Home lab
Home AI Lab
One fast GPU in a quiet tower. For an individual running models locally - no rack, no server-room noise.
pricenoisepower draw
updated 16 Sept 2026Total price
$3,838
VRAM
32 GB
RAM
64 GB
Storage
2 TB
Est. power draw
919 W
AI capability
Roughly ~13B params in fp16, ~26B in 8-bit, or ~49B in 4-bit quantization.
small-model inferencemid-size inferencelocal RAG
Components
- GPUNVIDIA GeForce RTX 5090memory: 32 GB GDDR7 · bandwidth: 1792 GB/s · interface: PCIe 5.0 x16$1,999Check price ↗
- CPUAMD Ryzen 9 9950Xcores: 16 · threads: 32 · clock: 5.7 GHz boost$599Check price ↗
- MotherboardASUS ProArt X670E-Creator (AM5)chipset: X670E · memory: 4 x DDR5, up to 192 GB · pcie: 2 x PCIe 5.0 x16 (x8/x8)$300Check price ↗
- RAMG.Skill 64 GB (2x32) DDR5-6000 CL30capacity: 64 GB (2 x 32) · type: DDR5-6000 · ecc: non-ECC$175Check price ↗
- StorageSamsung 990 PRO 2 TB NVMecapacity: 2 TB · bus: PCIe 4.0 NVMe · read: 7450 MB/s$130Check price ↗
- PSUbe quiet! Dark Power 13 1200 W 80+ Platinumcapacity: 1200 W · rating: 80+ Platinum · standard: ATX 3.1 / PCIe 5.1$230Check price ↗
- CoolingNoctua NH-D15 G2 (air, dual-tower)type: air dual-tower · dissipation: up to 250 W · noise: 24 dBA$150Check price ↗
- NetworkingIntel I226-V 2.5GbE (onboard-class NIC)speed: 2.5 GbE · ports: 1 · bus: PCIe 3.0 x1
- ChassisFractal Define 7 XL (ATX tower)form factor: E-ATX tower · gpu clearance: 2 triple-slot GPUs realistic · drive bays: up to 18 (with kits)$220Check price ↗
Component prices are hand-sourced and dated (checked 2026-09-16), not scraped. Confirm on the vendor's site before buying.
Why this build
- · A single RTX 5090 (32 GB) runs 70B models in 4-bit and leaves headroom for 13-34B in fp16 - the ceiling of what one consumer card does well.
- · Ryzen 9 on AM5 is the widest consumer platform: enough PCIe for one GPU, up to 192 GB of DDR5, and a cheap upgrade path.
- · A sound-dampened tower with a dual-tower air cooler keeps it desk- or living-room-quiet. No liquid loop to leak, no rack to house.
- · 1200 W Platinum leaves ~25 % headroom over the measured peak - the margin that keeps a PSU quiet and long-lived.
Pros
- + Silent enough for a shared room
- + Lowest entry price for 32 GB of VRAM
- + Standard parts, trivial to service
Limits
- − One GPU: no tensor-parallel path to 70B in fp16
- − Non-ECC memory - fine for inference, not for long training runs
- − 2.5 GbE only - not built to serve many concurrent users
Alternatives
Cheaper
RTX 4090 (24 GB) saves ~$200 and 125 W for a small VRAM cut.
Balanced
128 GB RAM + 4 TB NVMe: room for a model library and local RAG.
Buy vs rent vs hybrid
Over 36 months, at this build's power draw and the live cloud rate for its GPU.
Own it
- Hardware
- $3,838
- Electricity
- $1,992
- Total cost of ownership
- $5,830
Rent it (cloud)
- Equivalent cloud, monthly
- $397/mo
- Equivalent cloud, total
- $14,309
- Break-even
- 11 months
- · Break-even is 11 months, comfortably inside the horizon at this utilization - buying wins.
- · Over 36 months, owning saves roughly $8479 versus the cloud equivalent.
"Buy on Amazon" links are affiliate links - as an Amazon Associate, Obolith earns from qualifying purchases.