Compute should be something a community owns, not something four companies rent back to it.
Potluck is a peer-to-peer AI compute network. Run inference on your own machines first. Tap your friends’, your team’s, or your homelab when you need more. Opt into the public network when you want to share spare cycles or borrow them.
Read the protocol →Or install the beta on your machines →
Bittensor must be public. exo must be local. Only Potluck lets users choose per request. That is not a feature. It is a tokenomics constraint our competitors cannot escape.
§2.5 · Network economics
A token will exist. The mesh ships first.
Potluck plans to launch a two-asset model. A utility credit, priced in compute hours contributed and consumed, settles the work of running the network. A separate governance token, allocated to members and contributors, decides how the protocol evolves. The two assets are kept apart on purpose: utility on one side, value capture on the other.
The mesh ships before the token because supply density determines whether credits have anything real to coordinate. Launching a token before the network has substance is how decentralized infrastructure projects become memes. We would rather ship the boring part first and earn the right to ship the rest.
Governance covers the parts that actually need member control: routing policy for the trusted pool, slashing rules for nodes that misbehave on the public side, and treasury allocation for shared infrastructure the cooperative funds together. The full design will be published before any token ships.
§2 · What’s running today
Six months of measured numbers.
The peer-to-peer network is the long-run product. The mesh underneath it has been running for six months and the numbers below are measured, not promised.
| Measurement | Value | Reference |
|---|---|---|
| Cross-machine memory latency, p50, over WireGuard | 20 ms | [a] |
| Memory mesh characterized to | 10,000 items | [b] |
| Memories held at flat p50 | 5,000+ | [c] |
| LongMemEval recall_any@10 | 0.98 | [d] |
| Verified platforms | Mac · Linux · Win | [e] |
These are the numbers we’d want a skeptic to check, so they’re the first thing on the table. The 0.98 LongMemEval recall is the hardest one to earn. That is the one we’d ask a peer to reproduce first.
See the full benchmarks →§3 · Why peer-to-peer
Four companies should not be the only place AI compute happens.
Hyperscaler capex is capped. Consumer GPU FLOPS per dollar just crossed the threshold for useful local inference. A generation of homelab operators is online. The window for peer-to-peer AI compute is open right now, and it will not stay open forever.
Anthropic, OpenAI, Google, and Microsoft run almost all of the AI inference happening on Earth right now. Routing every prompt through those four companies is a structural choice most users never made consciously.
Potluck is for the people, teams, and organizations that own hardware and want the option of running their AI on it: alone, with people they trust, or as part of a broader network they help govern.
Centralized inference is cheap and fast today; that’s true. It’s also a structural dependency. Potluck doesn’t replace it for everyone. It gives the people who want an alternative a credible one.
The numbers are on the table. The first hundred seats are limited.
What only Potluck does
- Cross-machine memory mesh — shipped. Bittensor, Akash, exo, and Together AI have none.
- No token required. Bittensor requires TAO. Akash requires AKT. Potluck’s trusted pool runs on no token at all.
- Trusted private mesh. Your machines plus invited ones — not a public stranger network, not LAN-only. Bittensor, Akash, and Together AI have no equivalent.
See full comparison
| Feature | Potluck | Bittensor | Akash | exo | Together AI |
|---|---|---|---|---|---|
| Cross-machine memory mesh shipped | ✓ | ✗ | ✗ | ✗ | ✗ |
| Token required to participate | No | Yes (TAO) | Yes (AKT) | No | N/A |
| Trusted pool (private mesh of your own + invited machines) | ✓ | ✗ | ✗ | ✓ (LAN only) | ✗ |
| Public pool (open contributor network) | ✓ (opt-in) | ✓ | ✓ | ✗ | ✗ |
| Local-only inference | ✓ | ✗ | ✗ | ✓ | ✗ |
| Centralized | No | No | No | No | Yes |
| Feature | Potluck | Bittensor | Akash | exo | Together AI |
|---|---|---|---|---|---|
| Cross-machine memory mesh shipped | ✓ | ✗ | ✗ | ✗ | ✗ |
| Token required to participate | No | Yes (TAO) | Yes (AKT) | No | N/A |
| Trusted pool (private mesh of your own + invited machines) | ✓ | ✗ | ✗ | ✓ (LAN only) | ✗ |
| Public pool (open contributor network) | ✓ (opt-in) | ✓ | ✓ | ✗ | ✗ |
| Local-only inference | ✓ | ✗ | ✗ | ✓ | ✗ |
| Centralized | No | No | No | No | Yes |
On tokens: the public-pool side of Potluck will use a cryptoeconomic incentive mechanism designed to be utility-anchored: compute hours contributed and consumed, not speculation. The trusted-pool side requires no token. We will publish the incentive design before we ship it. Read more in the manifesto →
§4 · Early access
Request a place at the table.
The first hundred users help shape the protocol. Tell us what you’d build with Potluck and what hardware you’d bring. We pick by fit, not first-come.
No spam · no investor updates · unsubscribe in one click
Sources & receipts