Hardware for private LLMs in Australia
Buy for the model class you will actually run — not a thumbnail. Prefer storage and RAM upgrades before a GPU when you are still learning. Always compare live AUD street pricing.
Affiliate disclosure (this page)
Amazon Australia search links use our Associates tag. We may earn a commission on qualifying purchases at no extra cost to you. Other retailer links may be ordinary deep-links until we enrol in their programs. Blocks below are product classes, not lab-tested “best SKU” awards. Full disclosure.
What might fit this box?
A conservative filter for local weights only. Not a benchmark, not a warranty, and not a shopping list. Ollama
Cloud and :cloud tags are out of scope.
Hardware tiers (orientation)
Match rough budget to model ambition. Specs are practical learning bands, not lab certificates.
| Tier | Budget (AUD hint) | Model class | GPU / memory | System RAM | Storage | Best for |
|---|---|---|---|---|---|---|
| Entry / hobbyist | Often $0 extra (use what you have) to low four figures for upgrades — verify live | Small open models first (common “few-billion to low-teen-billion parameter” class when quantised) | iGPU / CPU / Apple Silicon with moderate unified memory / entry discrete GPU | 16 GB often workable · 32 GB more comfortable for learning | Tens of GB free at minimum · prefer NVMe as the library grows | Learning a runner, chat, light coding experiments |
| Prosumer | Often mid hundreds to several thousand AUD depending on generation and stock — verify live | Mid-size open models when quantised; very large models may still need offload or more VRAM | Mid-to-high consumer VRAM (16–24 GB class common; 32 GB flagship if you can actually buy one). RTX 40- and 50-series both appear in AU shops — generation is less important than VRAM, PSU, and street price. | 32–64 GB commonly discussed for comfortable local work | 1 TB+ NVMe is a common comfort zone for multiple models | Daily local driver, private coding, multi-doc experiments |
| Business / workstation | Workstation/server quotes vary widely — get current supplier pricing | Larger models, multi-user serving, longer context — still not automatic frontier parity | High VRAM / multi-GPU / high unified Mac Studio-class / workstation cards | 64 GB+ often discussed for heavier multi-user or embedding workloads | Fast local or NAS storage for knowledge bases | Team RAG pilots, offline-oriented trials, predictable internal tooling (with real ops) |
Australian realism filters
What overseas “buy this GPU” posts often skip.
Electricity & heat
A high-end GPU can draw hundreds of watts under load. In Queensland and humid NSW summers that becomes heat you pay for twice — power and cooling. Prefer efficient quant models and idle-friendly setups for always-on boxes.
Stock & AUD street pricing
GPU stock and pricing move weekly. RTX 50-series (Blackwell) is on Australian shelves alongside leftover 40-series. Mid-range 16 GB cards (e.g. 5080-class) are often the practical “buy if in stock” band; 32 GB flagships have been scarce or scalped at times — do not assume MSRP. Check Amazon AU, Scorptec, Mwave, PLE, Centre Com, and Umart. “Check current price” always beats a stale screenshot.
NBN & download sizes
A 30–40 GB model on a constrained evening NBN plan is a night of downloading. Prefer starting with 4–8 GB models. Mirror overnight; keep a checksum habit.
Apple Silicon
Unified memory is a genuine advantage for mid-size models without a discrete GPU. Mac Studio / high-RAM Macs can feel simpler than PC GPU hunting — at a different price curve. On a Mac, MLX is a native option beside Ollama. Apple Intelligence / on-device Foundation Models are a different product (vendor stack), not “download Qwen.”
Quantisation
GGUF Q4/Q5 is the practical default for quality-vs-size. Q8 is heavier. “Runs in VRAM” beats “technically loads with 40 GB of RAM thrash.”
Amazon AU search examples (affiliate)
Each button opens an Amazon Australia search for a hardware class. Results change. Read the listing.
Australian retailers to compare
Use multiple shops for stock and price. Affiliate status is labelled honestly.
Associates (tagged)
Amazon Australia
Primary Associates AU program (live tag). GPUs, RAM, NVMe, mini-PCs, UPS, cooling — stock varies.
Affiliate pending
Mwave
Strong AU components. Affiliate application pending approval (2026-08). Until tagged: deep-link only.
Network (when enrolled)
Scorptec
Strong GPU / component stock. Affiliate historically via FlexOffers (~4%) — apply via network.
Network (when enrolled)
PLE Computers
WA-based stock comparison. Affiliate historically via FlexOffers — verify active before counting on payouts.
Deep-link only
Centre Com
Retail + online street pricing. No clear public affiliate program — deep-link for comparison only.
Deep-link only
Umart
QLD-strong shopping. No clear public affiliate program — deep-link for comparison only.
Macs and compact workstations
Unified memory on Apple Silicon is still the straightforward Australian way to run mid-size models without a discrete GPU. Class language only — current shipping Macs, not an M-series shoot-out. On a Mac you can use Ollama or a native MLX path. Compact personal-AI boxes (NVIDIA DGX Spark-class and similar desktop appliances were marketed through 2026) are a business-tier option: verify live Australian availability and price before you treat any dollar figure as real. Learners should upgrade RAM and fast local storage before a hero GPU. Electricity and heat: running costs.
Power supply & cooling checklist
- PSU wattage with headroom for GPU peak + CPU + drives (use manufacturer guidance)
- Case airflow; high-power cards need real ventilation
- UPS for long downloads if your power is unreliable
- Room heat plan for QLD/NSW summer if always-on