24 GB of VRAM for the price of a mid-range GPU — the smartest on-ramp to local AI I know. A used RTX 3090 runs 7B to 13B models at speed and a 32B comfortably, all on your own hardware, with no per-token bill and no prompt ever leaving the building. Pair it with 64 GB of RAM for offloading and a Gen4 NVMe so models load in seconds.
Plans
Choose a plan that fits your needs and budget
Item List
10Cooling & Peripherals
3 itemsPlatform & Storage
4 items| Item | Category | Specs | Qty | Price | Link |
|---|---|---|---|---|---|
 | NVMe Storage | Capacity2 TB Read7000 MB/s InterfacePCIe 4.0 | 1 | $130 | View Shop |
 | Power Supply | Wattage850 W Rating80+ Gold ModularYes | 1 | $130 | View Shop |
 | Motherboard | ChipsetB650 FormATX Slots4x DDR5 | 1 | $150 | View Shop |
 | PC Case | FormATX Fans included4 Front panelMesh | 1 | $80 | View Shop |
Core Compute
3 items
Plan Totals$1,940
Cooling & Peripherals
3Platform & Storage
4Core Compute
3Total$1,940
FAQ
Common questions about this kit
Can it really run a 70B model?
A 70B needs about 40 GB VRAM at 4-bit. One 3090 (24 GB) runs it with layers offloaded to RAM — slowly — or runs a 32B comfortably. For full 70B speed, add a second 3090.
Why a used 3090 over a new 4060?
For local LLMs, VRAM is everything. A used 3090 gives 24 GB for roughly the price of an 8 GB 4060. There is no contest for inference.
Is the liquid cooler for the GPU or CPU?
The CPU. The 3090 has its own blower cooler. Undervolt the 3090 and it drops 15 to 20 percent in both heat and power draw.
Why 64 GB of RAM?
Layers that spill off the GPU land in system RAM during inference. 64 GB keeps large models from thrashing and slowing to a crawl.
User Reviews
5.0 / 5.0
Sort:
VRAM is king, agreed. The undervolt tip earned the fifth star.
Undervolt the 3090? I do the same to my 4060 Ti. Budget VRAM gang.
Sat next to a guy on a flight who built this. Confirmed the 3090 is the value play.
I run hillshade and weather models on this. A used 3090 was the right move — it runs cool undervolted.
Budget VRAM