Rent out your idle AI compute
DarkGPT is a peer-to-peer AI network: your Mac answers other people's prompts while you're not using it and earns DCN for every token it generates. Prompts arrive encrypted to a key that never leaves your Secure Enclave, and DarkGPT's servers never see them.
- Mac Studio · M1 / M2 / M3 Ultra — up to 512GB, runs the 671b flagship
- MacBook Pro, Mac Studio · M1–M4 Max — 64–128GB, 70b class
- MacBook Pro, Mac mini · M1–M4 Pro — 32–64GB, 32b class
- MacBook Air, Mac mini, iMac · M1–M4 — 16–32GB, 7b–8b class
- H100 and H200 — Hopper, SXM and PCIe
- B200 and GB200 — Blackwell, TEE-I/O
- RTX PRO 6000 Blackwell Server Edition
- Linux, in a confidential VM — AMD SEV-SNP or Intel TDX host, NVIDIA remote attestation
What happens to a prompt here
Four steps, no middleman, and nothing that survives the request except a proof that some work was done.
Encrypted to one provider's public key before it ever touches the network.
Discovery runs over a DHT with hashed topics. No broker, no directory of users.
Decrypted only inside the enclave. The operator cannot read it, by construction.
With a zk proof you verify in 200ms. Nothing is logged on either end.
Models on the network
FREE DURING BETA / PER-TOKEN PRICING IN PHASE 2Private by architecture, not by policy
A privacy policy is a promise. This is a network with nowhere to keep your data, run by people who could not read it if they tried.
Every answer runs on somebody's machine. Get paid to be that machine.
Providers earn DCN for every token they generate. Earnings accrue to your ledger from the first request; on-chain payouts switch on with payments in phase 2. You never see a prompt, which is exactly the point.
What is live, and what is next
The beta is free on both sides of the network. Payments switch on in the next phase, and nothing about the privacy model changes when they do. Installed apps update themselves over the swarm. Read the release log.
Node identity keys created inside the Apple Secure Enclave with verifiable attestation. Prompts encrypted to one provider's key end to end. App updates delivered peer to peer over Hyperdrive with checksum verification. DCN reward ledger running on testnet.
Desktop app for Apple Silicon, signed and notarised, one download, then updates itself over the swarm. Chat on the network, or switch to provider mode and serve models from your own machine. Inference is free for clients during the beta, and providers' earnings accrue per token to their ledger.
DOWNLOAD FOR MACOS →Pay per token in DCN. The client locks payment in escrow before any work starts; it releases when the proof verifies and refunds if it doesn't. Providers get paid out to their own wallet per request, no thresholds and no payout schedule.
Windows and Linux clients. NVIDIA confidential GPU providers, H100, H200 and B200, running in attested confidential VMs so the same rule holds everywhere: every provider sits inside a hardware trust boundary.
GET NOTIFIED AT EACH PHASE →Questions people actually ask
Is DarkGPT free?
Yes. The macOS app is free and inference is free during the beta. The chat box on this page gives five free messages a day without an account; the app has no daily limit.
What do I need to run a node?
Any Apple Silicon Mac (M1 or later) on macOS 14 or newer. The app installs Ollama for you and recommends a model that fits your memory. NVIDIA confidential-computing GPUs (H100, H200, B200, GB200, RTX PRO 6000 Server) are next. GeForce cards are not supported because they have no trusted execution environment.
How much can I earn with an idle Mac?
Rewards accrue at 0.1 DCN per 1,000 tokens your machine generates, on a test network during the beta. The app measures your tokens per second and shows a daily ceiling; what you actually earn depends on demand. Test-network DCN has no cash value yet, so treat it as a running tally, not income.
Where do my prompts go?
Your prompt is encrypted on your machine to the public key of the provider that will serve it, and the answer comes back encrypted to you. DarkGPT's servers never see either side. The provider's key lives in its Secure Enclave and its system posture is attested before it can serve.
Is it safe to run a node?
Your machine runs a language model on text; it never executes other people's code. Prompts are processed by Ollama locally, provider keys stay in the Secure Enclave, and you can stop serving at any moment by switching modes or quitting the app. The code is MIT-licensed and public.
Which models can I use?
Whatever providers are serving right now, which the app shows live. Today that is mostly 7B-class open-weight models such as WizardLM 2 and Qwen 3, chosen because a flat per-token rate makes them the best fit for most Macs.
Does it work on Windows or Linux?
Not yet. The app is macOS on Apple Silicon only. NVIDIA confidential-GPU nodes on Linux are the next platform; Windows is after that.