Dawnstack

Frontier-quality AI is technically available everywhere and practically available to people with a dollar-denominated credit card. That is a payments problem wearing an infrastructure costume.

Dawnstack is an OpenAI-compatible inference API where you top up a prepaid wallet with mobile money, get an API key, and are billed per token in local currency. No dollar card, no monthly minimum, no commitment. You change one URL in code you already wrote.

What is interesting about it

The metering. Balances are integer micro-dollars, never floats. Deduction is a single atomic conditional update: subtract the amount where the balance is at least the amount, and treat “no rows changed” as insufficient funds. Reading a balance and comparing it in application code loses under concurrent requests against the same key, which is exactly the situation an API is in. There is a double-entry ledger behind it and idempotency keys in front.

Licences as an executable gate. The model registry is a typed constant rather than a database table, specifically so that marking a model as commercially usable requires a reviewed diff. Two models are blocked as non-commercial and deliberately left in the file, so the verdict stays auditable rather than disappearing.

The tokenizer tax. Billing is per token, and tokenizer vocabularies are built mostly from English text. So the same sentence costs more tokens in some languages than others, which is a silent price multiplier applied to anyone working in an under-represented language. I measured it across seven tokenizers and published the result as language-cost data on every model in the live API: Swahili runs about 1.65x English, Yoruba about 2.45x. You can read it yourself at GET /v1/models, no key required.

Then the obvious conclusion turned out to be wrong. Routing to the model with the best tokenizer produced worse bills, because per-request cost is not per-token cost: one model burned 800 output tokens to return what another said in 99. Cost per answer is the number that matters, and I only found that by measuring it.

Built solo, 161 commits, first commit to live in nine days.