A fixed budget forces honesty. At a mid-range local AI budget you cannot buy maximum parameters, silent cooling, and infinite headroom at once. You can buy a usable chat and light RAG box if you write requirements before you open shopping tabs.
Start from workload, not from GPU marketing
Decide whether the box is for one power user or a small team taking turns. Decide context length targets and whether embeddings must run alongside chat. Map those needs with match tools and the VRAM table before you fall in love with a card. Compare the same money against budget-3000 and budget-5000 style builds, and against Mac options if mobility matters more than slot upgrades.
Spend where failures hurt
Power supply headroom, case airflow, and a healthy SSD for weights prevent "mystery reboots" that look like model bugs. DIY assembly notes under diy are as important as the GPU line item. Leave a slice of budget for cables, a surge-protected strip, and dust filters. If you buy second-hand GPUs, fund a return window and a burn-in afternoon.
Software expectations
Plan the runtime—Ollama, llama.cpp, or a desktop wrapper—using tutorials and install. Productized desktop flows are summarized at microai-one. Do not promise multi-user heavy RAG on day one if VRAM is already tight with a single chat model.
Extra
Write the tradeoffs on a sticky note taped to the case: what this box will not do. When someone asks for a bigger model, point at the note and the match page instead of improvising. For assisted procurement or custom machines see services and contact.
A budget box that meets a written goal beats an expensive box that disappoints quietly.
Appendix: sample allocation sketch
Reserve roughly: GPU and cooler as the primary line; PSU with headroom next; case and airflow; SSD for weights; leftover for cables and filters. Revisit the sketch after you price second-hand cards. If the quote drifts into fantasy 70B promises, return to match tools and rewrite the sticky note on the case. A written non-goal list protects the budget from friendly pressure.
Appendix B: procurement conversation starters
Ask vendors about return windows for GPUs, whether the PSU is new, and whether the case fits your desk depth. Bring the budget-5000 page as a conversation anchor rather than a shopping cart. If a salesperson pushes a card that breaks the VRAM plan, pause and reopen the VRAM table. Write the final BOM into the same repo as your runbooks so future you knows what was actually bought. Include a line for dust filters and a surge-protected strip so those are not forgotten at checkout.
Published by MicroAI One (本地AI 1号网), a Shenzhen team working on local LLM hardware, software and deployment.