The machine is only one part of an internal AI project. A useful deployment also needs approved data, a licensed model, identity and permission controls, integrations, evaluation, backup and a team that can operate it. These preliminary GHT planning estimates are for the stated scopes, not vendor quotations or guaranteed delivery commitments.
Three purchase-based starting scopes
| Scope | Hardware allowance | Implementation allowance | Deployment planning |
|---|---|---|---|
| Private AI pilot | $5,000–$10,000 | $8,000–$20,000 | 2–4 weeks |
| Department AI deployment | $15,000–$35,000 | $20,000–$60,000 | 4–8 weeks |
| Integrated company AI | $50,000–$150,000+ | $60,000–$150,000+ | 8–16+ weeks |
The pilot assumes one document source and one workflow with a small evaluation group. The department scope adds several agreed sources, identity integration and operational handover. The integrated scope addresses multiple departments or business systems and a more substantial compute environment. Initial combined allowances are $13,000–$30,000, $35,000–$95,000 and $110,000–$300,000+ respectively.
These ranges exclude tax, travel, recurring support/software/cloud charges, major data cleanup, electrical/cooling work and separately scoped high availability. A private hosted design replaces some capital purchases with recurring infrastructure costs and needs its own cost model. GHT confirms actual scope, staffing and pricing in a written proposal.
Hardware price anchors are not complete-system promises
As checked September 6, 2026, NVIDIA lists DGX Spark at $4,699 with 128GB unified system memory and 4TB storage. This can be a compact pilot or development candidate. Unified memory is not the same specification as dedicated GPU VRAM, and the listed 90-day AI Enterprise license is not perpetual production licensing. Shipping availability must be confirmed. NVIDIA DGX Spark listing
NVIDIA's RTX 5090 Founders Edition listing shows $1,999 for the GPU alone and was out of stock when checked. It does not establish the price or availability of a complete qualified workstation. NVIDIA RTX 5090 listing
For professional or server deployments, we can evaluate Lenovo ThinkStation and ThinkSystem configurations. Lenovo specifies supported RTX PRO 6000 configurations for systems such as the SR650a V4, with different card-count and power limits. We request a current configured-system quote rather than deriving a server price from an individual GPU. Lenovo supported GPU/server configurations
Size for actual requests
Memory must accommodate more than model weights: context and attention-cache requirements affect capacity. Several GPUs do not automatically act like one transparent memory pool. We benchmark the selected runtime, model, context length and simultaneous requests before promising service capacity. Hugging Face cache documentation
What controls the calendar?
The planning clock begins after scope approval, available hardware, approved data access and usable samples. Procurement lead time comes before deployment when equipment is not available. Data cleanup, connector limitations and delayed reviews can extend the calendar.
Work moves through scope and data review, a representative pilot, infrastructure and application integration, then acceptance and handover. Fine-tuning is an optional additional phase: allow $10,000–$35,000 in implementation and 3–6 extra weeks for a bounded task with usable examples. Extra training compute is quoted separately. Training a foundation model from scratch is outside these ranges.
The proposal also names recurring maintenance, model/runtime updates, source refresh, backup tests and support responsibilities. Review the complete AI planning page or discuss a private pilot.
Loading comments…