Bittensor (TAO)

History, technology and key data — no opinions, no advice

Historical price evolution of Bittensor

Daily historical data. Updated automatically.

Bittensor was conceived in 2016 by Jacob Steeves, former Google researcher with a degree in Mathematics and Computer Science from Simon Fraser University, whilst working as a machine learning researcher at Knowm Inc. In 2019, Steeves formally co-founded the project alongside Ala Shaabana, who holds a PhD in Computer Science from McMaster University, through the Opentensor Foundation. The whitepaper was published in 2020, and the project was initially conceived as a Polkadot parachain.

In January 2021, the "Kusanagi" mainnet was launched with a complete fair launch: no ICO, no pre-sale and no pre-mined tokens. Initially, only the founding team and the first external miner, Robert Myers, operated the network. In November 2021, the "Nakamoto" update was implemented to improve protocol stability.

The project experienced significant architectural changes in 2023. In March, Bittensor pivoted from being a Polkadot parachain to becoming an independent blockchain built on Substrate, called Subtensor, which was launched on mainnet. In October of the same year, the "Finney" update introduced the subnets architecture, creating independent specialised markets for different types of artificial intelligence, expanding the number of subnets from 1 to 32.

During 2024, Bittensor attracted significant institutional investment. Polychain Capital, DCG (Digital Currency Group) and dao5 led rounds totalling approximately $350 million. Grayscale launched the Bittensor Trust, whilst Deutsche Digital Assets introduced the STAO ETP in Europe. TAO reached its all-time high of approximately $767 in March 2024, with a market capitalisation of $4,660 million.

On 13th February 2025, Dynamic TAO (dTAO) was launched, the largest update to date, where each subnet received its own alpha token tradeable against TAO in an internal AMM. The distribution of emissions shifted from a centralised validator voting model to a pure market mechanism, expanding the number of subnets from 32 to more than 120 in a few months. In December 2025, the first TAO halving occurred, reducing daily emissions from approximately 7,200 to 3,600 TAO, whilst Grayscale filed the S-1 form with the SEC to convert the Bittensor Trust into a spot ETF under the ticker GTAO on NYSE Arca.

The developments of 2026 marked important technical milestones. On 10th March, Subnet 3 (Templar) completed the training of Covenant-72B, the largest language model trained in a decentralised manner to date, with 72,000 million parameters distributed across more than 70 independent nodes under Apache 2.0 licence. This model surpassed LLaMA-2-70B in the MMLU benchmark, receiving public praise from Jensen Huang, CEO of NVIDIA, on the All-In podcast. The Targon subnet (SN4) was accepted into the NVIDIA Inception programme. On 14th March, Bitwise filed its own TAO spot ETF application with the SEC. On 3rd May, the Opentensor Foundation implemented the "Robin τ" expansion, doubling subnet capacity from 128 to 256, whilst AI usage revenues on the network reached approximately $43 million in the first quarter of 2026.

Bittensor operates as a decentralised artificial intelligence marketplace built upon Subtensor, a Substrate-based layer 1 blockchain. The network is structured into independent subnets, which function as specialised markets where miners compete by producing different types of AI outputs —from language models to financial predictions or image generation— whilst validators evaluate the quality of these outputs. The distribution of rewards within each subnet is governed by the Yuma Consensus mechanism: validators assign weights to miners according to their performance, and the protocol aligns validator incentives by making their reward proportional to how closely they approximate the network's median consensus, which prevents potential collusions.

With the dTAO update scheduled for February 2025, each subnet will issue its own alpha token tradeable against TAO through an internal automated market maker (AMM) that utilises the constant product model. This change fundamentally modifies the allocation of TAO emissions between subnets: instead of depending on validator voting, emissions are distributed according to net capital flow, such that subnets attracting greater staking receive more emissions. TAO functions as the universal entry currency to the ecosystem, as to obtain exposure to any subnet, participants must first acquire TAO and deposit it in the corresponding pool. The token has a maximum supply of 21 million units with a biennial halving schedule identical to Bitcoin's, without having had a premine or initial coin offering.

Data verified against external sources. Some values may have changed since the last update.