Bittensor is the largest decentralized AI network by market cap and one of the most-discussed crypto projects of the past two years. As of mid-2026, it runs 128 active “subnets,” each a specialized market for one kind of AI work, with a market capitalization in the billions and a steady stream of both institutional inflows and public controversies.
The basic idea: Bittensor is a blockchain network where contributors run machine-learning models inside subnets, validators score the outputs, and the network distributes TAO tokens in proportion to how useful the work is. It’s the closest thing crypto has to a working, token-incentivized AI economy at scale, and it sits at the center of the broader decentralized AI conversation.
Quick disambiguation up front: when people say “mining TAO,” they don’t mean Bitcoin-style hash mining. They mean running an AI model and submitting its output to a subnet, where validators rate it and the protocol pays out rewards.
What Is Bittensor?
Bittensor is a decentralized network that uses blockchain incentives to coordinate machine-learning work. Contributors run AI models inside specialized subnets, validators score their outputs, and the network distributes TAO tokens as rewards based on how useful that work is.
The project was co-founded in 2019 by Jacob Steeves and Ala Shaabana and is maintained by the nonprofit Opentensor Foundation. The current main network, called Finney, launched on March 20, 2023, after two earlier iterations. Bittensor is built on Substrate, the same framework used by Polkadot, and runs on its own purpose-built chain.
The most important thing to understand about Bittensor is what it isn’t. It isn’t training a single frontier model the way OpenAI or Anthropic does. There’s no “Bittensor LLM.” Instead, the network is an economy of independent subnets, each focused on a different AI task: LLM inference, image generation, prediction markets, distributed training, GPU compute, audio, search. Each subnet is its own miniature competition; the network coordinates them.
That structure is what makes Bittensor the most-cited example of “AI on a blockchain,” and the source of most of the confusion about what it actually does.
How Bittensor Works: Subnets, Miners, and Validators
The Bittensor stack has four moving parts that are easy to confuse until you’ve seen them once.
Subnets
A subnet is an independent sub-network focused on one kind of AI work. Subnet 3 (τemplar) trains large language models in a decentralized fashion. Subnet 64 (Chutes) provides GPU compute.
Others handle prediction markets, embeddings, image generation, financial signals. Subnets are created by subnet owners, who define what counts as “good work” and write the scoring rules. As of mid-2026, there are 128 active subnets, with a planned expansion to 256.
The three roles
Each subnet has three types of participant:
- Miners run the AI model and produce outputs: text, images, predictions, whatever the subnet demands. These are AI miners, not hash miners. Hardware requirements depend on the subnet.
- Validators evaluate miner outputs and assign scores. They stake TAO to participate; their scoring influence is proportional to their stake.
- Subnet owners create and maintain the subnet’s rules, incentive design, and codebase.
Yuma Consensus
Yuma Consensus is Bittensor’s mechanism for translating validator scores into TAO rewards. Validators rate miners; the network aggregates ratings weighted by validator stake; rewards distribute proportionally to miners, validators, and subnet owners. The math is designed so bad-faith validators, those whose scores deviate too much from the consensus, earn less and get diluted by honest ones over time.
Token flow
A fixed amount of TAO is emitted per block. After the December 2025 halving, that’s 0.5 TAO per block, or roughly 3,600 TAO per day. Emissions are split across subnets based on market-driven signals, and within each subnet they’re split between miners, validators, and the subnet owner.
TAO, Dynamic TAO, and the Token Economy
TAO is the native token of the Bittensor network. It’s used for staking (delegating to validators to earn a share of emissions), subnet creation (subnet owners lock TAO to register), payment for AI services, and governance.
The supply schedule is modeled directly on Bitcoin: a hard cap of 21 million TAO, with halvings that cut emissions in half. Bittensor’s first halving occurred in mid-December 2025, dropping daily emissions from approximately 7,200 TAO to 3,600 TAO (0.5 TAO per block).
The biggest economic change in Bittensor’s history came in February 2025, when the Dynamic TAO (dTAO) upgrade went live. Before dTAO, a council of 64 validators determined which subnets got TAO emissions. After dTAO, every subnet has its own alpha token and its own liquidity pool.
Users who want to support a subnet stake TAO into that subnet’s pool, receive its alpha token in return, and earn emissions proportional to that pool’s size. Subnets with more staked TAO earn more emissions. In one line: dTAO turned each subnet into its own micro-economy, with TAO sitting at the top of the stack.
For live emissions data, staking yields, and subnet-level statistics, taostats.io is the standard source.
What Bittensor Subnets Actually Do
Subnets fall into a handful of broad categories. Some look like infrastructure; others look more like products.
Compute and distributed training
Subnets focused on GPU compute and decentralized model training. The most prominent example is τemplar (SN3), which trained Covenant-72B, what was, in early 2026, the largest collaboratively pre-trained model on any decentralized network. Chutes (SN64) and Targon (SN4) offer rentable GPU capacity for inference and short training jobs.
Foundation models and LLMs
Subnets training and serving open-weight language models for inference. They compete on output quality and cost; validators benchmark responses against scoring rubrics. The economics work best for mid-sized models where decentralized serving can undercut centralized cloud pricing.
Image, audio, and multimodal
Generative-media subnets: image generation, audio synthesis, speech, and increasingly video. Miners run diffusion or transformer models; validators score outputs against the subnet’s quality criteria.
Prediction, finance, and structured signals
Subnets producing market predictions, sentiment data, sports forecasting, and structured financial signals. Clearer success metrics (was the prediction right?) make scoring more straightforward than for generative tasks.
Search and retrieval
Subnets focused on web search, embeddings, and retrieval-augmented systems – useful as middleware for AI applications that need fresh data. Subnet quality varies widely. Some generate real revenue and have product-market fit. Others are early-stage, speculative, or close to abandoned. The dTAO market is supposed to surface this distinction, and to a degree it does, but it also creates sharper boom-bust cycles for alpha tokens.
The Honest Take on Bittensor
Bittensor draws strong opinions in both directions. Three points are worth holding simultaneously.
What’s real
Bittensor has a working technical stack, a halving-based emission schedule, a market mechanism (dTAO) for allocating rewards across subnets, and major institutional backing – longtime Polychain Capital involvement and a Grayscale Bittensor Trust (GTAO) with a spot ETF conversion filing pending at the SEC. Several subnets ship usable AI services today and generate real revenue.
What’s contested
Subnet quality varies wildly. Some are genuine AI products; others are emission farms with minimal output. The network has also had material security incidents: in July 2024, a malicious package uploaded to PyPI led to roughly $8 million stolen and the chain being halted for 10 days. In May 2025, a runaway batch call attack put the network into “safe mode” for two days. Both were resolved, but they’re part of the operational record.
The Covenant exit and subnet operator risk
In April 2026, Covenant AI (operator of Subnet 3 and the team behind the Covenant-72B training run) publicly exited the network, sold roughly 37,000 TAO (about $10 million), and accused Bittensor’s leadership of unilateral control.
A Bittensor co-founder publicly disputed the characterization, and the network continued operating. The incident triggered a significant price drop and remains the clearest example of how subnet operator behavior can affect the whole network. Bittensor is one of the few crypto projects with a working answer to “what would AI on a blockchain look like.” It’s also one of the most volatile.
Conclusion
Bittensor is the cleanest answer crypto has produced to “what would AI on a blockchain look like” – 128 specialized subnets, a Bitcoin-style halving schedule, real revenue from some workloads, and serious institutional money behind it. It’s also one of the most contested projects in the space, with security incidents, subnet operator dumps, and wide quality variance. Any serious read on TAO has to hold both things at once.
