Bittensor’s Subnet 2, operated by Inference Labs, has generated more than 300 million zero-knowledge proofs as it develops a decentralized infrastructure layer for verifiable AI inference. The network’s current documentation also reports more than 1,500 unique miners and a 10x improvement in average proof-generation time during 2025, providing measurable indicators of how its zkML proving system has scaled.
According to the official Subnet 2 documentation, the network converts AI models into zero-knowledge circuits so miners can return both an inference result and a cryptographic proof of how it was produced. The proof is designed to establish that the specified model processed the committed inputs and generated the reported output without requiring a validator to rerun the full inference.
Miners Compete to Produce Faster zkML Proofs
Subnet 2 separates proof generation from verification. Validators distribute inference requests, while miners execute supported models through zk circuits and return the output together with a proof. Validators then check that proof using the relevant verification key and score miners according to validity, response time, proof size and output accuracy. The incentive system therefore rewards the efficiency of cryptographic verification rather than simply accepting an AI output at face value.
The architecture has evolved beyond the earlier implementation of verifiable ZK-ML proofs. Inference Labs now uses its DSperse orchestration layer to divide models into segments that can be proven in parallel, while supporting multiple proving backends including JSTprove, built on Polyhedra’s Expander, and Circom-based systems. Parallel proving is intended to reduce one of zkML’s main operational constraints: the time required to generate cryptographic evidence for increasingly complex model computations.
The network’s documentation says competition among miners drove average proof times down by an order of magnitude during 2025. That 10x figure represents a project-reported improvement in proof generation, not proof that every AI workload now receives low-latency verification. Actual performance depends on the model, circuit, proving backend, hardware and complexity of the inference being proven.
Subnet 2 Builds a Verification Layer for AI
The distinction between proof verification and on-chain execution is also important. Miners and validators communicate outside Bittensor’s consensus execution path, with validators ultimately setting performance-based weights on-chain. Inference Labs additionally uses decentralized proof storage for selected proofs and has developed Proof-of-Weights infrastructure aimed at making validator scoring independently verifiable. Bittensor supplies the incentive and coordination layer, while the computationally intensive zkML proving occurs through Subnet 2’s distributed infrastructure.
That model differs from other specialized Bittensor deployments, including private inference infrastructure developed around Subnet 46. Subnet 2’s primary role is cryptographic verification of computation, rather than simply supplying private or general-purpose AI inference. Its validator APIs can also accept external requests, creating a path for applications outside the subnet’s own incentive traffic to use its proving network.
The next meaningful milestone will be broader external use of that infrastructure rather than another aggregate proof count alone. Third-party models deployed through Studio, sustained organic inference requests and continued reductions in proof time and cost would provide stronger evidence that Subnet 2 is becoming production infrastructure for verifiable AI. For now, the 300M+ proof count and 1,500+ reported miners demonstrate operating scale, while the practical economics of zkML for larger real-world models remain the more important test.
Natalie Pierce tracks the parts of crypto that move fast and rarely wait for everyone to catch up. From South Africa, she covers DeFi, AI crypto, hacks, airdrops, sentiment and emerging narratives, especially when user behavior and protocol risk start to overlap.
Her reporting is built for messy sectors. Natalie looks at incentives, reactions, security concerns, social momentum and early signs of traction without pretending every new trend is already proven. Her voice is clear and accessible, but careful enough for areas where excitement can outrun the facts very quickly.