On 15 December 2025 Bittensor had its first halving. Since then only half as many new TAO are created each day as before. Grayscale, which runs a TAO trust, expected the halving to support the price, according to The Block. The argument is familiar from bitcoin: at most 21 million coins, the flow of new ones halves every few years, so the asset becomes scarce.
In 2026 it looks as if that is working. By 5 October TAO was up 40 percent this year, while bitcoin had lost 2 percent.
If scarcity is why you buy TAO, one question is worth asking: how scarce is TAO today, measured against bitcoin?
What Bittensor does: an open competition for AI work
Bittensor is its own blockchain, called Subtensor, running since 2021. It hosts up to 128 so-called subnets. Each subnet is an open competition around one AI task: serving language models, renting out GPUs, training models, writing code. Anyone can join, without sign-up and without a contract.
Miners deliver the work, validators score it, and those who do well receive new TAO. Since February 2025 every subnet also has its own token. Since November 2025 a subnet receives more new TAO the more money investors put into it on a net basis.
What is technically different: it pays for the result, not the hardware
Most crypto projects for AI sell compute time. On Akash providers bid for your job in an auction, on Render you pay fixed prices per hour of rendering. Nobody there checks whether anything useful comes out. Gensyn and Ritual go one step further: they prove that a computation was carried out correctly.
Bittensor pays for something else: the result. A subnet's validators test the miners under rules written by the subnet's owner, for example how well a model solves tasks. Their scores go on the blockchain. A procedure called Yuma Consensus turns them into a stake-weighted median. If a validator rates a miner far above the majority, its score is cut back to that level, and neither of them earns anything on the excess. The point is that validators and miners cannot simply talk each other up.
On the Affine subnet it works like this: a new model has to beat the current best in every test task by a margin. Only then does the reward go to the challenger.
The weaknesses are in Bittensor's own documentation. Validators can copy other validators' scores instead of testing themselves. Submitting scores encrypted and revealing them later helps, but according to the docs only if miner performance changes fast enough. The scores are also only as good as the test rules, and miners optimise for the test. And because weighting follows stake, the large validators set the benchmark.
Bittensor does not rent out GPUs. It runs a competition and pays the winner.
What comes out of that competition?
What is already running
| Area | Subnet | What it does | What is documented |
|---|---|---|---|
| Serving language models | Chutes (SN64) | call models through an API without your own servers | provider for 6 models on OpenRouter, including Kimi K2.6 and GLM-5.1 (6 Oct 2026) |
| Renting GPUs | Lium (SN51) | GPU rental | an Nvidia H100 costs $2.15 to $2.25 an hour (lium.io, 6 Oct 2026) |
| Confidential compute | Targon (SN4) | models run in sealed machines, the operator sees neither data nor model | technology described in a paper with Intel co-authors (March 2026); the only customers named so far are other Bittensor teams |
| Training models | Templar (SN3) | training spread across the internet, anyone can contribute compute | Covenant-72B: 72 bn parameters, at least 70 participants, paper from March 2026 |
| Fine-tuning models | Gradients (SN56) | competition for the best fine-tuning | in 180 of its own tests, better than Hugging Face AutoTrain 83 % of the time |
| Analysing genomes | Minos (SN107) | tunes settings of existing genome-analysis software | 5.3 % of new issuance (4 Oct 2026) |
The most interesting result is training. Covenant-72B is the first language model of this size that anyone could help train without permission. Earlier attempts such as INTELLECT-1 only admitted selected participants. On the MMLU knowledge test the model scores 67.1, Meta's 70-billion-parameter Llama 2 scores 65.6. That is the level of 2023, not of today's leading models. The team behind Templar left Bittensor after a dispute in April 2026. Community miners restarted the subnet under the name Teutonic; no finished new model has been documented since.
For GPU rental, Lium is cheaper than the big clouds. Against specialist providers such as Vast.ai (from $1.49 for an H100) or RunPod (from $1.99) it sits in the same range, not below.
The products are real. They are not at the frontier yet.
How many people use them?
Who uses Bittensor: a few thousand addresses a day
On 5 October 2026, 89,895 transactions were signed on Bittensor, by 6,229 different addresses. That is the result of counting all 7,200 blocks of the day. More than half of it was the machine itself: 31,404 scores from validators, 15,815 entries in which miners register their models, plus server sign-ups. Around 2,400 addresses staked or traded that day, and around 2,200 sent money.
| Chain, 5 October 2026 | Transactions | Active senders |
|---|---|---|
| Solana | 164.3 m | 5.17 m |
| Tron | 11.6 m | 3.43 m |
| Ethereum | 1.77 m | 283,000 |
| NEAR | 866,000 | 96,000 |
| Avalanche C-Chain | 215,000 | 52,000 |
| Bittensor | 90,000 | 6,200 |
Sources: Artemis (Solana, excluding vote transactions), Tronscan, Etherscan, NearBlocks, Avalanche Metrics, own count of all blocks (Bittensor). For Solana and NEAR these are active accounts; definitions differ slightly.
That gives Bittensor a twentieth of Ethereum's transactions and an eighth of Avalanche's active senders, and Avalanche itself shrank sharply in September. To be fair: the actual AI work, meaning computing, training and serving models, does not happen on the blockchain. What lands there is scores, rewards and staking.
On-chain, Bittensor is mostly a machine scoring itself. The humans number a few thousand a day.
So far customers pay for only a small part of all this. According to an analysis by SubConnect, reported by Crypto Briefing on 21 September 2026, subnet revenue covers about 9 to 12 percent of what Bittensor issues in new TAO each year.
The technology runs. It is paid for mainly with new TAO.
So where is the opportunity for that to change?
Where the opportunity in the AI market lies
The gap Bittensor is aiming at is real. Three numbers show it:
| Figure | Source | |
|---|---|---|
| Cost of the largest training runs | growing 2.4-fold a year since 2016, runs above $1 bn expected from 2027 | Epoch AI, 3 Jun 2024 |
| Nvidia data center revenue | $193.7 bn in the fiscal year to January 2026, gross margin 71.1 % | Nvidia filing, 25 Feb 2026 |
| Share of open models in all requests on OpenRouter | about one third | OpenRouter and a16z, December 2025 |
Training keeps getting more expensive, the hardware for it is sold largely by one company at a very high margin, and open models that anyone may run have a firm place in the market. Whoever offers compute openly, cheaply and without a contract has room here. That is exactly what Chutes, Lium and Targon are trying.
What is not yet proven: that decentralised training can keep up with the leading models. According to Epoch AI (Jaime Sevilla, 29 Dec 2025), the largest decentralised training runs used about 1,000 times less compute than Grok 4. Decentralised training is growing 20-fold a year, the frontier only 5-fold. Epoch still considers catching up this decade unlikely. Confidential compute for business customers and coding agents are also promises without independently checked results so far.
The gap in the AI market is real. Whether Bittensor fills it is up to paying customers.
Until they pay, Bittensor funds its competition with new TAO. How many arrive each day?
The supply: bitcoin's schedule, started later
TAO copied bitcoin's supply plan almost one to one. At most 21 million tokens. The first halving comes when 10.5 million have been issued, the second at 15.75 million, and so on, exactly as with bitcoin. And after the first halving both add 3,600 new coins a day. For bitcoin that meant 144 blocks of 25 BTC each from November 2012; for TAO it is 7,200 blocks of half a TAO each.
The difference is the start date. Bitcoin began this path in 2009, TAO in 2021. On 6 October 2026, 11.59 million TAO had been issued. That is how many bitcoin existed on 20 August 2013.
| TAO today | Bitcoin today | |
|---|---|---|
| issued | 11.59 m | 20.09 m |
| new coins per day | about 3,563 (measured, 7 days) | 450 |
| new supply per year | 11.2 % | 0.82 % |
| bitcoin had this many coins on | 20 August 2013 |
So if you buy TAO, you are not buying a supply like bitcoin's today. You are buying bitcoin's supply from the summer of 2013.
TAO is as scarce as bitcoin was in 2013, not as bitcoin is today.
When does that change?
What comes next: below 1 percent only around 2037
Because the halvings are tied to the amount issued, the path can be worked out quite precisely.
| Halving | at issued | about | new TAO per day | new supply per year | bitcoin was there in |
|---|---|---|---|---|---|
| 1st | 10.5 m | December 2025 | 3,600 | 11.2 % (today) | November 2012 |
| 2nd | 15.75 m | end of 2029 | 1,800 | 4.2 % | July 2016 |
| 3rd | 18.375 m | end of 2033 | 900 | 1.8 % | May 2020 |
| 4th | 19.6875 m | end of 2037 | 450 | 0.83 % | April 2024 |
On this plan, TAO only gets below one percent new supply a year, where bitcoin is today, at the end of 2037. From its August 2013 level bitcoin took eleven years to get there. On this path TAO runs about 13 years behind bitcoin.
In dollars: about 3,563 new TAO reach the market every day. At $305.90 that is almost $1.1 million a day, or about $400 million a year. Someone has to buy those tokens just for the price to hold.
For comparison: all the known large holders together, meaning the Grayscale trust and three listed companies buying TAO as a reserve, hold about 213,000 TAO according to their filings. That is 60 days of new issuance.
The large buyers you can name hold two months of new supply.
Is the plan itself at least fixed?
The plan is software, and software can change
On 15 September 2026 at 21:42 UTC, 172,001 new TAO were created in a single block. That is as much as usually arrives in 48 days.
Behind it was a software update to the blockchain, version 459. The reason is in the source code: stakers on the main subnet had been credited dividends in the past without the matching TAO sitting in the account. Whoever tried to unstake last would have come up empty. The update closed the gap by simply issuing the missing TAO. Total issuance rose by exactly that amount.
The update was pushed through a key with special rights, a so-called sudo key. According to Bittensor's documentation it is held by three keys of the Opentensor Foundation. On bitcoin, a change like this would need the whole network to go along.
The update also shifted the schedule. Because the halvings are tied to the amount issued, the 172,001 TAO bring the next halving about 48 days closer.
With TAO the cap lives in the code. And the Foundation can change the code.
But does any of this matter for the price?
The obvious objection
"In 2013 bitcoin also had more than 10 percent new supply a year, and it still rose a thousandfold afterwards."
That is true. But scarcity was not the reason back then. Bitcoin rose because more and more people wanted bitcoin, and that demand outgrew a supply increase of more than 10 percent. For TAO it means the same: the price depends on whether enough people want to buy TAO, because they use the subnets or believe in them. The 21 million cap helps with that eleven years from now at the earliest.
The halving itself was no starting gun either. In the week after it TAO fell 20.6 percent while bitcoin rose 0.5 percent. The big rise of 2026 came later, above all in March with +67 percent.
And if you stake TAO, you get part of the new issuance back yourself. If you only hold TAO on an exchange or through a trust, you carry the 11 percent in full.
A halving does not make TAO scarce. Demand can make it more expensive.
Who holds TAO today?
Tokenomics: who has TAO and where it sits
Bittensor launched in 2021 without a presale, and by its own account without pre-mined tokens. How much the early miners received back then can no longer be checked.
64.5 percent of all TAO is staked, as of 4 October. Three quarters of that sits on the main subnet, a quarter in the pools of individual subnets. Most of the stake runs through a few operators: four validators manage 58 percent of the stake on the main subnet, among them the exchange Kraken.
The large holders look like this:
| Holder | TAO | As of |
|---|---|---|
| Grayscale Bittensor Trust (GTAO) | 46,332, plus later purchases | 30 June 2026 |
| TAO Synergies (TAOX) | about 83,800 | 11 August 2026 |
| TaoWeave (TWAV) | about 23,300 | 30 June 2026 |
| xTAO (Canada) | about 60,000 | 25 November 2025 |
Together that is just under 2 percent of all TAO issued.
The stake sits with a few validators. The large buyers hold two percent.
And how has the price done against bitcoin?
The price against bitcoin
Binance has listed TAO since 11 April 2024.
| TAO | BTC | |
|---|---|---|
| $1 on 11 April 2024 became | $0.49 | $1.23 |
| Up months | 13 of 29 | 16 of 29 |
| Median month | −4.4 % | +2.4 % |
| Months in which BTC fell | 13 | |
| … TAO did better | 5 | |
| … TAO itself was negative | 11 | |
| Largest fall from the peak | −79.5 % | −53.0 % |
| Annual volatility | 104 % | 45 % |
| Year | TAO | BTC |
|---|---|---|
| 2024 (from 11 April) | −29.3 % | +33.7 % |
| 2025 | −50.3 % | −6.3 % |
| 2026 (to 5 October) | +39.6 % | −2.1 % |
Over the whole period TAO lost 60 percent against bitcoin, half of its value in the halving year 2025 alone. In 2026 TAO is catching up, without the supply having become any scarcer.
What moves TAO's price is demand, not the schedule.
What was announced and what can be checked
| Claim | Who, when | What the data show | Verdict |
|---|---|---|---|
| The halving supports the price | Grayscale according to The Block, December 2025 | week after −20.6 % (BTC +0.5 %); new supply afterwards 11.2 % a year | does not hold as a scarcity argument |
| 21 m like bitcoin | Bittensor | cap and halvings like bitcoin; on 15 Sep 2026 an update issued an extra 172,001 TAO | cap yes, plan changeable by update |
| Grayscale US spot ETF, decision in August 2026 | Grayscale filings 30 Dec 2025 and 02 Apr 2026 | GTAO still trades over the counter; new custody agreements with BitGo and Coinbase filed on 05 Oct 2026 | did not happen |
| Bitwise TAO Strategy ETF | Bitwise, filed 30 Dec 2025 | postponed again and again since, most recently on 25 Sep 2026 | not in effect |
| Companies buy TAO as a reserve | TAO Synergies, xTAO, 2025 | TAOX holdings documented in its quarterly report, xTAO only by press release | partly |
| Subnets earn $28–35 m a year | SubConnect, end of August 2026 | a range, only 15 of 24 subnets with high confidence, individual figures not disclosed | partly |
| Customers such as PwC France and Dropbox | Crypto Briefing, 21 Sep 2026, unattributed | no confirmation by the customers, no contract found | not verifiable |
| Coding agent reaches 80.3 % on SWE-bench | Ridges (SN62) on X, 31 Aug 2025 | no entry on the official SWE-bench leaderboard found; much weaker on a different test set | not verifiable |
| Founder Jacob Steeves has no special rights | Steeves, The Block, 10 Apr 2026 | sudo key active, network halted by the Foundation after a hack in July 2024 | partly contradicted |
What sits in the code and in filings can be checked. The big customer names cannot.
What this means for you
If you buy TAO for its scarcity: think of bitcoin in 2013, not bitcoin today. Every year a good 11 percent new TAO is added, about $400 million at today's price. That only drops below 1 percent around 2037, and the plan can be changed by an update.
If you buy TAO for the subnets: then that is the real bet, and it is not made up. Chutes, Lium and Templar show that the technology runs. But demand for these services has to find buyers every year for a new supply that is larger today than everything the known large holders own together.
If you hold TAO without staking: the 11 percent of new issuance passes you by. Staking makes up for part of it, through a validator of your choice.
21 million is the cap. TAO only becomes scarce when demand swallows the new supply.
Three numbers show where this goes: new TAO per day, about 3,563 today, the status of the two ETF filings, and subnet revenue in dollars. The open question is whether the subnets earn enough by the next halving at the end of 2029 for buyers to absorb $400 million of new supply a year without the price giving way.
FAQ
Why 3,563 and not 3,600 new TAO a day? 3,600 is the plan. What was measured is the actual increase in the amount issued over the seven days to 6 October 2026. A small part flows back, for example when TAO is burned to register a new subnet.
How were the halving dates calculated? Using today's issuance of about 3,563 TAO a day up to the next threshold. If issuance changes, through such flows back or updates like the one on 15 September, the dates shift. According to Taostats the next halving falls around 12 December 2029.
What was the update on 15 September? A one-time fix called migrate_fix_root_pot_shortfall in the Subtensor code, version 459. It issued 172,001.44 new TAO and booked them to the main subnet's account, so that it holds as much TAO as had been credited to stakers. Stakers' balances did not change.
How are the monthly returns calculated? Month-end close against the previous month-end close, TAO and bitcoin each against USDT on Binance, without fees. April 2024 from the close on 11 April.
All TAO monthly returns in US dollars
| Year | Jan | Feb | Mar | Apr | May | Jun | Jul | Aug | Sep | Oct | Nov | Dec |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2024 | −37.7 | 0.1 | −30.1 | 11.1 | −6.9 | 105.6 | −16.2 | 39.9 | −34.9 | |||
| 2025 | 1.9 | −27.4 | −31.4 | 58.7 | 21.1 | −22.2 | 5.5 | −10.6 | −4.0 | 58.7 | −39.9 | −24.2 |
| 2026 | −9.7 | −7.5 | 67.2 | −18.6 | 2.6 | −20.9 | −4.4 | 19.2 | 30.6 |
Figures in percent. April 2024 from the close on 11 April.
Not investment advice, not a recommendation, not a forecast — historical patterns are no guarantee.
Study the Past — Improve your Future 🥋