Accelerators and models
A procurement office can specify accelerators, benchmark them and award a contract. Models can be procured or licensed the same way.
02Chapter 2.0 · Market Diagnostic
The intelligence asymmetry
Why licensed, provenance-bearing human knowledge — not compute — is the durable constraint, and where the protocol begins.
2.1 · The data wall
A depleting stock, not a growing one
The supply of high-quality training text behaves like a depleting stock, for three reasons that compound.
01
Exhaustion
Indexed public web text is finite, and its highest-quality portion — edited prose, technical documentation, refereed literature — was crawled early. What remains is disproportionately low-signal, templated or commercially gated.
02
Contamination
As generative systems publish a growing share of new text, the ratio of human-authored material in the crawlable pool deteriorates. Recursive training degrades coverage, and the tail goes first.
03
Legal
Unlicensed scraping is closing unevenly across jurisdictions. Regulated procurement increasingly requires documented provenance, turning an informal cost advantage into a formal cost for anyone who cannot produce it.
Falsification test
If large enterprise buyers keep accepting undocumented training data without a price or compliance penalty, the premium the protocol depends on narrows.
2.2 · The representation gap
Speakers are not the same as value served
Training material across languages is far more skewed than the distribution of speakers. A small set of languages holds the overwhelming majority of indexed text, while most living languages have effectively no representation.
Where representation is absent, capability is absent — and users compensate manually for degraded output. That compensation never appears in a published statistic, which is why the market looks smaller than it is.
Material is produced locally at local cost, collected without payment, aggregated elsewhere and returned as a product priced for high-income markets. This is a pricing observation: the first party to pay for the material should face lower collection costs than the last party to take it for free.
Falsification test
Contributor retention after the incentive taper — a measured quantity the Chapter 11 phase gates require.
2.3 · The missing layer
Compute is procurable. Data is not.
States now fund AI as public infrastructure — compute programmes, national models, talent pipelines. What these programmes systematically under-provision is the data layer.
Local adequacy
The material that makes either useful in a given language and administrative context does not exist as inventory. It is latent capacity spread across a population — hundreds of thousands of contributor-hours across hundreds of languages and dialects.
Capacity without local adequacy.
Sovereign programmes routinely fund the two layers that can be bought and leave the layer that must be produced. A protocol does not replace a national programme and does not compete with one. It supplies the input those programmes need and cannot generate internally.
The commercial consequence is a counterparty that is not another data network, and a procurement motion that does not run through developer adoption at all.
Falsification test
If no sovereign or quasi-sovereign programme contracts for external corpus supply within the first two phases, Layer I revenue rests on private licensors alone.
2.4 · The anchor market
Selected against six criteria, not chosen
The criteria are stated in full so a reader who disagrees can substitute their own weighting and see whether the conclusion survives.
| Criterion | Why it matters | What disqualifies a market |
|---|---|---|
| Linguistic fragmentation | Sets the size of the unserved corpus and the collection surface | Monolingual markets offer no marginal collection value |
| Micro-enterprise density | Sets the depth of Layer II demand and task-generated evidence | Formal-enterprise economies offer few agent-deployable workflows |
| Mobile payment rail maturity | Decides whether Layer II can bill without building a collection stack | Cash-dominant markets make small-ticket billing cost more than it earns |
| Policy momentum in AI | Decides whether licence demand includes a public counterparty | No state AI programme removes the highest-margin licence class |
| Mobile and network penetration | Decides whether contributors can join without hardware subsidy | Populations without device access cannot enter the Commons |
| Cost of agent delivery | Decides whether Layer II margins can fund Layer I collection | High inference, connectivity and support costs invert the model |
The anchor is described as a high-density maritime emerging market archetype. Using the archetype rather than an identity forces every downstream claim to be stated as a property — testable against any market that shares it.
The assumption is testable in Phase I and its failure is recoverable: the collection mechanism ports across markets in the same archetype, while licence relationships do not.