Founded in 2023 by machine learning experts Ron Chan and Colin Gagich, Inference Labs is solving one of AI's most critical challenges: how to verify AI outputs without exposing sensitive data. The Canadian startup operates Omron (Subnet 2) on the Bittensor network, now the world’s largest zkML proving cluster.
Through partnerships with EigenLayer, EZKL, Bagel, and Arweave, Inference Labs is building the cryptographic trust layer for verifiable, decentralized AI.
The team has raised over $6.3 million from Delphi Ventures, Mechanism Capital, and DACM—alongside a $1 million public sale on Legion in June 2025.
This article explores how Inference Labs solves the AI trust problem, the Omron subnet, and Arweave’s role as the decentralized storage layer for AI.
As AI adoption accelerates and autonomous agents emerge, there's a pressing need for trustless systems to verify model outputs and ensure agent actions are auditable by third parties. This requires verifying the computing party, the specific model used, and how input data is handled. However, performing such verification without privacy protections may expose internal model data and compromise user privacy.
"The AI trust problem is a $50 billion market waiting to be solved. When autonomous agents are managing your finances or making medical decisions, you need mathematical guarantees, not just promises." — Colin Gagich, Co-founder
To solve this problem, Inference Labs has built a zkML system as a subnet on Bittensor (Omron) that verifies model outputs without trusting the operator or revealing the model’s inputs or weights. Let’s explore how Omron works.
Omron is a decentralized network that incentivizes model distillation (reducing model size for verifiable AI) and prover optimization. It uses a Proof-of-Inference mechanism to cryptographically verify zero-knowledge proofs of AI outputs, confirming a specific model produced a result without exposing the model or input data.
Zero-knowledge proofs are inherently compute-intensive and often impractical for large models. Omron addresses this challenge by incentivizing miners in the Bittensor network to distill and run models that generate fast, valid proofs inside zero-knowledge circuits (programs that prove a model produced the correct output without revealing how). Model distillation and optimization significantly reduce proof generation time and costs, making verifiable inference viable at scale.
While model distillation enables scalable zkML, it doesn’t guarantee that AI outputs can be trusted.
To ensure each inference is computed as claimed, Inference Labs built an optimized proof network where subnet validators verify model outputs and use an on-chain scoring system to reward miners based on correctness and latency.
How Proof of Inference works:
Users or applications submit AI inference tasks to the network, which validators broadcast to miners.
Miners select a model they believe will perform best, run the inference, and generate a cryptographic proof of correctness. The proof reveals neither the model weights nor the input data.
Miners return both the model output and the proof to the validator.
Validators verify each proof to ensure the inference was executed using the specified model and that the output is valid. Only outputs with valid proofs are accepted.
The validator returns the verified result and proof to the user, and then scores miners based on proof validity and latency. These scores determine on-chain rankings and TAO rewards. Miners who consistently fail to provide valid proofs or perform slowly receive reduced rewards.
Proof of Inference economically incentivizes miners to compute truthfully and optimize their hardware for proof acceleration, creating a competitive marketplace for verifiable compute that can scale with demand.
To understand how Omron fits into the broader Bittensor ecosystem, explore the Bittensor documentation and subnet development guide.
While Omron stands as the first fully decentralized zkML proving network, Inference Labs realized that making AI inference outputs transparent and persistently verifiable requires storing proofs on a permanent, decentralized storage network.
To solve this, Inference Labs partnered with AR.IO to develop the Proof of Publishing system to permanently record proofs on Arweave’s Permaweb. This system allows Subnet 2 validators to automatically publish inference proofs as permanent Arweave transactions. Each proof is timestamped and linked to a specific model, input, and miner.
Unlike centralized storage, Arweave uses cryptographic guarantees and economic incentives to ensure data permanence. As a result, every model output from the Omron zkML network stored on the Permaweb becomes publicly accessible, tamper-proof, and independently auditable by anyone worldwide.
For Inference Labs, the choice of Arweave was strategic. "Arweave gives us something no other storage solution can—mathematical certainty that these proofs will exist forever. When you're building trust infrastructure for AI, permanence isn't optional," says Colin Gagich.
How the Proof of Publishing works:
Validator Submission: After receiving inference results and their corresponding zero-knowledge proofs from miners, validators forward both to the Proof of Publishing System (PPS).
Upload to Arweave: The PPS uses AR.IO's open-source Turbo SDK to bundle the data and upload it to the Permaweb via a subsidized server configured by Inference Labs.
Permanent Record: Once confirmed, each proof is stored immutably on Arweave. Validators receive a transaction ID that serves as a permanent, content-addressed pointer—allowing anyone to retrieve and independently verify the data at any time.
Inference Labs’ integration with Arweave provides Omron with the following benefits:
Real-time verification, allowing APIs, applications, and DeFi protocols to instantly access published proofs for re-verification and risk mitigation.
Enterprise-grade auditability, making AI outputs suitable for industries such as finance, healthcare, and governance.
Since launching on Bittensor, Omron achieved the following:
Operates the world’s largest zk-inference marketplace via a global peer-to-peer proving cluster.
Enables verifiable AI across EVM-compatible chains.
Reduced median proving times from 15s to 5s.
Supports multiple proving frameworks such as EZKL, Circom (Groth16), and a16z’s Jolt.
In addition to Arweave, Inference Labs has partnered with other leading decentralized AI protocols to build zkML solutions:
EigenLayer: Integrated its economic security infrastructure into Inference Labs’ ZK‑VIN, creating the Sertn AVS, which achieves 76% faster zkML proof generation. See the EigenLayer blog post for more details.
EZKL: Developed Metal bindings to enable zero-knowledge proofs on mobile devices; Inference Labs integrated these into Subnet 2 to support privacy-preserving applications like on-device age verification. See the EZKL blog post for more details.
Bagel: Collaborated with Inference Labs to integrate verification protocols into Bagel’s decentralized fine-tuning marketplace, ensuring model integrity and protecting intellectual property. See the Inference Labs blog post for more details.
These partnerships position Inference Labs as a leader in zkML, as they build the infrastructure layer for verifiable, transparent AI outcomes.
Inference Labs combines Bittensor’s decentralized inference with Arweave’s permanent storage to create a scalable security stack for autonomous agents. As AI agents surpass human users online, trustless infrastructure becomes essential. Inference Labs leads decentralized AI and zkML development, aiming to scale AI while preserving privacy and ownership over personal and proprietary models.
Learn more in Inference Labs’ blog post on their Arweave integration, or follow their updates on Twitter/X.
Developers need a foundation that’s not only decentralized but also secure, private, and persistent. Whether you're building autonomous agents, AI protocols, or data-heavy apps, Arweave ensures your data stays verifiable and immutable forever.
Future‑proof your AI stack with Arweave:
About the author: Vitti is a Web3 writer who turns complex ideas into clear content.
Follow Vitti on X for insights into AI and Web3.