Decentralized AI inference: who actually serves models peer-to-peer
Short answer: as of November 2026, models have been trained across volunteers’ machines (Covenant-72B, Pluralis-8B), but no public network serves a large model across untrusted consumer GPUs with verified answers and private prompts. What exists is datacenter GPUs rented through a token, your own devices pooled together, and small models that fit on one GPU. Obscurify.ai’s P2P sharing is the last kind: a browser tab serves an open model and earns credit for any other model.
“Decentralized inference” gets used for three different things, and most confusion comes from mixing them up:
- A marketplace of datacenter GPUs. Anyone can list a server, you pay per token, often in a crypto token. The machine that answers you is still a server in a rack, and its operator can read your prompt.
- Your own devices as one cluster. Your laptop, desktop and phone split a model between them. Private, because every device is yours, but it isn’t a network you share with anyone.
- A real peer-to-peer swarm. Strangers’ consumer GPUs serve strangers’ requests, nobody needs permission to join, and the network checks the answers. This is the one people mean when they say the future of private AI is peer-to-peer, and it is the one that isn’t here yet.
Who is doing what (checked 2026-11-06)
| Project | What it does today | Serves inference across strangers’ GPUs? | Who can read your prompt |
|---|---|---|---|
| Covenant AI (formerly Templar) | Trained Covenant-72B across 70+ permissionless contributors. Chat runs on Basilica, its own GPU cloud. | Not yet: “serving open models on distributed hardware is the piece we are working on now” (August 2026) | The Basilica host |
| Pluralis | Trained Pluralis-8B across 330 nodes; model-parallel training on consumer GPUs. Research on inference where no node holds the full weights. | Research only | n/a |
| Prime Intellect | Centralised inference API; TOPLOC for checking that a provider ran the model it claims. Its peer-to-peer protocol repo was archived in November 2025. | No | Prime’s servers |
| Gensyn | IR3DE-AXL prototype: anyone can serve or use models, with no verification or incentives yet. | Prototype | Whoever serves you |
| Chutes (Bittensor subnet 64) | Live pay-as-you-go inference on permissionless datacenter miners, with TEE attestation. | Yes, but datacenter GPUs | Nobody, if you trust the TEE |
| Petals | Volunteer swarm that split big models across home GPUs. Last code change September 2024. | Inactive | Every peer in the chain |
| exo | Pools your own devices into one cluster. | Your devices only | Only you |
| Obscurify.ai P2P sharing | Browser tabs serve open models up to 9B on their own GPU via WebGPU; accounts earn credit for any model. | Yes, models that fit one GPU | The tab that serves you sees the text, never who sent it |
Sources: Covenant-72B paper (arXiv 2603.08163), Covenant’s inference roadmap post, Pluralis-8B report (arXiv 2607.13332), Gensyn’s IR3DE-AXL write-up, Chutes, Petals, exo.
The two problems nobody has solved yet
Your prompt is visible to whoever runs the model
A GPU that computes the answer has to see the input, or at least the activations of the layers it runs, and activations can be inverted back into text. Splitting a model across many peers spreads your prompt across many strangers instead of hiding it. Confidential computing (TEEs) fixes this on datacenter GPUs like the H100 and Blackwell, but consumer graphics cards and browsers have no attestation.
What can be separated is the prompt from the person. On Obscurify.ai, the tab that serves a P2P model gets the conversation text and nothing else: no account, no address, no time zone. That is the same line Obscurify.ai draws with the big model providers, who also read prompts but never learn who sent them.
Checking that a peer ran the real model
A peer can save electricity by returning made-up text. The known fixes are re-running a sample of the work (TOPLOC, Gensyn’s Verde), hardware attestation (TEEs), and spot checks with known answers. Zero-knowledge proofs are still far too slow for language models. P2P sharing uses spot checks with known answers plus a 24-hour hold on earnings, so cheating costs more than it pays.
What would change the picture
- Covenant AI ships distributed serving. It has already shown that strangers’ GPUs can train a 72B model together. Serving it the same way would be the first large open model available from a permissionless network.
- Pluralis serves a model no single node holds. That keeps the weights safe, which is what makes open, paid models possible on a public network.
- Verification gets cheap enough for consumer GPUs, so networks can pay peers without trusting them.
This page is updated when one of those happens. Obscurify.ai watches these projects’ code and announcements every two hours.
Common questions
Is there a peer-to-peer ChatGPT?
Not one that serves frontier-sized models from consumer hardware. The closest things today are open models up to about 9B parameters served from single consumer GPUs, and large open models served from datacenter GPUs rented through decentralized marketplaces.
Can I earn AI credits with my GPU?
Yes. With P2P sharing, a browser tab with WebGPU serves an open model to other people and your account earns 80% of what they pay, spendable on any model including GPT, Claude and Gemini. No install, no email, no wallet.
Is decentralized inference more private than a normal AI provider?
Not by itself. More machines means more parties who could read your prompt. What makes it private is separating the prompt from your identity: no account tied to you, payment that can’t be traced back, and a relay so the peer never sees your address.
What happened to Templar on Bittensor?
The team behind Templar, now Covenant AI, left Bittensor in April 2026. See Templar AI and Covenant-72B.
Last updated 2026-11-06. Plan prices and model lists on this page are read live from the site.
Try it now: guest chat needs no account, no email and no cookies. When you want more models, an account is just a random number.
Lend your GPU from a browser tab See pricing