§1 — The brief.
The brief.
The AI compute market does not lack GPUs. It lacks a way to buy inference that is verifiably real, private, and priced below the centralized clouds. Y.AI’s answer is a Layer-1 blockchain where the work that secures the chain is the product: instead of burning energy on throwaway puzzles, miners serve open-model inference requests. The protocol verifies the work, settles payment on-chain, and pays the GPU twice for the same request — the client’s fee plus the block reward. That subsidy is designed to let the network price below them.
Zpoken builds it end to end — the chain, the verification pipeline, the execution and privacy model, and the platform API, from the codebase up, through devnet, toward public mainnet.
§2 — What we shipped.
What we shipped.
[01]
Layer-1 chain — score-weighted proof-of-useful-work consensus; one contributor class, three roles drawn dynamically (miner · computation validator · block proposer)
─ DEVNET
[02]
Inference verification (IPE — Inverse Prompt Engineering) — three-layer pipeline: structural filtering · loss-integrity scoring · log-probability fingerprinting (KL divergence against the designated model’s statistical fingerprint)
─ DEVNET · HARDENING FOR MAINNET
[03]
Privacy & execution model — private routing, isolated sessions, RAM-only execution, opaque on-chain receipts
─ DEVNET
[04]
YAI Platform — OpenAI-compatible REST API (api.y-ai.io/v1): chat completions, embeddings, tool calling; workload routing to qualified GPU supply
─ BUILT
§3 — Engineering decisions.
Engineering decisions.
Verification as a property of mining, not an add-on.
IPE reframes GPU work as a search problem with a verifiable outcome: the network presents a target text, miners search for a prompt that reproduces it, and every submitted share carries the prompt, output, loss value, and log-probability fingerprint. Client inference requests flow through the same pipeline — route, execute, verify, settle — so the same three verification layers cover both the consensus workload and paid work. Each layer rejects a different class of attack; the fingerprint check defeats the cheapest one, substituting a smaller model for the designated one. Validators re-run verification before a share lands in a block.
One contributor class, three dynamic roles.
Most chains split miners, validators, and stakers into permanently distinct populations with separate incentive curves — which turns into a political problem over time. Y.AI has a single contributor class scored by a composite performance formula; miner, computation validator, and block proposer are roles drawn from that class dynamically, not castes.
A purpose-built, inference-native L1.
Routing, model registry, verification, and settlement are first-class protocol operations, not contracts retrofitted onto a general-purpose chain. Open models — Qwen, DeepSeek, Kimi — can be swapped in without coordinated downtime. The intelligence layer is replaceable; the ledger keeps ticking.
Operational privacy now, cryptographic privacy on the research track.
Today: private requests route over point-to-point channels instead of the public task stream, execution is RAM-only in isolated sessions, and on-chain they appear as opaque hash-only receipts — work and settlement only, never content. The research track (AloePri, equivariant weight encryption) is designed so the executing server sees neither weights, plaintext input, plaintext output, nor intermediate activations. The published materials state plainly which guarantees are live and which are roadmap.
§4 — Where it stands.
Where it stands.
The verification algorithms are implemented and validated as a dedicated service; consensus-path enforcement runs on devnet and is hardening for mainnet. The published roadmap targets public mainnet in October 2026.
§5 — Tech stack.
Tech stack.
~/zpoken/work/y-ai/stack
Purpose-built L1 (score-weighted proof of useful work) · IPE verification service · open models (DeepSeek · Qwen · Kimi) · YAI Platform REST API (OpenAI-compatible, api.y-ai.io/v1)
§6 — Public surface.
Public surface.
§7 — Founders.
Founders.
Engagement led by Mike Yezhov and Anton Yezhov.