When an AI lab encounters a Full Syllable Inventory for the first time, the question they ask is “what is the licensing fee for the dataset?”
That is the wrong frame. And it is the single most expensive mistake an AI lab can make about Bantu — not about one Bantu language, about the whole family.
There is an FSI for Swahili. There is an FSI for Zulu. For Xhosa, Shona, Bemba, Nyanja, Lingala, Kinyarwanda, Luganda, Sesotho, Setswana, Lozi, Tonga, Umbundu, Bisa — and on through the long tail. 700+ Full Syllable Inventories — 459 released for commercial licensing today. Each one is a closed, enumerable, standardised inventory. Each one is infrastructure for its language.
The data mental model.
The data mental model goes like this: there is a corpus, you license it, you train on it, you ship the model, you move on. Data is consumed. It depreciates the moment a better corpus appears. It competes on volume. It is a transaction.
If you think of a Full Syllable Inventory as data, you will treat it like the Common Crawl Bantu shard — scrape it once, run a tokenizer over it, archive it in object storage, forget about it. You will negotiate price per gigabyte. You will ask whether you really need all 700+ or just the top fifteen. You will treat it as background.
Infrastructure is different.
Infrastructure gets used, not consumed.
The 26-letter English alphabet is not data the typewriter trained on — it is the substrate every English word references. TCP/IP is not data the router trained on — it is the protocol every packet uses. UTF-8 is not data your stack trained on — it is the encoding every string queries. The periodic table is not data the chemist trained on — it is the index every experiment indexes into.
None of these depreciate. None of them get cheaper at scale. None of them compete on volume. They compete on completeness, reliability, and standardisation.
- ·Consumed (trained on)
- ·Depreciates as better corpora ship
- ·Competes on volume — bigger is better
- ·Priced per gigabyte / per record
- ·Acquired in one transaction
- ·Lives in object storage
- ·Used (queried, referenced)
- ·Compounds — better the longer it operates
- ·Competes on completeness + reliability
- ·Subscribed to as a continuing dependency
- ·Versioned, governed, standardised
- ·Indexed into at runtime
A Full Syllable Inventory is exactly the second column. It is the closed, enumerable set of legal syllables that every Bantu word, every recording, every morpheme, every downstream linguistic operation references. You don’t train on the FSI — you query it. You don’t consume it — you reference it. You don’t out-scale it — there is exactly one for each language, and it is finite.
Three properties make this concrete.
It is closed.
Every Bantu syllable inventory has a fixed number of entries. The catalog runs from inventories of a few hundred entries to inventories of several thousand — Bemba is 480, others vary — but each language has one number, and that number is final. You can enumerate it once and reference it forever. Adding a new entry would change the language, not the dataset.
Every downstream operation indexes into it.
Tokenisation, syllabification, ASR re-ranking, TTS unit selection, morpheme parsing, loanword detection, code-switch detection — all of these read the FSI, none of them mutate it. It is the queried index, not the consumed corpus.
It is standardised.
Every published FSI in the BantuNomics catalog conforms to BTS-S100. Schema, provenance, validation. Like RFC 791 for IP. Like CIE for colour. Like ICANN for DNS. The standard is what makes it infrastructure rather than someone’s personal corpus.
Closed + indexed-into + standardised. That is the definition of infrastructure. It is the definition of what an FSI is. It is not the definition of any dataset that has ever been licensed.
What this changes commercially.
Accept that FSIs are infrastructure and the commercial model becomes obvious.
You don’t buy infrastructure outright. You subscribe to it. It gets versioned, maintained, updated, governed. The party operating it has a continuing obligation — new languages curated, new recordings ingested, new BAB rounds run against new frontier model versions. The party using it has a continuing dependency.
This is why BantuNomics is a $1.75M / yr Founding Partner subscription, not a $X-per-FSI license. We operate the registry. We run BAB against frontier models. We curate the recordings through Amina. We publish the standards. Your model depends on the substrate the way your stack depends on DNS — not the way it depends on a training corpus.
Labs that treat FSIs as infrastructure will index into them at tokenisation, at decode, at every layer where the substrate matters — for Swahili, Zulu, Shona, Bemba, Lingala, Kinyarwanda, and on through the 700+-language catalog. They will reference each inventory the way their English systems reference UTF-8 — quietly, always, per-language, without re-deriving it.
Labs that treat it as a dataset will keep asking for bulk pricing, keep training on web-scraped Bantu, and keep shipping models that score 702 BABS at best — Founding Partner candidate tier, not Production.
FSIs are not data.
They are the periodic table of the Bantu language they describe. They are the alphabet every Bantu reader carries in their head. They are the closed, enumerable, standardised index every downstream linguistic operation will reference.
They are infrastructure.
Subscribe accordingly.