We build private AI deployments around your approved information and business workflows, with custom applications, document retrieval and optional model fine-tuning. Choose on-site, private hosted or hybrid infrastructure with clear data and support boundaries.
Retrieval-augmented generation, or RAG, finds approved company information and supplies it to a model when answering. It supports source citations and refreshed documents without changing model weights. We usually evaluate this first for internal knowledge.
Train an existing model further on approved examples when the evaluated task needs more consistent behavior, formatting or classification. This requires suitable data, a compatible model license and a separate evaluation. It does not replace access controls or current source information.
Build the screens, identity integration, document connectors, review queues and reporting your team needs. Start with read access; changes to business records follow an agreed permission and human approval process.
Training a foundation model is a separate research and infrastructure engagement. It requires its own feasibility study, data strategy, compute budget and delivery plan; it is outside the implementation ranges below.
These preliminary GHT planning ranges describe a purchase-based deployment. They are estimates for the stated scope, not manufacturer quotes, fixed offers or guaranteed delivery dates. A workload review and written proposal establish the actual price.
| Starting scope | Hardware | Implementation | Combined initial | Deployment |
|---|---|---|---|---|
| Private AI pilot ↗ | $5,000–$10,000 | $8,000–$20,000 | $13,000–$30,000 | 2–4 weeks |
| Department AI deployment ↗ | $15,000–$35,000 | $20,000–$60,000 | $35,000–$95,000 | 4–8 weeks |
| Integrated company AI ↗ | $50,000–$150,000+ | $60,000–$150,000+ | $110,000–$300,000+ | 8–16+ weeks |
Timeframes start when the scope is approved, hardware is available, data access is authorized and usable samples are supplied. Procurement lead times and delayed client review extend the calendar. Allowances exclude tax, travel, recurring support/software/cloud costs, major data cleanup, electrical/cooling changes and separately designed high availability. Hosted options replace some hardware purchases with recurring infrastructure charges and require their own quote.
One approved document source, one defined workflow and a small evaluation group.
Hardware approachCompact local AI system, such as an NVIDIA DGX Spark, or a qualified GPU workstation. Selection follows a model and workload test.
Acceptance checkAnswer representative questions with sources, reject restricted documents, measure response time and record the pilot decision.
Explore package details ↗Several agreed sources, identity integration, permission-aware retrieval and one department workflow.
Hardware approachConfigured professional GPU workstation, including Lenovo ThinkStation options, plus backup and power protection as scoped.
Acceptance checkVerify user permissions, citation quality, simultaneous requests, content refresh, recovery and administrator handover.
Explore package details ↗Multiple departments or business-system integrations with stronger operating and recovery requirements.
Hardware approachValidated multi-GPU server architecture, including Lenovo ThinkSystem candidates. Power, cooling, storage and networking are designed together.
Acceptance checkTest each integration, approval boundaries, load, recovery and the agreed operational monitoring. High availability is a separately scoped design.
Explore package details ↗Allow an additional $10,000–$35,000 in implementation and 3–6 weeks for a bounded task with usable approved examples and held-out evaluation. Suitable existing compute may be reused; additional training compute is quoted separately.
We recommend this phase only after a baseline test demonstrates a problem that training can reasonably address.
Local hosting alone does not establish privacy or compliance. Each design records where prompts, documents, logs, backups and support access reside. Model answers can be wrong; the workflow includes review and escalation appropriate to the task.
