DSperse
SN2AI inference you can mathematically verify, with proof that the model actually ran your request
Every AI inference call today requires you to trust the provider's word that the right model ran and the result was not tampered with. Omron removes that requirement by attaching zero-knowledge proofs to every inference output, making it mathematically verifiable that a specific model produced a specific result. Despite currently earning zero emissions, 797 TAO flowed into the subnet in the past week, reflecting market interest in the underlying technology.
// Verifiable AI inference with ZK proofs
Omron (SN2) is a Bittensor subnet providing verifiable AI inference using zero-knowledge proofs. Miners convert AI models into zero-knowledge circuits and generate predictions alongside zk-proofs that cryptographically verify the source model and inference integrity. Validators distribute inference requests, verify the proofs, and score miners on proof size and response time.
The simple version: Every AI prediction comes with a math proof that the right model ran and the output was not altered, no trust required.
Centralized equivalent: OpenAI, Anthropic, Google DeepMind (inference APIs)
How it works:
- Miners convert AI models into zero-knowledge circuits, run inference on input data, and return both predictions and zk-proofs to validators. They compete on proof generation speed and proof size. Currently CPU-intensive, but the network incentivizes development of GPU-optimized proving systems.
- Validators produce input data and distribute inference requests to miners. They verify the authenticity of returned zero-knowledge proofs and score miners based on proof size and response time.
- The problem it solves: AI model inferences cannot be cryptographically verified today. Users of AI APIs must trust that the provider ran the claimed model and returned an unaltered result. There is no technical mechanism to confirm this.
- The opportunity: Verifiable inference is increasingly relevant for applications where the integrity of AI outputs matters: financial decisions, medical queries, on-chain agents, and any context where provability has real value.
- The Bittensor advantage: Bittensor's incentive layer funds a competitive network of miners developing efficient zk-proving systems for AI models. The network directly incentivizes GPU-optimized ZK prover development, a hard technical frontier that centralized labs have limited reason to prioritize.
- Traction signals: 797 TAO net inflow over 7 days despite zero current emissions. GitHub shows 888 commits from 9 contributors, with the last commit on 2026-04-01, active development right up to publication. Omron holds netuid 2 in the Bittensor ecosystem, reflecting early positioning.
Category: Inference and Compute | Centralized Competitor: OpenAI, Anthropic, Google DeepMind (inference APIs)
Zero-knowledge proofs for machine learning inference represent one of the harder frontiers in applied cryptography. The core challenge is that neural network computations are not naturally expressed as arithmetic circuits, and converting them requires significant engineering effort per model architecture. Omron's subnet incentivizes miners to solve this problem by rewarding proof size efficiency and speed.
Mechanism:
Validators produce input data and distribute inference requests to miners. Miners have converted AI models into zero-knowledge circuits, run predictions on the provided inputs, and return outputs accompanied by zk-proofs. Validators verify that the proofs are authentic, confirming the claimed model ran and the output was not altered. Scoring is based on proof size (smaller is better) and response time. Miners compete to build more efficient proving systems, with the network explicitly incentivizing GPU-optimized ZK proving as a development target.
The subnet has 888 GitHub commits from 9 contributors with a last commit on 2026-04-01, making it one of the more actively developed subnets by commit volume. The current emission rate is 0.0%, meaning miners earn no alpha token emissions at this time. Despite this, 797 TAO flowed into the subnet over the past 7 days. The 30-day price is down 23.4%, reflecting the zero-emission context. Root proportion is 0.174, indicating mostly organic liquidity. Official documentation is available at sn2-docs.inferencelabs.com.
- Zero emissions currently: SN2 is earning 0.0% of network emissions at the time of writing. Miners have reduced incentive to maintain operations without emission rewards, which could affect network reliability.
- Technical difficulty: Converting AI models into ZK circuits is computationally demanding. CPU-intensive proving limits throughput and practical model sizes until GPU-optimized proving matures.
- Price decline: Down 23.4% over 30 days. The combination of zero emissions and price decline means the subnet needs a catalyst, whether a new emission allocation, a proving breakthrough, or notable adoption.
- Adoption dependency: Verifiable inference is only valuable if application developers integrate it. Demand depends on use cases where proof-of-inference is prioritized over convenience and cost.
IntoTAO, out.
