## Why EU first The wedge is open-weights inference, OpenAI-compatible, served from somewhere that isn't us-east-1. The fastest path to that is two GPU racks in Helsinki and Frankfurt — neoclouds we trust, residency that satisfies most enterprise contracts on day one, and ~40ms to most of the EU and West Africa over the right paths. Africa-resident GPU is the strategic destination. We are explicit about this: it does not make sense to commit to a Lagos POP before we have ten paying customers asking for one. EU is where we earn the right to build the African footprint we actually want. > We chose the unsexy thing first. EU residency, a clean OpenAI shim, and a > pricing page that does not say "starting at." Africa POPs follow demand, > not press releases. ## What shipped in v1 The launch surface is small on purpose. Four things are in production: - **OpenAI-compatible `/v1`** — `chat.completions` (streaming and non-streaming), `embeddings`, `audio.transcriptions`. Pass the same payload as OpenAI; change the base URL. - **42 open-weights models** across chat, embeddings, rerankers and speech, with a deliberate bias toward African languages: `afrispeech-tts`, `yoruba-instruct`, `swahili-bge` and friends. - **Two regions** — `hel1` and `fra1` — with active-active routing on the public endpoint and per-region overrides via `X-Tomoul-Region`. - **Per-second compute and per-token serverless**, on the same API key. The pricing page is one page. There are no tiers. ## The numbers Eight weeks of internal load testing, plus three weeks of design-partner traffic. Median numbers from the last 30 days of production: The tail is the work. p99 TTFT sits at 184ms across both regions today — that is acceptable for chat, generous for embeddings, and the number we are most focused on shrinking in the next two months. A post-mortem from March on a 38-minute regression is up [here](/blog). **EU residency is genuine.** No US fallback, no shadow replicas. If `hel1` and `fra1` both go down, requests fail rather than silently leave the region. We think that is the right default for residency claims. ## The Africa sequence We are not the African AI cloud. We are an inference cloud that takes Africa seriously as a capability — in models, in payments, in regions, and in language coverage. The order matters. Models and payments come first because they unlock customers; regions follow customer demand. EU launch · hel1 + fra1, note: "Production traffic. OpenAI-compatible API, Stripe + USDC payments.", }, { done: true, date: "June 2026 — DONE", title: "Flutterwave and Paystack billing", note: "Local-currency invoicing for NG, KE, ZA, GH and CI. M-Pesa rolling out next.", }, { date: "Sep 2026 — Q3", title: <>lagos1 private beta, note: "Tier-3 colo, single H100 rack. Invitation-only for design partners with West African residency requirements.", }, { date: "Dec 2026 — Q4", title: <>lagos1 GA · nairobi1 private beta, }, { date: "2027 — H1", title: <>joburg1 · cairo1, note: "Subject to demand validation. We will say no to regions that don't have customers.", }, ]} /> ## What's next Three things on the near-term roadmap, in order of how loudly customers have asked: 1. **Fine-tuning.** LoRA first, full-parameter later. Region-pinned, billed per-second. 2. **Function calling and structured outputs** across every chat model, not just the ones whose authors shipped it. 3. **The `tomoul` CLI** moves from public beta to GA. Same flags as `ollama`, points at local or cloud with one switch. If you want to build on top of this, the docs are at [docs.tomoul.io](/docs) and the console is at [console.tomoul.cloud](/signup). Or grab the CLI: `brew install tomoul`. --- Questions or feedback — [drop us a line](/contact). We read every email.