Your idle GPU could power the next generation of AI.
An open marketplace connecting unused computing power with developers who need affordable AI inference.
Pre-launch. 0 live workers — the network is being built in the open. No earnings are promised; no billing is active.
What Nodefield is
Idle hardware, meet real demand
Millions of capable GPUs sit idle in gaming PCs, workstations and small servers while developers pay premium prices for AI inference. Nodefield is building an open marketplace that connects the two: GPU owners contribute capacity they are not using, developers call open-weight models through a simple OpenAI-compatible API, and the scheduler routes each request to a single suitable machine.
For GPU owners
Earn from your unused GPU — when your hardware actually completes paid work. You stay in control: pause any time, set resource limits, uninstall any time. No earnings are guaranteed.
For developers
Run AI models at lower costs through one API. Open-weight models, streaming, token usage reporting, and a base URL you can swap in place of your current provider.
For businesses
Access flexible AI infrastructure without owning GPUs. No hardware to buy, no capacity to manage — pay for the inference you use, once paid billing exists.
How it works
Three steps, no hype
- Providers install the Nodefield agent. It detects your GPU and VRAM, runs a local benchmark, and reports available capacity to the coordinator. It never runs arbitrary customer code and never touches your personal files.
- Developers send a normal API request. One endpoint, OpenAI-compatible. The scheduler assigns the whole request to one suitable machine — we do not split a single model across geographically distributed consumer GPUs.
- Work completes, usage is reported. Tokens in and out are counted, results return to the developer, and — once paid billing exists — the provider is paid for completed work under the published split.
Honest status
Where we actually are
These are real counts from an empty pre-launch system, not placeholders waiting to be inflated. Live values, when the backend runs, are on the status page.
Launch-candidate models
Small, open-weight, runnable on consumer GPUs
We prioritise smaller quantised models that genuinely fit in consumer VRAM. All five launch candidates are planned — not yet serving live traffic.
Qwen3-4B
Apache 2.0 · ~2.5 GB (Q4) · ~4 GB VRAM
Qwen3-8B
Apache 2.0 · ~5.2 GB (Q4) · 6–8 GB VRAM
Llama-3.1-8B
Llama 3 Community License · ~5 GB (Q4)