Private AI on your own servers
Running open-weight models such as Llama, Qwen, DeepSeek and GLM on servers you control.
AI Solutions
Private AI and self-hosted AI models
Some organisations cannot send confidential files to an outside AI provider, usually abroad. We run open-weight models on GPU servers you control, linked to your documents, so prompts, files and outputs stay in your environment.
Who this service is for
A good fit if
- Policies, clients or regulators forbid outside AI providers processing your data, or sending it abroad.
- You need control of where AI runs (including in Nigeria), what is logged and for how long.
- You want AI over legal, health or financial files that no third party sees.
- Heavy use makes per-request pricing, usually in US dollars, costly, or you want less vendor dependence.
What is included
- Requirements: sensitivity, location (NDPA transfers, sector or NITDA rules), users, workloads, targets
- Model shortlist tested on your documents; licences checked for commercial use
- Sized GPU hosting: premises, Nigerian colocation with power backup, or your cloud abroad
- vLLM or SGLang inference with an OpenAI-compatible private endpoint
- Private networking, single sign-on, encryption, no public endpoint
- Retrieval over your documents; Microsoft 365, file share, Prometheus or ERPNext connections
What this service is
Most AI use sends prompts and documents to a provider, almost always to data centres outside Nigeria. Business terms exclude your data from training, some with zero data retention, and for many that is enough. Law firms, health organisations, banks and other financial institutions, public bodies and companies protecting product designs may need stricter, whatever the contract says: confidential information does not leave our environment. Private AI gives staff an assistant and applications an endpoint, with no outside AI provider seeing a prompt.
| Option | Where the model runs | Who sees your prompts | Best for |
|---|---|---|---|
| Enterprise AI service | The provider's cloud, usually outside Nigeria | The provider, under business terms; with zero data retention, not stored after the response (limited exceptions) | Most organisations, when the terms and a lawful transfer basis suffice |
| Managed model in your cloud account | Your cloud's AI service, in a region abroad you choose | Your cloud provider under your agreement, not the model developer | One-cloud organisations relying on transfer safeguards |
| Self-hosted open-weight model | GPU servers you control: premises, a Nigerian data centre or cloud | Only your organisation | Data that must stay in your environment or Nigeria, steady high volume, full control |
We are not tied to one approach. Many mix a commercial assistant for everyday work with a private model for what must stay in-house.
What we deploy
A private AI platform is more than a model on a server.
GPU servers
Sized for the model and peak use: your cloud, a dedicated server or your hardware, with UPS and backup power on premises.
Inference server
vLLM or SGLang serves many users through an OpenAI-compatible API.
Gateway
Your identity provider's sign-in, usage limits, logging under your rules.
Document retrieval
Answers from your policies, files and records, with sources (how it works).
Applications
A staff assistant, AI in your portal or ERP, or AI agents on your private model.
Your existing tools
Many apps, coding tools and frameworks can use the private endpoint; each is tested.
Choosing a model, and tool compatibility
- There is no single best open-weight model. We shortlist by your tasks, languages, hardware and licence terms, then test on your own questions and documents.
- Smaller models answer faster, need less GPU memory and are often enough for document questions and extraction. Larger models handle harder reasoning but need more powerful servers.
- New releases arrive often, so we keep your evaluation set and re-test when a better candidate appears.
- Licences differ, even between versions of one family, from permissive (Apache 2.0, MIT) to custom terms with conditions, so we check the licence of the exact version you will rely on.
- OpenAI-compatible is not identical: chat formats and tool calling differ between models, which is why each integration is tested.
Security, data location and running it
Closed by default
Reachable only from your network or approved apps, behind single sign-on.
Trusted model files
Official sources, safetensors, no unreviewed model code, files recorded; injection and leakage tested.
Your retention rules
You decide what is logged and for how long, even no prompt content.
Keeping AI in Nigeria
No hyperscale cloud region here, so in-country AI runs on your own or colocated servers; nothing routed abroad, every location documented.
Where the model can run
- On GPU servers on your premises, colocated in a Nigerian data centre, or in your own cloud account (such as AWS, Microsoft Azure, Google Cloud or DigitalOcean).
- Most AI services and cloud GPUs process data abroad. AWS's Lagos Local Zone offers only entry-level NVIDIA T4-class GPUs, so larger cloud GPUs mean a region abroad.
- Safetensors stores only the model's numbers, not code that could run when the model loads.
Data protection law still applies. Self-hosting keeps data in your control, but the Nigeria Data Protection Act 2023 (NDPA) and the NDPC's GAID 2025 still govern personal data used in AI, including a data protection impact assessment (DPIA) and documented safeguards; servers abroad also need NDPA transfer safeguards. From 1 January 2027, NITDA's National Cloud Technical Guideline applies data-residency rules to federal public institutions and regulated-sector organisations. General information, not legal advice; see AI governance and security.
A private model is a system you own, so it needs an owner. We hand over with a runbook, or run it under a support or managed IT services agreement: patching, monitoring (availability, speed, GPU use, cost), capacity, backups and testing new releases first.
Delivery
How the work is delivered
Assess
What data the AI touches, where it must stay, who uses it, and whether enterprise terms are enough.
Output: Recommendation, shortlist, hosting options.
Prove
Candidates on a temporary GPU server in your environment, tested on your questions beside a commercial model.
Output: Accuracy, speed, running-cost results.
Build
Production servers, private networking, sign-in, logging, document retrieval and integrations.
Output: Production-ready private AI platform.
Pilot
A pilot group uses it; we review answers and feedback and tune models, prompts, retrieval and capacity.
Output: Pilot report, improvements.
Run and improve
Monitor, patch and back up; new model releases tested against your evaluation set before switching.
Output: Runbook, monitoring, update plan.
Scope at a glance
Not included unless agreed
- You buy GPU servers, hardware and cloud usage (often in US dollars); we can manage them
- Training new models from scratch
- Legal advice on licences or data protection law: we flag terms for your adviser
What we need from you
- Your data classification and data protection, security and data location requirements
- A cloud account, data centre or server room, or approval to set one up
- Example questions with expected answers, and an owner who judges quality
Before handover
- Endpoint reachable only from your network or approved apps; every user signs in
- Tests confirm no prompts, files or outputs reach an outside AI provider
- Accuracy and peak response times meet targets; retention, costs and scaling documented
What affects the cost
- Model size and the GPU memory needed
- Users, peak load and response-time target
- Always-on or not; cloud, colocation, or owned hardware and power
- Retrieval, integrations, security and ongoing operation
No package prices: estimates in naira from an agreed scope, taxes shown separately.
Questions buyers usually ask
What is an open-weight model?
A model whose developer publishes the trained weights, so you can download it and run it on your own hardware under a licence. Families include Meta's Llama, Alibaba's Qwen, DeepSeek, Z.ai's GLM, Mistral, Google's Gemma, OpenAI's gpt-oss and IBM's Granite.
Are open-weight models as good as ChatGPT or Claude?
For many business tasks (answering from documents, summarising, drafting, extracting data), yes. Independent evaluators report the best open-weight models trail leading commercial ones by a few months on average, more on the hardest reasoning and coding. We test on your questions, not benchmarks.
If we run DeepSeek or Qwen, does our data go to China?
Not when you host it: a self-hosted model sends nothing to its developer, unlike its app or API. We know of no Nigerian rule specific to Chinese-developed models, but your policies, clients or international partners may restrict some developers, so we check first and security-test every model.
Is self-hosting cheaper than paying per use?
Only at steady, high volumes: a GPU server costs the same for ten questions or ten thousand, so for light use an enterprise service with zero data retention is usually cheaper. Both are usually priced in US dollars, so exchange rates affect each (how to work it out).
Related
Self-hosted AI model or AI service?
Self-hosted open-weight model or AI service? Compared on privacy, quality, cost and effort.
Generative AI development and custom AI assistants
Custom AI assistants that answer from your approved documents and data, with sources, permissions and testing.
Need AI that never leaves your environment?
Tell us what data the AI will use, where it must stay and who uses it. We will arrange a conversation and suggest private hosting, enterprise terms or both.