delivery by a meter either party could read. Metering made billing honest. Honest billing made contracts writable. Writable contracts made markets. Markets built the grid. Everything we now call the electricity economy sits on top of an agreement about what one unit of the product is.

Compute — the input the world is currently spending hundreds of billions of dollars a year to acquire — has no such unit.

What compute is sold in instead

On the supply side, compute sells by the GPU-hour. On the demand side, it sells by the token — a unit each vendor defines differently, that cannot be compared across models, and that the seller's own tokenizer choices can inflate. In between sit requests, credits, and "instances." None of these measures work delivered. None converts into any other. A CFO cannot compare two AI invoices the way they can compare two power bills, and a procurement team cannot compare two GPU contracts the way a refinery compares two crude cargoes.

It gets worse, because the GPU-hour isn't even a consistent unit of capacity. Eight H100s on a shared Ethernet fabric and eight H100s in an NVLink domain with non-blocking InfiniBand are different goods — a training job that runs on one crawls on the other — yet both sell as "8× H100." In our first survey of 16 providers, the same H100 rented for $1.99 and for $12.29 in the same week. A 6× price ratio between sellers of "the same" good is not a spread. It is several different goods sharing a name.

The financial layer just made this urgent

On October 5, 2026, CME lists futures on GPU rental prices. ICE has announced a competing suite. All of it settles in cash against price indexes — which means the market now has a price for a unit nobody has defined, and no instrument that delivers one. A hedger holding a compute future during a shortage receives money, not GPUs. The gap between the index print and what you can actually take delivery of — the basis — is unmeasured, unpublished, and, for anyone signing real compute commitments, unpriced risk.

Every mature commodity market solved this with two layers: a price layer and a delivery layer. Crude has Brent and it has cargo specs, delivery points, quality differentials. Power has a forward curve and it has firm-vs-interruptible tiers and transmission rights. Compute, as of next month, has an index — and nothing underneath it.

The path to a kWh for compute

The unit can't be defined in one step, because work delivered depends on capacity, model, and workload together. The sequence is forced:

Step one: a unit of capacity. Define what a buyer actually receives when they take delivery: silicon, scale-up domain, fabric bandwidth and topology, contiguity, power firmness, terms. Grade whether cluster A can substitute for cluster B for a given class of work. This is ASSAY-1 — a specification, a registry of real capacity graded against it, and a weekly report on the gap between index prices and deliverable reality. It exists as of this month. This is the step where the fabric engineering lives, and it is the step the financial layer cannot do itself.

Step two: a unit of work. Quality-adjusted delivered intelligence — comparable across silicon, fabric, and model, the way a kWh is comparable across coal and hydro. This is genuinely hard: a token from a stronger model is worth more than a token from a weaker one, and quality is contested and moves monthly. It may be intractable in full generality; the honest version is a unit rigorous within workload classes that refuses to compare across them. ASSAY-1 already reserves the slot for it (§4.3), so the historical record survives the transition.

Step three: the economy on top. Once a unit is comparable, everything electricity got follows: metering, honest billing, contracts that reference the standard, settlement against it, physical delivery, clearing. None of it is ours to build alone — a standard succeeds precisely when other people build on it.

Why believe a spec can matter

Because it has happened exactly this way before, repeatedly. Platts started as a trade paper publishing oil prices; its assessments now settle billions in contracts. Underwriters Laboratories started by testing products nobody trusted; its mark became a purchasing requirement. SPEC benchmarks became the vocabulary of an entire hardware industry. The pattern is always the same: publish free, be right in public, become the reference, become the standard others settle against. There is no shortcut, and the sequence starts with being right — measurably, checkably right — every week.

That is what the Delivered Compute Report is. The registry describes what capacity actually exists and what a buyer would actually receive. The spec defines the terms. The record is public and every number carries its source. When the gap between the index and the deliverable becomes visible, it becomes priceable — and when it becomes priceable, compute finally starts becoming what everyone already calls it: a commodity.

Compute is pre-kWh. Someone has to write the definition.

Compute Assay publishes the Deliverable Compute Registry spec (open, versioned, free forever) and the weekly Delivered Compute Report. Corrections welcome — the registry improves by being wrong in public.