Problems addressed
- Model ambitions that exceed available VRAM or memory
- Poorly balanced GPU, CPU, storage and power choices
- Unclear privacy, access and software-stack requirements
- Buying expensive hardware before proving the workload
AI / 03 Urquhart / capability
Local AI computing
Plan local AI around models, memory and operating reality.
Architecture for local inference and AI development systems, with attention to VRAM, RAM, storage, cooling, power, privacy and the upgrade path.
Who it is for
A defined service for a recognisable technical problem.
- Developers experimenting with local models
- Businesses exploring private inference
- Researchers who need a bounded proof-of-concept environment
Included
- Use-case, model and concurrency assumptions
- Hardware architecture and constraint analysis
- Software-stack recommendation
- A staged route from feasibility to deployment
Not included
- Model performance or commercial outcome guarantees
- Training large foundation models from scratch
- Third-party API, model, licence and hardware costs
- Production deployment without separate security and operational scope
Delivery process
Reduce uncertainty in deliberate stages.
Bound the use case
Define models, context, throughput, users, privacy and budget.
Test assumptions
Identify memory, compute and software constraints before committing.
Specify
Produce a balanced architecture and staged purchasing route.
Configure
Optional runtime setup and initial validation follow the approved architecture.
Starting prices
Choose the smallest useful commitment.
Models, APIs, hosted services, licences and hardware are separate third-party costs.
See the complete price listAI or Local-LLM Workstation Architecture
Model, VRAM, RAM, storage, cooling and upgrade path
AI Workstation Software Setup
Local AI runtime and initial environment
Local AI Feasibility Review
Hardware, model, privacy and workload
Evidence and limits
Credibility comes from what can be checked.
- Architecture is tied to named model classes and explicit workload assumptions.
- Memory limits, quantisation choices and operational trade-offs are documented.
- A feasibility review can precede hardware purchase when uncertainty is material.
Questions
Before you enquire.
Do I need the most expensive GPU?
Not necessarily. The right choice depends on model size, precision, concurrency, latency and budget.
Can local AI keep data private?
Local processing can reduce external data transfer, but privacy still depends on access controls, logging, backups and the wider system design.
Can you guarantee a model will meet my needs?
No. The correct route is to define an evaluation and test representative material before treating a model as production-ready.
Start with evidence
Describe the current system and the outcome you need.
Include the main constraint, the people affected, the timescale and any existing specification. UDS will confirm the most proportionate first step.
Send the project brief