Generative AI on your own GPUs, inside your own perimeter

Sending customer, payment or internal data in a prompt to a service hosted abroad is processing of that data outside Nigeria, which the directive is designed to prevent. Whether a specific use falls in scope depends on the data involved. Running the model on your own infrastructure removes the question entirely. This is not legal advice; confirm the interpretation with your compliance function.

For retrieval-augmented assistants, document processing, classification and summarisation, current open-weight models in the 8B to 70B range perform at the level required, and we measure that on your own evaluation set before committing. For scoring and anomaly detection, the models are yours and were never dependent on a hosted LLM. Where a hosted model is genuinely better, the gap is usually smaller than the compliance cost.

A private assistant for a few hundred users runs comfortably on two to four data-centre GPUs with a quantised 8B to 32B model. Larger models or thousands of users need more; fraud scoring on Triton often needs none, running on CPU. We size from measured token throughput and latency targets rather than rules of thumb, and MIG lets one GPU serve several small workloads.

New weights, images and patches are downloaded on a connected staging host, hash-verified, scanned and signed, then transferred on encrypted media into the enclave, where the signatures are checked again before registration. Each import is recorded with who did it and when. On a monthly cadence this becomes routine.

Every model in the registry carries a model card with its licence, training data description, evaluation results and approver. Every inference carries the model version, and prompts and responses are logged to an audit index. For scoring models, feature importance and drift reports are produced for the model-risk function.

NuxFamily engineers provide the support directly on the inference servers, GPU stack, vector index and MLOps components, from first response to root cause. Coverage includes security patches, version management, controlled model updates and, on the Mission Critical tier, a named engineer and assistance in CBN inspections and model-governance audits. The exact SLA terms are set out in the support agreement.

Yes. The same GPU pool supports parameter-efficient fine-tuning (LoRA) for classification, extraction and domain style, tracked in MLflow. We recommend starting with retrieval and prompt design, which solve most assistant use cases, and fine-tuning only where an evaluation shows a measurable gain.