We are building a full-stack AI company in public.

The top of the stack is where the value actually lands. Domain agents, reusable skills and vertical applications, running on your own infrastructure instead of somebody else's API.

This is the layer people usually buy first and regret later, because an agent is only as sovereign as the compute underneath it. Building down from the application is how organisations end up with their data in someone else's region. We are building up to it instead, which is slower and is the right order. This layer is in design and not yet available.

#agents #fullstack

Sovereign data, modelled as connected objects and the actions on them. Not a warehouse you query, but an ontology that knows what a patient, a meter, a claim or a student is inside your organisation, and what may be done to each.

Above it sit the models: open-weight and custom, including African-language models. A continent whose languages are missing from the frontier models will keep paying to translate itself into somebody else's. Both layers are in design.

#dataontology #africanlanguages

Security is not a layer, it is a fabric. It wraps every other layer, from the metal to the applications, and it is quantum-safe by design rather than by later migration.

The reason to settle this at the start is that harvest-now, decrypt-later is already happening. Traffic captured today can be stored until a machine exists that can open it. Anything carrying a twenty-year confidentiality requirement, which covers most government and health data, has to be post-quantum now or it is already exposed.

#postquantum #sovereignty

Managed inference, metered by the token. You send requests and pay for tokens, and we deal with the GPUs, the batching and the model serving underneath.

The part that matters commercially is that it is portable across GPUs. You are buying inference rather than a particular card, so a generational change in silicon becomes our problem to absorb instead of your migration to run. This is also the layer where owning the foundation pays for itself, because the token price is set by our power cost and our PUE rather than by a hyperscaler's margin.

#tokenfactory #inference

Committed capacity is open on the cloud. CPU, GPU and storage on our own metal, from one control plane. You commit to a monthly floor in GPU-hours, vCPU-hours or storage terabytes and pay a discounted unit price on it for a one-year or three-year term.

Anything above the floor bursts at the on-demand rate, so a busy month costs you more compute and not a renegotiation. Scheduled maintenance and announced regional events do not draw the floor down. If you move from a one-year commit to a three-year one, the spend you already locked carries forward against the deeper discount with no reset and no penalty.

#privatecloud #committedcapacity

The first 400 kW module is taking reservations. That is twenty racks at 20 kW and above, which is enough density for GPU training and inference rather than the 5 kW-per-rack assumption most regional halls were built around.

If you would rather bring your own hardware than buy our cloud, this is the layer you buy. You take the capacity you need now and add modules as you grow, so you are not paying rent on an empty hall while you wait for demand. Expansion zones on the campus are already planned, which means growth does not involve a relocation or a maintenance window.

#colocation #tieriii

Carrier neutrality is a design decision, not a brochure line. The campus has a Meet-Me Room and a separate telco transmission room, so carriers land their own equipment and you cross-connect to whichever of them you want.

Liquid, Safaricom and SEACOM are the anchors. You choose your routes, you multi-home across more than one of them, and you can switch carrier later without moving a single rack. If a provider owns your only route out, it owns your pricing too.

#carrierneutral #meetmeroom

The data centre is built in a factory, then trucked in. A complete Tier III facility arrives as three insulated steel modules that were assembled, witness-tested and commissioned before they left the production line.

This is why the timeline is nine to twelve months instead of the twenty-four to thirty-six a traditional build takes. Civil works happen on site at the same time as the modules are being made, so the two schedules run in parallel rather than end to end. Our manufacturing partner, Delta Electronics, has delivered more than 300 Tier III deployments across five continents.

#modular #witnesstested

Power is the reason this site exists. Naivasha sits in Kenya's geothermal heartland, so roughly 95 percent of the baseload is renewable, and more importantly it is steady. Geothermal does not stop at night or wait for wind.

A stable renewable baseload plus a verified PUE of 1.29 is what brings the cost of a compute-hour down, and that number carries all the way up the stack into the price of a token. The land is a thousand acres inside a Special Economic Zone, which gives duty-free equipment imports and a clear jurisdiction for data residency, with 24-hour security, perimeter control and CCTV by Securex.

#geothermal #sez

You cannot build a sovereign AI company on rented compute. Every layer above the metal inherits the jurisdiction, the pricing and the failure modes of the metal it runs on. If that metal belongs to someone else, in someone else's region, the sovereignty is a contract term rather than a physical fact, and contract terms get renegotiated.

So we started at the bottom and are building up. Nairobi is at capacity, demand for compute is not, governments are legislating for data residency, and Tier III supply outside a handful of hubs is thin. The window to secure regional capacity on your own terms runs from now to 2030.

One more thing worth saying at the start. If you only want the foundation, that is a complete product and not a trojan horse. We run the building, the power and the cooling, and what you build above the rack is yours to own, brand and sell. Everything higher up the stack is something you opt into. Your customers stay your customers.

#builtinpublic #sovereigncompute