# Into: DSperse Subnet 2 is one of Bittensor's oldest slots, and it just changed its name: Omron is now DSperse, a zkML proving cluster that aims to hand you a cryptographic receipt proving an AI model actually ran your request. // Proof that the model really ran --- > New to Bittensor? Start here. Experienced users can skip to the full analysis. ### What is DSperse? DSperse is a network for verifiable AI inference. When a normal AI service answers a question, you have to take its word that the answer came from the model it claims to use. DSperse turns that promise into a proof: it runs the model and attaches a zero-knowledge proof showing the output really came from that specific model, untampered. **The simple version:** It's like a tamper-evident seal for AI. Anyone can check the seal and confirm the answer came from the right model, without trusting the company that served it. **Centralized equivalent:** A hosted API like an OpenAI or AWS inference endpoint. You send a prompt, you get an answer, and you trust the provider that the right model produced it. DSperse replaces that trust with a proof you can verify yourself. **How it works:** - **Miners** run AI models that have been compiled into zero-knowledge circuits, generate the prediction, and return it together with a proof. - **Validators** hand out the work, then verify each miner's zero-knowledge proof and score it on proof size and response time. --- ### Why This Matters - **The problem it solves:** AI output is unauditable. When a model makes a financial call, a moderation decision, or an oracle reading, you cannot independently confirm which model ran or that nobody swapped the result. DSperse makes that checkable. - **The opportunity:** Verifiable inference is a building block other systems can sit on top of, from on-chain oracles to AI used inside agreements where the parties do not trust each other. The new "DSperse" branding leans into this, positioning the subnet as a source of verifiable oracles for any computation. - **The Bittensor advantage:** Generating zero-knowledge proofs is expensive, so it suits a decentralized cluster of miners competing on speed and efficiency. Bittensor's incentive mechanism pays for exactly that: a standing fleet of provers rather than one company's servers. - **Traction signals:** Development is active. Inference Labs, the team behind the subnet, has raised over $6 million for verifiable AI, according to The Defiant. The codebase carries over 2,100 GitHub stars. --- ## Full Analysis **Category:** Inference and Compute | **Centralized Competitor:** Hosted inference APIs (OpenAI, AWS, Google), trusted-hardware attestation Subnet 2 has been around since April 2024, run by Inference Labs, and for most of that time it went by Omron. The on-chain identity now reads DSperse, and the website frames it as "the world's fastest zkML proving cluster" delivering "verifiable oracles for any computation." The work underneath is the same lineage of verifiable inference, with a sharper product name and a specific technical bet attached to it. **Mechanism:** The miner and validator are native Rust binaries that talk over HTTP and QUIC. According to the subnet's repository, miners receive input data from validators, run custom AI models that have been converted into zero-knowledge circuits, and return the prediction along with a proof. Validators produce the input, distribute the requests, verify the authenticity of each returned proof, and score miners on cryptographic integrity, proof size, and response time. Because the heavy lifting is the proof rather than a giant GPU job, the design lets non-GPU machines participate, while still rewarding faster, more efficient proving systems. The DSperse name points at the team's approach to a hard problem: proving a whole model is costly. As community write-ups describe it, DSperse leans on targeted verification, splitting a model into slices and proving only the critical ones rather than the entire pipeline, which cuts the proof cost and memory needed to make zkML practical. The repository reflects this in its handling of model slices and per-slice circuits. There is a second thread here worth knowing about. The same proving capability backs Proof of Weights, where validators on other subnets can produce a proof that they ran their incentive mechanism correctly. The subnet's own documentation describes Subnet 2 positioning itself as an "execution layer" for that work, citing a capability of producing over 300,000 proofs per day. Treat that figure as the team's stated capacity rather than an independently measured number, but it captures the ambition: a general-purpose proving cluster other parts of the network can lean on. On the market side, the alpha token trades around 0.00531 TAO with a market cap near 27,000 TAO and roughly 10,000 TAO of root liquidity in the pool. Emissions are a softer spot. Net staking flows have been negative over the past week, and under Taoflow, the flow-based model Bittensor adopted in November 2025, negative net flows leave a subnet with little to no current emission. A subnet that built genuinely useful infrastructure still has to attract net staking to earn its share, and right now that signal is muted. --- ### Risk Factors - **Deregistration:** Subnet 2 is well past its 4-month immunity window, so it is exposed to automatic deregistration if its EMA price sits at the bottom of the non-immune set. With recent net outflows and minimal current emission, this is the risk to watch. - **Adoption over demos:** Verifiable inference is technically real, but the value depends on other systems actually consuming the proofs. A proving cluster needs paying demand, on-chain or off, to matter beyond the benchmark. - **Execution and rebrand risk:** Renaming Omron to DSperse and narrowing toward verifiable oracles and targeted verification is a bet. The work is genuine, but the product framing is newer than the subnet, and reputational continuity has to carry across the name change. - **Concentration:** Ownership of the alpha token is concentrated, with a high gini reading across the top positions. A few large positions moving in or out can swing a thin pool. --- Another subnet, unpacked.