Bittensor’s subnet architecture is moving beyond abstract incentive design toward systems built to execute and verify AI inference. Subnet 2 provides a concrete example through infrastructure that lets miners generate AI outputs and validators cryptographically verify that the expected model produced them, turning decentralized compute into an executable inference workflow.
The public Subnet 2 repository from Inference Labs describes the project as a Proof-of-Inference system operating on Bittensor netuid 2. Its software includes production miner and validator binaries that communicate over HTTP and QUIC and can register directly on Bittensor mainnet. The code therefore documents an operational inference architecture, although it does not provide independent figures showing how much external demand the subnet currently handles.
Subnet 2 Links Inference to Cryptographic Proofs
Under the model, validators distribute inputs to miners, which run AI models and return both the prediction and a zero-knowledge proof. Validators then verify those proofs and score miners using metrics such as proof integrity and response time. The objective is to prove that an inference came from a specific model instead of requiring users to trust the compute provider’s claim about what actually ran.
Bittensor’s own subnet directory currently describes Subnet 2, branded DSperse, as providing verified AI inference powered by zero-knowledge proofs. The directory also shows active miners and validators on the subnet, but those figures measure network participation rather than customer demand. Running infrastructure is therefore observable, while persistent commercial usage remains a separate question that available data does not yet answer.
That distinction matters for evaluating Bittensor’s broader model. A subnet can successfully coordinate miners, validators and incentive weights without necessarily establishing that outside applications are routing significant production workloads through it. The next credibility threshold is not simply demonstrating that inference works, but showing recurring users willing to depend on and pay for the service.
Other Subnets Move Toward Pay-Per-Inference Models
Subnet 53, now known as Engy, offers another example of that shift toward practical inference infrastructure. Its public documentation describes an OpenAI– and Anthropic-compatible gateway where GPU miners serve model requests and responses carry cryptographic verification. Engy’s scoring model counts paid requests and incorporates token usage into miner rewards, explicitly tying Bittensor incentives to inference that customers actually purchase.
Engy is still rolling out parts of its verification and incentive stack in stages, illustrating why individual implementations should not be treated as evidence that the entire Bittensor ecosystem has reached production maturity. What Subnets 2 and 53 do demonstrate is a structural move toward networks where AI inference can be routed, priced and independently checked rather than merely benchmarked for emissions.
For Bittensor, that transition could prove more meaningful than raw subnet counts. The key measure will be whether verifiable inference systems develop durable paid workloads beyond their own incentive mechanisms, turning decentralized AI subnets from functioning technical experiments into infrastructure that outside developers and applications routinely use.