The AI industry runs on infrastructure owned by a handful of companies. Every chatbot, coding assistant, and image generator you use sends its requests to a server controlled by OpenAI, Anthropic, or Google. Bittensor is a bet that a permissionless protocol can break that monopoly — one machine-learning competition at a time.
What Bittensor Is
Launched in 2021 by Jacob Steeves and Ala Shaabana and incubated by the Opentensor Foundation, Bittensor is a peer-to-peer protocol that coordinates machine-learning work across a global network of computers. The core idea: instead of one company owning the AI stack, anyone who runs a useful model can earn TAO tokens if independent scorers judge their outputs to be valuable.
Bittensor is not a single AI model. It is a coordination layer — applying what a blockchain is to the AI model economy. The actual computational work happens inside specialist sub-networks called subnets, each targeting a specific machine-learning task.
The underlying chain is called Subtensor, built on the Substrate framework. It records weights, scores, and token emissions on-chain. One important caveat upfront: Subtensor currently validates transactions through a small set of nodes controlled by the Opentensor Foundation, running under a Proof of Authority model — a centralisation point that critics flag as contradicting the protocol's decentralisation narrative.
Subnets: Specialised Incentive Markets
Each subnet is an independent incentive market focused on a specific ML task: text generation, image synthesis, financial prediction, decentralised inference, GPU compute, and more.
A subnet has two actor classes. Miners produce the ML work — running models, generating predictions, serving inference requests. Validators evaluate miner outputs and submit numerical weight vectors to the Subtensor chain.
Subnet 1 (text prompting) is the canonical example. Validators issue prompts and generate reference answers; miners respond; scores are assigned by comparing miner responses to those reference answers. Every subnet defines its own scoring criteria, as long as validators and miners can agree on what "better" looks like.
As of March 2026, Bittensor runs 128 active subnets, up from 32 before a major protocol upgrade in early 2025, according to CoinGecko. Expansion to 256 subnets is projected later in 2026.
Yuma Consensus: How Scores Become Tokens
Individual scores do not flow directly into token rewards. The on-chain algorithm Yuma Consensus (YC) aggregates all validators' weight vectors into emission allocations using stake-weighted median clipping.
The mechanism matters. A validator whose scores diverge significantly from the honest majority — through error or deliberate coordination — gets penalised, and the divergent signal is discounted. Validators also build exponentially decaying "bonds" to the miners they score near consensus; maintaining those bonds increases validator rewards, creating a mutual accountability loop between the two actor classes.
Yuma Consensus imposes real costs on obvious collusion. It does not eliminate gaming entirely. Validators and miners who coordinate off-chain can still calibrate outputs to score well without producing genuinely useful model results, because the network has no trustless mechanism to verify whether an output is actually valuable versus adversarially optimised for the benchmark.
dTAO: Open Markets Replace the Senate
Before February 2025, a root network of roughly 12 high-stake validators — informally called the Senate — decided which subnets received emissions. Critics described the structure as a cronyism risk and an inconsistency for a protocol claiming decentralisation.
The dTAO (dynamic TAO) upgrade deployed on Bittensor mainnet on approximately February 13–14, 2025, after Senate approval. It replaced that allocation authority with open capital markets.
Under dTAO, every subnet now receives its own alpha token and a TAO/alpha liquidity pool. Users stake TAO into a subnet's pool and receive alpha tokens in return — effectively voting with capital for which subnets deserve more resources. Emissions follow net capital flows: subnets attracting net TAO inflows receive more of the daily emission; subnets with net outflows receive less or none.
By May 2025, emissions to alpha subnets exceeded those going to the root network for the first time. Total subnet market cap grew from roughly $4 million before dTAO to over $690 million within months. Annualised staking yields settled around 16–19% as the market found equilibrium.
A "Subnet Zero" exists for passive holders who do not want to select individual subnets — stake there to receive a blended cut across all subnets automatically.
TAO Tokenomics: Bitcoin's Playbook
TAO has a hard supply cap of 21,000,000 tokens, deliberately mirroring Bitcoin's scarcity model. No tokens were pre-mined or allocated in a private sale; every TAO in existence was earned through network participation. Emissions split roughly 50/50 between miners and validators within each subnet, modulated by Yuma Consensus.
Bittensor's first halving occurred in mid-December 2025, cutting daily emissions from 7,200 TAO to 3,600 TAO and reducing annualised inflation from approximately 10% to approximately 5%. Future halvings follow a four-year cadence — the next is expected around 2029 — until the supply asymptotically approaches 21 million around 2140.
As of mid-June 2026, approximately 11.06 million TAO are in circulation — roughly 52–53% of the maximum supply — according to CoinMarketCap. TAO's market capitalisation sits in the $2.2–2.6 billion range, with the token trading around $200–270 per TAO. The combined market cap of all Bittensor subnet alpha tokens crossed $1.4 billion by mid-2026, according to CoinGecko.
Notable Subnets: Where Real Utility Lives
Templar (SN3) is the largest subnet by market cap (~$134.9M as of March 2026). It focuses on large-scale LLM pre-training. The Templar team trained a 72-billion-parameter model using decentralised compute coordinated through this subnet — a measurable output rather than a roadmap promise.
Chutes (SN64) is the second-largest (~$132.9M). It is a serverless AI compute platform offering decentralised inference, the first subnet to surpass $100M in market cap, and it reported approximately $22,000 per day in revenue at peak. Chutes competes with centralised inference APIs on cost and censorship-resistance.
Targon (SN4), operated by Manifold Labs (~$91.8M), provides confidential GPU compute for enterprise AI and has co-authored a whitepaper with Intel engineers. Lium (SN51) operates as a peer-to-peer GPU marketplace for model training.
Many of the 128 active subnets remain low-activity — sometimes called "zombie" participants. The December 2025 halving reduced the subsidy available to unproductive subnets. Whether the dTAO market model filters them out faster than the old Senate model is still being demonstrated; capital allocators can misprice subnets on narrative rather than output quality. Current rankings and revenue data are available at taostats.io.
Bittensor as Infrastructure for AI Agents
AI agents require inference — they must call language models, vision models, or domain-specific models to execute tasks. Today, almost all such calls route through OpenAI, Anthropic, or Google APIs: centralised, permissioned, and increasingly expensive at scale.
Bittensor subnets like Chutes and Targon offer a potential alternative: permissionless inference and GPU compute accessible via standard APIs, priced by market competition rather than corporate tier structures. The decentralised compute networks covered in Part 4 of this series provide additional context on the GPU markets that subnet infrastructure ultimately depends on.
An autonomous agent wallet could, in principle, route inference calls through Chutes rather than a centralised provider — paying in TAO with no accounts, rate limits, or terms-of-service restrictions to navigate. Under dTAO, an agent could also stake TAO into the subnets whose outputs it finds most useful, creating a feedback loop between agent demand and subnet emissions. This is the speculative long-run vision. Whether it materialises depends on whether the incentive design reliably produces outputs cheaper and more verifiable than centralised alternatives.
Honest Assessment: Centralisation and Incentive Gaming
Bittensor is the most structurally serious attempt to decentralise the AI model economy. That does not mean the work is complete.
Subtensor uses Proof of Authority validation through Opentensor Foundation-controlled nodes. The Foundation can theoretically censor transactions. Founder Jacob Steeves described an 18-month roadmap toward a permissionless validator set in 2025 — a clock that has been running without a confirmed completion date.
Governance remains concentrated. Upgrades are proposed by a three-member Foundation "Triumvirate" and voted on by the 12-validator Senate. In April 2026, Covenant AI publicly exited the Bittensor network, accusing the protocol of "decentralisation theatre" and alleging unilateral founder control, according to ainvest.com. The defection carries weight: Covenant had been training a 72B-parameter model inside the network, making it a significant rather than peripheral participant.
The honest read for anyone evaluating Bittensor today: assess subnets by verifiable revenue and model outputs, not token prices or governance promises. Chutes' ~$22K/day revenue and Templar's 72B-parameter training are concrete evidence. The protocol's broader decentralisation claims require ongoing scrutiny.
Key Takeaways
- Bittensor coordinates machine-learning work across a peer-to-peer network by paying TAO to miners who produce useful outputs and validators who score them accurately using Yuma Consensus.
- The February 2025 dTAO upgrade replaced a 12-validator Senate with open capital markets: each subnet now has its own alpha token, and TAO emissions flow proportionally to net staking inflows.
- TAO has a hard cap of 21 million tokens; the December 2025 halving cut daily emissions from 7,200 to 3,600 TAO; approximately 11M circulate with a market cap of $2.2–2.6 billion as of mid-2026.
- Real utility exists in specific subnets — Chutes (~$22K/day revenue), Templar (72B-parameter model trained), Targon (enterprise GPU compute) — but most of the 128 subnets remain low-activity.
- Subtensor runs on Foundation-controlled Proof of Authority nodes; the April 2026 Covenant AI exit surfaced governance risks that a promised decentralisation roadmap has not yet resolved.
Sources
- Bittensor dTAO upgrade — CryptoBriefing
- TAO tokenomics — taostats.io
- Bittensor first halving — Grayscale Research
- TAO market data — CoinMarketCap
- Top Bittensor subnets and dTAO — CoinGecko
- Bittensor Subnets category — CoinGecko
- Centralisation concerns and Covenant AI exit — ainvest.com
- Yuma Consensus — learnbittensor.org
