Breaking

PC Built for $330 With Used V100 Runs gpt-oss-20b Locally

June 29, 2026 at 22:36 EDT

Infra & Chips
Open Source
Foundation Models

A method of sourcing retired data-center NVIDIA Tesla V100 GPUs on eBay for roughly $200–$300 and assembling a PC for about $330 to run OpenAI's open-weight gpt-oss-20b model locally is drawing attention among DIY builders. Against the trend of considering a $4,200-plus Mac Studio for local LLMs, it is spreading as a configuration that exploits the price edge of used GPUs.

Local AI on a budget

A $330 retired data-center GPU now runs OpenAI's GPT-OSS 20B at home

Once a $10,000 DGX-server card, NVIDIA's 2017 Tesla V100 is being dumped onto the used market for $200–$300 — and modded builds are beating far newer consumer GPUs at local LLM inference.

~$330

Total modded-V100 build cost

900 GB/s

HBM2 memory bandwidth

130 tok/s

gpt-oss-20b on Ollama

Cost to start local AI inference

Approx. build price · taller column = more expensive

~$330

Modded
Tesla V100
16/32GB HBM2

~$700

Used
RTX 3090
24GB GDDR6X

~$4,200

Mac Studio
high-memory
128GB-class unified

The V100 build lands at roughly one third the price of a used RTX 3090 — and a fraction of a high-memory Mac Studio.

Inference throughput — gpt-oss-20b

Tokens per second, one reported test (higher is better)

130

Modded V100

RX 7800 XT

Why it works

HBM2 bandwidth consumer cards lack
112–125 TFLOPS Tensor · 5120 CUDA cores
GPT-OSS 20B fits 16GB via MXFP4 (~o3-mini level)
SXM2→PCIe adapter costs only ~$50

The caveats

180–250W draw, loud blower fans
Needs Above 4G Decoding enabled in BIOS
CUDA 11 advised, driver quirks reported
SXM2 needs adapter + custom cooling

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.