robotnurse.io

./gear/best-local-ai-setup-under-3000 — buyers' notes

Best local AI setup under $3,000 — three builds that work

Three builds that work: under $800 runs 7–8B models, under $2,000 runs 13B comfortably, and a full $3,000 buys a 24GB GPU that runs 30B-class models at q4. Spend on VRAM first; everything else is secondary.

Three price tiers — under $800, under $2,000, and the full $3,000 — with the parts list, what each runs comfortably, and what to skip.

✎

Street prices move weekly — treat every number here as ±15%. What doesn't move is the logic: memory first, compute second, everything else last.

Already own a gaming or desktop PC? Before you buy any of this, read turn an old gaming PC into a local AI box — you may only need a used GPU, not a whole new build.

[ tier 1 — under $800: the quiet starter ]

the box Mac mini M4, 32GB — 7B models at q8 and 13B at q4 run comfortably, in silence, on 30W
what it won't do 30B-class models, or more than one model resident at once
don't buy the 16GB config to save $200 — it's the one upgrade you can't make later

This is the right first machine for most people. It answers "is local AI useful to me?" for the price of a mid-range phone, and if the answer is yes it becomes your always-on small-model server when you upgrade.

[ tier 2 — under $2,000: two honest paths ]

path A — GPU PC RTX 5070 Ti 16GB + mid-range CPU, B650 board, 64GB DDR5, 2TB NVMe, 850W PSU — roughly $1,600–1,900 built. Fastest 7B–13B tokens/sec in this whole guide.
path B — bigger Mac Mac mini M4 Pro, 64GB — slower decode than the PC, but runs 30B q4 without breaking a sweat and stays silent doing it

Choose on workload, not spec sheets: mostly chat and coding assistants at 13B and below → build the PC. Want 30B-class models or an always-on home server → buy the Mac. The fork is explained properly in Mac Studio vs GPU PC.

[ tier 3 — the full $3,000 ]

the pick Mac Studio, M4 Max, 64GB — 30B at q4 with long context, headroom for two models resident, silent at full load
stretch option used or sale-priced 96GB configs — worth stalking; that memory tier is the difference between "30B fits" and "30B is comfortable"
why not a bigger GPU PC at $3,000 a PC still tops out at one high-VRAM card; crossing into real 70B territory means multi-GPU complexity this budget can't carry honestly

At this budget the constraint stops being compute and becomes addressable fast memory — which is exactly Apple's home turf. Config-by-config detail lives in the Studio config notes.

[ the part everyone underbuys ]

Every tier above fails or succeeds on memory, not on the chip. Before you pick a tier, do the math for the model you actually want to run: RAM vs VRAM — the full memory guide covers weights, context costs, and offload penalties; the Apple-specific budgets are in how much RAM you actually need. Buy for one model class above today's favorite — the next one is always bigger.

CHECK PRICES see options at the retailer →

AFFILIATE DISCLOSURE: links in this post are affiliate links — if you buy through them I earn a small commission at no extra cost to you. RobotNurse is independent; I only recommend hardware I would actually deploy. No vendor paid for placement.

WANT IT SET UP FOR YOU?on-site /work

Picking the hardware is the easy half. I'll spec, install, and tune it end to end on your hardware.

start with the readiness audit →   or book the call