Custom PC for Local LLMs (2026): Components Guide & 9 Picks by Budget

Building your own machine is how you get the most VRAM per dollar — and VRAM is the spec that decides which models you can run. This guide covers every component that actually matters for local LLMs, then gives you three picks per budget tier.

Rather unbox and go? See the companion guide: Pre-Built Machines for Local LLMs.

Evergreen note: component prices move weekly. I re-verify the picks below on a regular schedule — check the last-updated date, and treat street prices as “verify before buying.”

The components that matter

VRAM: the one number that decides everything

Local LLMs live or die on how much model fits in fast memory. Rough mapping at usable quantization (Q4–Q5):

• 12–16GB → 7B models comfortably, 13B tight
• 24GB → 13B comfortably, 30B with quantization
• 32GB → 30B comfortably, dipping into 70B
• 48GB+ → 70B territory

Quantization (Q4/Q5/Q8) shrinks models with modest quality loss — it’s why a 24GB card punches above its weight. Raw speed (tokens/sec) matters far less than whether the model loads at all. Buy memory first.

GPU: NVIDIA, realistically

CUDA remains the ecosystem everything targets — every framework, every optimization lands on NVIDIA first. AMD’s 24GB cards offer more VRAM per dollar, but ROCm support (mostly Linux, mostly tinkering) is the catch. For a build you want to use rather than debug, it’s NVIDIA.

CPU: matters less than people think

For inference, the GPU does the heavy lifting. A mid-range modern CPU (Ryzen 5 / Core i5 class or better) is plenty — don’t spend GPU money here. What matters more: enough PCIe lanes if you ever want a second GPU.

System RAM: the overflow lot

When a model doesn’t fit in VRAM, frameworks offload layers to system RAM. It’s slower, but it lets a 24GB card attempt 30B+ models. 64GB of DDR5 is the sweet spot for a serious local-AI build; 32GB is the minimum worth bothering with.

Motherboard: PCIe slots and spacing

Get a full ATX board with two PCIe x16 slots (electrically x8/x8 or better) if multi-GPU is even a remote possibility — high-end cards are triple-slot monsters and need physical spacing. This is the component people cheap out on and regret.

PSU: do the wattage math

High-VRAM cards are power-hungry: an RTX 3090 can pull ~350W, a 4090 ~450W, a 5090 ~575W. Size your PSU at 1.5× your estimated max draw, buy 80+ Gold or better from a reputable brand, and don’t reuse a decade-old unit. This is the one component that can kill the rest.

Cooling & case: they’re big and hot

Triple-fan, 300mm+ cards need a case with real airflow and GPU length clearance — check the spec sheet, not the product photo. Undervolting is free performance-per-watt; most of these cards ship overvolted.

Storage: fast NVMe

Models are 10–40GB each and you’ll collect them. A 2TB NVMe SSD keeps loading times sane and gives your library room to grow. This is cheap; don’t starve it.

Low budget (~$1,400–$2,500 total): what to expect

7B models comfortably, 13B with quantization. The entry ticket to real local AI.

Rest of the build at this tier: a Ryzen 5 / i5-class CPU, 32GB DDR5, a decent B-series board, 750W Gold PSU, and any case with airflow. Budget ~$500–$600 for everything that isn’t the GPU.

Mid budget (~$2,700–$3,600 total): what to expect

13B comfortably, 30B with quantization, and your first taste of multi-GPU. The enthusiast sweet spot.

Rest of the build: step up to 64GB DDR5 (you’ll want the offload headroom), a quality ATX board with dual x16 slots, 850–1000W Gold PSU depending on the GPU choice.

High budget ($5,000+ total): what to expect

30B comfortably, 70B within reach. Models that rival hosted frontier models for many tasks, running privately.

Rest of the build: don’t cheap out anywhere — 64GB+ DDR5, a workstation-class board, 1200W+ Platinum PSU for dual-GPU, a full-tower case with serious airflow, 2TB+ NVMe.

Compare component prices on Newegg Canada — GPU pricing moves weekly and sales are frequent.

What I’d skip

• 8GB GPUs for LLMs. Fine for gaming, a dead end for meaningful local models.
• Chasing launch-day prices. Last generation’s flagship, bought used six months later, is the perennial smart buy.
• Overspending on the CPU. A $600 CPU doesn’t run models better than a $200 one — put the difference into VRAM.
• Cheap PSUs. The one component that can take the whole build with it.

The honest footnote

I don’t earn anything from hardware links yet — these programs are on the list for a future review round. The picks above don’t depend on that either way.