GPU / Compute Support
The Hub's core compute capability – shared, governed GPU access on an 8× H200 node, with scheduling, time-slicing, an open-source serving stack, and in-country data residency.
The Hub's foundational capability is shared, governed GPU access. A single 8× NVIDIA H200 SXM5 node is operated as a multi-tenant resource so that startups, researchers, and national use cases can reach high-end compute they could not economically provision alone. This is the capability that answers the core problem NAISH was created to solve – the compute-access gap that has forced Nigerian teams onto expensive foreign cloud, billed in dollars, with data leaving the country.
What the Hub provides
- High-end accelerators – 8× H200 SXM5, 1,128 GB total HBM3e, and a 900 GB/s NVLink full-mesh fabric for tightly-coupled multi-GPU training. This is training-grade hardware, not just inference silicon.
- Sub-GPU sharing – time-slicing via the NVIDIA Device Plugin lets one physical GPU serve several lower-intensity inference workloads at once, so small services do not each tie up a whole accelerator.
- Scheduled training windows – the Volcano / Kueue batch schedulers place heavy training and fine-tuning jobs off-peak, on a workload calendar, so perpetual inference services are not degraded by bursty training.
- A managed software stack – Kubernetes, vLLM (high-throughput inference), JupyterHub (notebooks), MLflow (experiment tracking), MinIO (on-node storage), and Keycloak (identity), with Prometheus / Grafana / DCGM telemetry. Teams get a working platform, not bare metal. See Software & Platform.
Who it serves
The node is deliberately multi-tenant. It supports the seven identified use cases, but also broader startup and researcher access, and startups matched with government institutions. Allocation across these tenants is what the governance framework exists to arbitrate.
How this differs from a rented cloud GPU
| Rented cloud GPU | NAISH shared node | |
|---|---|---|
| Allocation | First-come, pay-as-you-go | Governed quotas & priority tiers |
| Billing | Dollar-denominated, per hour | National resource, not per-hour billed to users |
| Data location | Off-shore | On-node (MinIO), in-country |
| Support | Self-serve | Embedded technical support |
The governed, sovereign model is the point: it keeps sensitive Nigerian data on national infrastructure and directs scarce GPU time toward high-impact national use cases rather than whoever swipes a card first.
For the full hardware and capacity picture, see The Compute Node, Workloads, and Capacity.
Data Readiness
How the Hub assesses whether the data behind a use case is available, curated, legally usable, and pipeline-ready – often the true pacing constraint rather than compute.
Voice & Local-Language Datasets
The Hub's capability to integrate Nigerian-language speech and text data – building AI that works in Hausa, Yoruba, Igbo, and more, for inclusive national reach.