Rent out your Mac's idle compute, or use everyone else's
DarkGPT is a peer-to-peer network where Apple Silicon Macs run open-weight AI models for each other. Clients get private answers; providers earn DCN for every token their idle machine generates. Here is the whole thing, end to end.
01What DarkGPT is
DarkGPT is a decentralized AI inference network. Instead of one company's data centre, the models run on ordinary Apple Silicon Macs owned by people who rent out the compute they aren't using. You can use it from either side: as a client who wants private answers from open models, or as a provider who wants an idle Mac to earn.
There is one app for both. It is free, open source under the MIT licence, and updates itself over the network.
02Using it as a client
Open the app, type a prompt, and it goes to a provider that is online and serving the model you picked. Automatic selection chooses the best available machine; manual selection lets you pick one and see exactly which models it serves.
Your prompt is encrypted on your Mac to that provider's public key. Only that machine can decrypt it, and its answer comes back encrypted to you. DarkGPT's own servers coordinate discovery and never see the text. Inference is free during the beta.
03Renting out your Mac
Switch the same app to provider mode. It installs Ollama if you don't have it, measures your memory and bandwidth, and recommends the model that pays best on your hardware. From then on your Mac serves other people's requests while you're not using it.
Each request is decrypted inside your machine, run through the local model, and the answer is encrypted back to the client. Nothing is stored. You never run anyone else's code, only text through a language model, and you can stop at any time by switching modes or quitting.
04What a node needs
Any Apple Silicon Mac, M1 or later, on macOS 14 or newer. More unified memory means bigger models: 8 GB runs a 3B to 7B model, 16 GB to 32 GB runs 7B to 14B comfortably, 64 GB and up can serve 30B-class models.
Apple Silicon is required, not just preferred. Provider keys are generated inside the Secure Enclave and the app attests the machine's security posture (System Integrity Protection, sealed system volume, secure boot) before it can serve. That hardware trust boundary is why GeForce cards aren't supported: they have no trusted execution environment. NVIDIA's confidential-computing GPUs (H100, H200, B200, GB200, RTX PRO 6000 Server) do, and they are the next platform.
05What it pays
Rewards accrue at a flat 0.1 DCN per 1,000 tokens your machine generates. The app measures your real tokens per second and shows a daily ceiling; actual earnings depend on how much demand reaches your node.
During the beta, rewards are on a DCN test network and have no cash value. Balances are a running tally, and the reward program can change. The download and run-a-node pages say the same thing, on purpose.
06How prompts stay private
Three layers. Encryption: every prompt is encrypted end to end to a specific provider's key, so the network and DarkGPT's servers only ever carry ciphertext. Hardware: that key never leaves the provider's Secure Enclave, and the machine's posture is attested with a nonce challenge before it serves. Policy: telemetry is a fixed whitelist of device and event fields with no prompt content, and providers agree in the terms not to log or retain anything.
What we can't promise: a provider who physically modifies their own machine could see the plaintext of prompts routed to it. So don't send secrets or regulated personal data through the network. The security page goes deeper.
07Get started
Download the app to chat or to run a node. The run-a-node page has the hardware table and earnings estimator, and the Discord is where providers compare notes.