A Common Basis
for AI.

I. The Problem

AI is unusually substitutable at the point of consumption, yet unusually difficult to model as a long-term economic resource.

Large language model inference has a useful architectural characteristic: the model is an enormous computational object, while the context supplied to request an inference is comparatively portable. At an abstract level, an application can send substantially the same context to Model A, Model B, or Model C.

But substitutable does not mean equivalent. Models vary in computational intensity per token, native tokenization, tokens required for a task, output quality, latency, specialization, reasoning effort, modality, and price. Their economics are also affected by the separate market for accelerators, energy, networking, infrastructure utilization, and inference optimization.

These relationships are unstable through time. Models and hardware turn over rapidly. A model that is economically attractive today may be superseded within months; the accelerators serving it may be replaced by a generation with materially different economics.

The value numerator is necessarily user-specific. A coding model can create enormous value for one business and little value for another. A universal standard should not decide that value. The cost denominator should be more observable — but today it is not standardized.

A business may know it will consume substantial AI in three or five years while having little ability to know which models it will use, what hardware will serve them, how many tokens a task will require, or what their relative prices will be. The problem is not only that future AI cost is uncertain. There is no common economic unit in which that uncertainty can be expressed.

II. What a Solution Requires

A useful standard for the economic cost of AI consumption would need to satisfy several unusual requirements at once.

• ObjectiveGrounded primarily in observable market variables, not a universal judgment about model intelligence or usefulness.
• Model-independentUseful as individual models gain or lose market share, improve, or disappear.
• Compute-independentNot tied to a particular GPU, accelerator, or hardware architecture.
• ObservableCapable of being calculated from market data and, ultimately, verified through quotes, transactions, and utilization.
• DenominableA discrete economic quantity against which spot and future prices can be quoted.
• Usage-weightedRepresentative of what the relevant market actually consumes.

III. What It Could Enable

A historical economic record

A generalized measure could make changes in AI cost comparable even as individual models and hardware generations turn over.

Long-term commercial contracting

Consumers could contract for future AI consumption without selecting the eventual model today; providers could express future capacity without permanently tying it to contemporary hardware.

Forward price discovery and risk transfer

Market participants could quote expected prices at future dates. Subject to market development and regulation, standardized references could eventually support structures for transferring future price exposure.

Relative model economics

A common cost basis would let each user compare a model’s premium or discount with the economic value that model creates for its own workload.

Budgeting, investment, and routing

Enterprises could model future AI consumption without independently forecasting every model and accelerator. A common unit could also support procurement across providers, model allocation, compute routing, and, eventually, more fungible contracted capacity.

IV. A Common Compute Basis

CCBT is proposed as the stable denominator: a market-derived common basis for heterogeneous AI consumption.

The standard does not need to predict which model will win, which accelerator will dominate, how quickly capability will improve, or what a unit of intelligence is worth. Attempting to answer those questions inside the standard would make it less objective and less durable.

Instead, CCBT proposes to combine observable standardized model-token prices within defined classes, weighting them according to qualifying utilization. The resulting denomination would evolve with the market rather than depend on today’s technologies surviving.

Real compute is not perfectly fungible. Geography, latency, capacity guarantees, scarcity, availability, credit, data residency, and provider quality matter. Those differences can be expressed as a basis — a premium or discount — around a common reference rather than embedded in the definition of the unit itself.

The objective is not to eliminate the complexity of AI economics. It is to give that complexity a price.