The thesis in one sentence
For two centuries, a company built capital on two pillars: labor (person-hours) and physical assets (machines, buildings, inventory). The balance sheet measured what was tangible; the income statement measured how much labor it took to produce revenue.
Generative AI breaks that framework at its foundation: cognitive work — writing, classifying, reasoning, deciding — is now produced in tokens, at a cost trending to zero and with a virtually infinite supply. If anyone can rent the same foundational model and generate the same tokens, the raw token is not capital: it is a commodity.
Capital emerges when the token stops being ephemeral and becomes something that compounds: a captured human judgment, an evaluation datapoint, a fine-tuned model that does tomorrow for free what today cost dearly.
A token spent and forgotten is expense. A token spent, governed and captured is capital. The difference is not in the model — it is in the architecture of the company.
Not all tokens are equal
The first accounting error of the AI-native company is treating all tokens as the same input. There are two classes, with opposing economics — and the company's health is measured by the ratio between them.
| Dimension | Foundational token (rented) | Local token (owned) |
|---|---|---|
| Marginal cost | High, per million tokens | ≈ 0 after hardware amortization |
| Nature | OPEX — consumed and gone | CAPEX — yields in perpetuity |
| Ownership | Not yours; depends on a provider | Yours; no network, no fee, no leakage |
| Sensitive data | Leaves your perimeter (GDPR, IP) | Never leaves your infrastructure |
| Optimal use | Hard reasoning, bootstrapping | Routine, high volume, distilled tasks |
The master metric is not "spend less on AI." It is the ratio of local work to foundational work. The more it shifts toward local, the greater the autonomy, the lower the cost, and the more proprietary knowledge compounds.
Three forms of durable capital
If the raw token is a commodity, where does capital accumulate? In three deposits that fill as a byproduct of daily work — but only if the company is built to capture them.
1. Captured human judgment. Every time a human approves or denies an agent action at the governance gate, they produce a label: a tagged decision about what was correct. That judgment is the most expensive data in AI — and here it is generated for free, as a residue of work.
2. Verifiable audit trail. Not just compliance: it is the evaluation dataset that no competitor can rent. A verifiable hash chain capturing every decision, every action, every outcome.
3. Fine-tuned models. The distillate of foundational work turned into local capability: a proprietary model that, once trained, yields tokens almost for free forever on already-mastered tasks.
The vertical wedge
Start narrow and deep. One agent excelling at variant classification for a handful of genes is worth more than a mediocre one attempting all of genomics. Depth is what makes the specialized local token outperform the generic foundational token — and only then does the wedge widen.
At QMetrika, that vertical is RNA thermodynamics. The nearest-neighbor ΔG engine (patent P202630522) feeds three research sub-lines, four publications with permanent DOIs, and a dual optimizer that outperforms commercial vaccines. But the Capital Q model is general — it is the same equation in any domain where depth builds a moat.
In QMetrika's science, the free energy (ΔG) of an RNA molecule is profiled to extract biological knowledge. In the company, computational energy (tokens) is spent to extract organizational knowledge. It is the same equation on two layers.
Eight operating principles
These are not aspirations — they are architectural decisions that must be made at build time, or not at all.
The token is capital, not expense
Every foundational spend must leave a deposit: label, eval, model. If it is consumed without capturing anything, it was pure cost.
Measure local vs foundational
The ratio of local work to rented work is the health metric. More local = more margin, more autonomy.
Control lives in the tool
The decision is deterministic code at the gate, not LLM judgment. Fail-closed by design.
Capture human judgment
Every approval is recorded, labeled and reusable. The label is a residue of work.
Audit from day one
Verifiable chain. Not just GDPR: it is the evals dataset that no generic can rent.
Narrow and deep
The vertical wedge is where capital compounds. Widen only when depth already yields returns.
The moat is the data
Models expire; captured judgments and evals do not. A better model only makes your capital yield more.
Design before spending
Canvas before code. Ambiguity is paid for in burned tokens.