By Gavin Stevens · Updated November 2026

The entry-level local AI PC: a complete parts list

Every part of a PC that runs 14B models at full speed and 30B models at a squeeze, chosen for the lowest price of entry in a year when memory and graphics cards keep getting more expensive, with the upgrade path laid out so nothing you buy today gets thrown away.

Illustration of a black mid-tower PC with a dual-fan graphics card inside, with spare memory sticks beside it

As an Amazon Associate I earn from qualifying purchases. Product links on this page go to Amazon; they are affiliate links, and the price you pay is the same.

The picks

Under $1,000
Illustration of a compact dual-fan desktop graphics card Under $1,000

NVIDIA GeForce RTX 5060 Ti (16GB)

The whole build exists to feed this card. 16GB, CUDA, and the lowest price of entry to the 14B class: about $765 in late October 2026, the only new 16GB card left under $1,000. Get the 16GB version.

For local LLMs, holding the model matters more than speed. This is the lowest price of entry to 16GB with CUDA. Get the 16GB version; the 8GB one is useless for this.

View on Amazon

Illustration of a square desktop processor with a silver heat spreader CPU

AMD Ryzen 5 7600 (6-core, AM5)

Six cores is plenty; the GPU does the work. AM5 means the next CPU drops in without a new board.

The cheapest AM5 chip worth buying. AM5 means a drop-in upgrade to a faster CPU later without changing the board.

View on Amazon

Illustration of a black ATX motherboard with an empty CPU socket, memory slots and a reinforced graphics card slot Board

B650 ATX motherboard (two M.2 slots, PCIe 4.0 x16)

One x16 slot, two M.2 slots, four DDR5 slots. That's the whole spec. Any reputable B650 ATX board with those three things is fine.

A B650 board does everything an AI PC needs. The extra money for X670 or X870 buys USB4 and PCIe 5.0, neither of which changes tokens per second.

View on Amazon

Illustration of a two-stick desktop DDR5 memory kit with black heat spreaders 32GB

32GB DDR5 kit (2 x 16GB, 6000MT/s)

Enough for the OS and a model that fits in VRAM. Replace with 64GB on two sticks when you start offloading; don't add two more sticks.

Keeps the entry build under the line in a year when memory is expensive. First upgrade when you outgrow it: replace, don't add, so you stay on two sticks.

View on Amazon

Illustration of an M.2 NVMe solid-state drive 2TB

Samsung 990 PRO 2TB NVMe SSD (PCIe 4.0)

Models are 5 to 20GB each at this tier. 2TB means you stop thinking about it for a year.

Load time is disk-bound. A fast NVMe turns a 30-second model load into a 5-second one.

View on Amazon

Illustration of a black modular ATX power supply 850W

Corsair RM850x 850W 80+ Gold ATX 3.1 power supply

More than the 5060 Ti needs, and exactly what a 5070 Ti or a used 3090 needs. Buy it once.

Most single-card builds never pull 600W at the wall. Paying for 1000W you'll never use buys nothing but a bigger fan.

View on Amazon

Illustration of an empty black mid-tower PC case with a mesh front and a glass side panel Case

Mid-tower case with mesh front (fits a 3-slot, 330mm GPU)

Mesh front, room for a 330mm card. The only part you'll still own in five years.

The case is the one part you keep through every upgrade. Buy for the 5090-sized card you might own in three years, not the 5060 Ti you're buying today.

View on Amazon

What this build does

Runs any model up to 14B at full quality at 60 tokens a second or more, runs the 20B class (gpt-oss-20B) comfortably, and runs 30B models at a tighter quantization. It will not run 70B models. It will run SDXL and Flux at FP8 for image generation. For most people starting out, that’s a year of use before the limits bite.

Where it lands on price

When this page was written in October it sat at the bottom of the Under $1,000 tier. It doesn’t now. The 5060 Ti 16GB was about $765 at the cheapest on October 26, with board partners reporting another $37 to $60 for November, and PC DRAM contract prices set a record in October. As of November 2026 the list as written lands in the $1,000 to $3,000 tier, closer to $1,500 than $1,000, and the card alone is most of the gap. The 32GB kit and a 1TB drive instead of 2TB pull a couple of hundred back, but not enough to cross the line. If the line matters more than the card, the used 3060 route at the bottom of this page is the way under it; if the card matters, buy it before the November prices land.

The parts, and why each one

Graphics card: RTX 5060 Ti 16GB. The entire build is justified by this card. It’s the cheapest new card with 16GB and CUDA, which means every tool works and the 14B class fits with room for context. The 8GB version of the same card exists and is useless for this; check the listing. If the budget stretches, the 5070 Ti is the same 16GB at about twice the speed and is a direct swap; it was about $1,229 in late October, so the stretch is now the price of the rest of the build.

CPU: Ryzen 5 7600. For inference the CPU moves data and runs the parts of a model that don’t fit on the GPU. Six fast cores do that. It’s on AM5, so the upgrade later is a chip, not a platform. Don’t buy a faster CPU for this build; the money belongs in the card.

Motherboard: any B650 ATX with two M.2 slots. The three things that matter: a full-length PCIe x16 slot for the card, two M.2 slots so you can add a drive later without pulling one, and four DDR5 slots. Every B650 board has the first and most have the other two. X670 and X870 add USB4 and PCIe 5.0, neither of which changes tokens per second.

Memory: 32GB DDR5 on two sticks. The minimum that leaves the OS comfortable while a 14B model sits in VRAM. It’s deliberately the first upgrade: when you want to run models bigger than the card, replace it with a 64GB two-stick kit. Two sticks hold their rated speed; four often don’t, so replace rather than add.

Storage: 2TB NVMe, PCIe 4.0. Load time is disk-bound and models pile up. 2TB at PCIe 4.0 is the right amount; PCIe 5.0 costs more and runs hotter for no gain here. If the budget’s tight, 1TB now and the 2TB as the second drive later, since the board has the slot.

Power supply: 850W ATX 3.1. The 5060 Ti would run on 550W. The 850 is here because it’s also exactly what the 5070 Ti or a used 3090 needs, and a power supply is the part you least want to buy twice. Native 12V-2x6 connector, no adapters.

Case: mid-tower with a mesh front. Buy the case for the card you’ll own in three years, not the one you’re buying today: 330mm length and three slots of clearance. Mesh front, two intake fans, one exhaust. Glass fronts look better and run hotter under the sustained load that inference and image generation produce.

The upgrade path, in order

  1. 64GB memory (replace the 32GB kit). Lets you offload models bigger than the card. First thing most people do.
  2. A 24GB card: used RTX 3090 or 4090. The 850W supply and the case already fit either. This is the jump to 30B models at full quality, and it’s the one that changes what the machine is for.
  3. A second 2TB drive in the spare M.2 slot.
  4. A faster CPU, only if you offload a lot. Same board.

Nothing on this list makes you throw away a part you bought today except the 32GB kit, and that’s the cheapest part here.

What this build is not

It’s not the cheapest possible PC that runs a model; a used office PC with a used 12GB RTX 3060 does that for a lot less and is a fine way to start. It’s the cheapest build where every part is one you’d still want in the machine after two upgrades. If you want the absolute floor, buy the 3060 route and skip this page. If you want to buy once, this is the list.