Token Factory
Layer 05 · Compute & Inference — Net-new this year
Serve every open-weight and custom model behind one governed, OpenAI-compatible API, metered by the token and running on Navon metal across NVIDIA or AMD GPUs, so no prompt or output ever leaves the country and no single GPU vendor can gate us. Buy it three ways: a shared endpoint, a Reserved Lane that holds a share of one model for a fixed monthly fee, or a dedicated endpoint.
Capabilities
Section titled “Capabilities”- One OpenAI-compatible API for chat, embeddings, and rerank
- Open-weight model catalog, including African-language models hosted nowhere else
- Serverless shared endpoints and dedicated endpoints with an SLA
- Reserved Lane, the tier in between: a fixed monthly fee for a share of one model’s serving capacity, with a token quota in and out, a stated p95, and our right to move the hardware underneath it
- Portable inference across NVIDIA and AMD GPUs, so silicon is a choice not a lock-in
- Optimized serving: prefix caching, paged attention, continuous batching, speculative decoding, quantization
- Two completion windows on one API, asap and flex, so work that can wait is priced for waiting
- Prefix-cache-aware routing, so repeated context is served from cache and billed at the cached rate
- Tokens per megawatt published per model and per system tier, with the method next to the number and a board that refreshes nightly across T1 to T4
- Residency policy on every project and every model route, enforced per request with no silent fallback
- Fine-tuning, LoRA, and distillation, deployed in one step
- Retrieval that resolves through the Data Lake ontology
- Per-token billing, zero retention by default, usage fully observable
- Every response records the route it took, so usage and jurisdiction are both auditable
Priority workloads
Section titled “Priority workloads”- Agentic coding: Coding assistants and agents for African software teams, where the model runs in-country and the code never leaves it.
- Documents and records: Extraction, classification and summarisation over bank, insurer and government records that are not allowed offshore.
- African-language speech and translation: Transcription and translation for languages no hyperscaler serves, using models trained and hosted here.
Model catalog at launch
Section titled “Model catalog at launch”| Family | Why it is here |
|---|---|
| African-language models | Trained and hosted on Navon metal, available nowhere else. |
| Llama (Meta) | General chat and instruction following across sizes. |
| Qwen | Strong multilingual and coding performance at low cost per token. |
| DeepSeek | Reasoning and coding, where throughput matters more than latency. |
| Mistral | Small, fast models for high-volume classification and extraction. |
| Whisper and successors | Speech to text, the base for the African-language speech work. |
| Embedding and rerank models | Retrieval that resolves through the Data Lake ontology. |
This product exposes an API. The reference is generated from its OpenAPI spec (openapi/token-factory.yaml).