A marketplace for AI inference where you don't have to trust the machine that served you. Anyone with a GPU can answer your prompt, and the network re-checks the answer itself before any money changes hands.
Every other GPU network leaves one gap open: the moment you send your prompt to a stranger's machine, you have no way to know it ran the model it promised, or a cheaper one wearing its name. OGONG closes that gap. Each answer ships with a small mathematical receipt of how it was produced, and because re-checking costs a few percent of generating, an independent peer re-runs nearly all of them. Money only moves when the receipt checks out.
You send a prompt to an OpenAI-compatible endpoint. From your side it's a normal POST /v1/chat/completions. Your payment is locked in escrow on-chain, not paid out yet.
The router forwards your request to a provider but removes your identity at the door. The machine that runs your prompt earns the fee without ever learning who sent it.
A provider runs the model on its own GPU and returns your answer. Alongside it, it commits a tamper-evident receipt: a Merkle root over the hidden states and token probabilities of exactly what it produced, signed with a post-quantum key and anchored with a validator even if the provider goes offline.
Because re-checking costs only a few percent of generating, the network can check nearly every answer, not just a sample. Two independent peers already running the same model each re-prefill the work and score it Accept or Reject, and the answer passes only if both agree, so a single colluding checker can't wave a bad reply through. The draw comes from an unpredictable randomness beacon, so a provider can't tell which answers are watched, and a cheaper model wearing the right name fails by an order of magnitude.
On a validator quorum (weighted by staked OGONG, so agreement means more than two-thirds of the stake, not a headcount) the escrow is released to the provider; every co-signing validator independently confirms the release from its own copy of the receipt before any money moves, so a co-signature is a real check, not a rubber stamp. A reply caught cheating forfeits its fee. Because nearly every answer can be re-checked this cheaply, honesty is the profitable move with no slashable bond posted, and the same scheme settles a model sharded across many machines, paying each for the layers it served. A protocol floor price keeps the fee-at-risk, and with it the cost of getting caught, from being competed to zero, and a consumer who wants certainty can force verification of its own request rather than trust the sampled rate.
OGONG is an open protocol and a set of CLI binaries you run yourself, headless, with nothing hosted in between. Four roles cover the whole loop above.
The permissionless marketplace. Consumers pay, providers serve and commit their work, validators audit a sample, and the chain settles escrow on a quorum. Verifiable-only supply. No privileged key on the money path: settlement, minting, and slashing are authorized by a co-sign quorum of staked validators alone, so no key can release, mint, or confiscate by itself, and none can veto or censor one either. Discovery is trust-minimized too: providers and routers register on-chain (endpoint, stake, and for a provider a hash commitment to its advertised models and prices), so a consumer verifies an advertisement against the chain and trusts no router. Anyone can enumerate the whole bonded set the same way and check each live advertisement against its on-chain commitment, so the network's state (which models are on offer, at what price, how much stake secures it, how much has been emitted) is a verifiable read of the chain, not a report you have to trust.
Serve a model from your GPU, commit a checkable record of what you produced, and earn. Fully headless. Composes modalities: text, image and audio at once, each its own subprocess. Set your own price per function (per token, per second of audio, per image) and override it per model.
A drop-in, OpenAI-compatible endpoint. Point your tools at it; identity is stripped at the router so the provider never learns who you are. Or run ogong-connect, a small local daemon that reaches the network for you: it finds a gateway automatically (bootstrap seeds, then it learns more over gossip and remembers them), verifies the provider's attestation, seals your request end to end, and opens the reply. Point any OpenAI or Ollama client at localhost.
A zero-signup, no-account server you run for yourself or a friend over an encrypted tunnel. Your own GPU, no third party to attest or audit.
# every role runs from the command line
ogong-validatord # validator / audit node
ogong-verifierd # re-runs committed work to score an audit
ogong-routerd # provider marketplace match engine
ogong-gatewayd # OpenAI-compatible consumer endpoint
ogong-connect # local client: verify + seal + reach the network from localhost
ogong-provider # provider daemon, also a local model server
ogong-segment-server # serves one layer-range shard in a cohort
ogong-lead # drives a cohort of shards across machinesMost networks decentralize who runs the GPU, then ask you to trust the answer. OGONG checks the answer itself, and settles only when it passes.
| OGONG | Bittensor | Venice | Dolphin | |
|---|---|---|---|---|
| Per-answer correctness | Nearly every answer re-checked (full coverage), tens of times cheaper (about 30-40x) than generating | Subjective validator consensus; no per-answer proof | Attests which code ran, not that the answer is right | Sampled per-answer logprob checks; probabilistic, not full coverage |
| Verification signal | Hidden-state sketch and logprobs, Merkle-committed, post-quantum signed | Subjective stake-weighted scoring | TEE attestation of code identity (Pro) | Sampled logprobs, model checksum, canaries |
| Modalities verified | Text, image, audio and video, all verified | Multi-modal across subnets; none proven per answer | Text, image, audio, video; correctness not verified | Text LLMs; verification is text-only, image and audio on the roadmap |
| Beyond a single GPU | Verified split inference: one model sharded across a cohort, every slice checked, zero bond | One model per miner; no verified sharding | Centralized; whole model per provider | Pooled nodes, whole model each; requests split, not the model |
| Provider capital | Zero correctness bond; honesty held by full-coverage audit | Recycled-TAO registration; penalties by lost emissions, not slashing | None (centralized provider) | Bonded POD with slashing (100k POD to validate) |
| Privacy | Optional TEE tier, hardware-attested, first-class | Opt-in on some subnets, validator-side | No-log policy by default; TEE is a Pro add-on | Software no-logs; no hardware attestation |
| Decentralization | Permissionless protocol; the verification itself is decentralized | Decentralized subnets | Centralized app, not a decentralized protocol; GPUs rented from third parties | Consumer-GPU DePIN (pooled) |
| Supply | Fixed 5B cap; 80% earned by verified work | 21M cap; distribution mechanism revised over time | Annual emissions plus buy-and-burn | Revenue buy-back |
Comparison reflects each project's public documentation as of June 2026. Projects evolve; check their current docs. Bittensor, Venice and Dolphin are trademarks of their respective owners.