AI infrastructure built in Lagos

Safety-alignment data
does not transfer
to African languages.

Frontier model safety refusal collapses 55 points from English to Igala on matched harmful prompts. The corpus that closes that gap — native preference data from owned African workforce, processed on sovereign GPU — does not yet exist. Until now.

Phase 1 deployment Q3 2026
Stack Labor · Data · GPU
Governed by Nigerian law
90%→35%
Safety refusal collapse, English to Igala (LSR Benchmark, arXiv:2603.19273)
64×
A100 80GB SXM4, Tier III Lagos datacenter, 30kW allocation
$0.05
Per kWh solar vs $2.37 diesel — solar-powered compute economics
$14.3B
What Meta paid for access to human feedback — the market UUAMNI serves
The structural problem

The refusal centroid is anchored to English. It does not transfer.

Every AI system on earth runs on human feedback. The constraint is not compute — it is genuine human intelligence at scale. Current frontier models refuse harmful prompts ~90% of the time in English. In Yoruba, Hausa, Igbo, and Igala, that refusal rate collapses to 35–55% on matched prompts.

The paper formalizes this as Refusal Centroid Drift: safety-alignment representations are anchored to English token sequences and do not transfer cleanly to tonal, low-resource West African languages. The same gap runs the other direction: multilingual preference data lifts win rates across all 23 languages in the training set, not just the targets.

The corpus that learns from more of the human world reasons better in every part of it. No one has built this work for any African language yet.
The complete infrastructure stack

Three layers. One site. Zero external dependencies.

If any layer is separated, someone else controls whether the teaching continues. Together, they make the generative future self-sustaining.

🧠

RLHF Annotation Workforce

Nigerian annotators trained on tonal accuracy, dialect identification, and cultural reasoning. First Igbo-origin DPO expert annotators onboarding now via SkillUpImo's trained graduate base and Masakhane networks.

$5–8/hrcost
$12–18/hrbill rate
40–55%margins
📚

Igbo-Origin Preference Data

The first open Igbo-origin DPO dataset, sourced from 1,500+ native Igbo proverbs (ilu) as reasoning prompts. Not translated English. CC-BY-4.0 public sample, commercial tier for frontier labs.

1,500+native proverbs
DPOformat
CC-BY-4.0public tier
🖥️

GPUaaS, Owned and Local

64× A100 80GB SXM4 at Tier III Lagos datacenter. Colocation locked. Solar-powered. +35% sovereignty premium for regulated customers — banks, NDPC, government.

$0.05–0.09/kWh solar
+35%sovereignty premium
Nigeriadata sovereign

Fair wages flowing into local economies. Careers, not gig work. Excess solar distributed to communities, hospitals, schools.

This is not just infrastructure. It is the economic argument for African language AI built by Africans, for Africans, on African soil.

Oil powered the last century.
Human intelligence powers the next one.

The infrastructure is being built in Lagos. The workforce is being trained now. The data layer is being released. The window is open — and it will not stay open indefinitely.