Today is not a loud day. There is no single change that rewrites how anyone gets paid, and the most consequential thing in the last few days was mentioned almost in passing at the start of a community call rather than announced anywhere. So this is a short edition, and the two things worth your attention are what quietly went live around root staking, and a queue of finished upgrades that nobody has switched on.
What has actually gone live
On a community call last week, Const, who founded Bittensor, said that Root Reborn and the curated staking products that come with it had gone live roughly two hours earlier. It is worth slowing down on this, because it touches the most ordinary thing a person does with TAO.
Start from the beginning. Staking means putting your TAO behind one of the machines that keeps the network honest, called a validator, and earning a share of the new TAO the network creates. Until recently you had two broad choices. You could stake into individual subnets, which are the independent projects that make up Bittensor, each running its own business and each with its own token. Or you could stake to root, which was the quiet, do-nothing option. Root Reborn connects the two: the yield your root stake earns now gets sold into a selection of subnets, so over time you end up holding a spread of them rather than nothing in particular.
The part that landed last week is the part that matters if you do not want a second job. Each validator can now choose which subnets it sells that yield into, and that selection is called a basket. Instead of researching more than a hundred teams yourself, you subscribe to a validator's basket and let people who have watched the network for years do the picking. It is a bit like handing your shopping list to someone who goes to the market every single day. Const was careful about the limits of that idea and so are we. A basket is not a fund, nothing about it is guaranteed, and you still end up holding subnet tokens that can fall as easily as they rise. He also said that baskets are not tokens in themselves, though they could later be made to behave a little like them.
Two smaller things arrived with it. The way yield is described on the chain was standardised, so the same holding is calculated the same way wherever you look. And the tooling for curating a basket, which teams like Crucible and Mentat Minds had built privately for themselves, now exists as a standard part of the protocol rather than as somebody's side project. Const said he had already set his own weights using Arbos, one of the curators.
You can watch that system working already. Subnet 23, Trishool, said this week that it has been added back into the Arbos basket after being dropped. It had been removed because it burns 80 percent of what it earns, meaning that share of its newly issued tokens is destroyed rather than paid to anyone, and someone read that as a sign the subnet was rewarding the wrong behaviour. Trishool says that reading was wrong. Whether it was or not, look at the shape of what happened: a human curator made a judgement about a subnet, that judgement moved real money, and then it was revised. That is the new system in miniature, and it puts a great deal of weight on a small number of people's opinions.
Bittensor writes down who really runs the chain
A documentation page was merged into Bittensor's core code that states plainly, in one place for the first time, how the chain itself is operated. The answer is Proof of Authority. A known, permissioned set of machines takes turns producing the blocks that record what happened, rather than anyone with enough support being able to take a turn. Think of it as a members-only rota rather than an open competition.
The same page sets out a non-binding aim to move to Nominated Proof of Stake within about a year, which is the arrangement where whoever produces blocks is chosen by how much stake is backing them, and is open to newcomers who can attract that backing.
Writing this down is not a scandal. Plenty of young networks run this way while they are still changing quickly, because a small trusted set is easier to coordinate and easier to fix. But Bittensor's whole pitch is intelligence that no single company controls, so it is fair to know that the ledger underneath it is currently kept by a permitted list, and that the plan to change that is an aim rather than a commitment with a date.
What is coming next
None of what follows is live. It can all still change, and some of it will never ship at all.
The most interesting thing in Bittensor's pipeline right now is that the pipeline is not moving. Core software is released in numbered versions, and each one goes live only after two approvals: the automated build system signs first, and then a small group of trusted key holders, known as the triumvirate, has to add the second signature. Versions 447, 448, 450 and 452 have all been built and have all received that first automatic signature. None of them has the second. Release 448 was submitted on 21 August. Version 452 got its first signature on 29 August and joined the same waiting room. Our own feed has reported this queue growing several times over the past week, which is really the story: not what any one release does, but that the pile keeps getting taller while nothing leaves it.
Some of what is stuck matters to ordinary people. Release 450 is the one that would have the network publish a single official price for staking baskets. At the moment every wallet, explorer and app works out what your basket is worth using its own arithmetic, so the same holding can show a different number in three different places, which is confusing at best and a nuisance if you are keeping records. Version 451 raises how long a linked order stays alive from seven days to a hundred and eighty. That sounds like nothing until you know that advanced orders are signed to last a hundred and eighty days, so instructions people believed were standing were built on a component that expired long before they did.
So who does the delay help, and who does it hurt? Requiring a human signature is exactly the guard you want on a network that holds real money. Nobody should want upgrades landing automatically because a robot approved them, and a signer who pauses to look properly is doing the job. But a queue four deep has costs. Fixes people are already relying on sit undelivered. And the longer a queue sits, the bigger the eventual change becomes, and a large combined change is harder to reason about and harder to unpick if something goes wrong. It is also where the Proof of Authority point bites: a network whose upgrades wait on a signature from a handful of people has a human bottleneck no matter what the machines are doing. Nobody has said anything is wrong, and there may be a perfectly good reason nobody has published. What we would watch is whether the queue clears in one batch, whether it stays put another fortnight, and whether anyone explains why it stalled.
The second proposal worth understanding is not code at all. It is a community idea called fractal emissions, and it is being discussed rather than proposed formally. Emissions are the new TAO the network creates every few seconds and hands out, split between the people doing the work on each subnet and the people backing them. When new TAO flows to a subnet, new units of that subnet's own token get created alongside it, which dilutes existing holders. Dilution simply means each token you already hold represents a slightly smaller slice of the whole than it did yesterday.
Fractal emissions would let each subnet voluntarily accept less than it is entitled to: all of it, half, a quarter, an eighth, or less. Take less, create fewer new tokens, dilute your holders less.
For a subnet with real income from paying customers, that could be attractive. It says out loud that the team does not need the subsidy, and the holders keep a larger share of what exists. For a young subnet with no customers, it would be close to suicide, because emissions are what pays the miners who actually do the computing. And that is where it could go wrong. If taking less becomes a badge of seriousness, teams may take less than they can genuinely afford in order to look confident, and the people doing the work get squeezed for a signal. Equally it may turn out that almost nobody opts in, and the whole thing is decorative. Either way, remember what emissions are for. A subnet that turns them down is betting that its customers will cover the bill instead.
Beneath those, a cluster of smaller proposals all point the same direction: root, baskets and delegation. One would let a staking proxy, which is a limited permission slip one wallet gives another, actually claim root rewards, because at the moment a delegated setup can quietly stop collecting them and the owner never signed anything wrong. Another adds a narrow new permission covering only the setting of root weights, so you can hand somebody the job of curating your basket without handing them anything else. Another lowers a fee quote for claiming root rewards that was being calculated as though there were more subnets than the network allows. There are also housekeeping fixes: a check that stops a wallet approving a payment it cannot actually afford once fees and minimum balances are counted, and a shortcut that lets a new machine joining the network start from a trusted signpost rather than replaying every block since the beginning.
The subnets
The clearest theme this week was money arriving from outside the network rather than from emissions. Numinous, subnet 6, made its first buy-back: it spent 4 TAO purchasing its own token on the open market, funded by 10 percent of what its paid forecasting service took in across July and August, and published the transaction so anyone can check it. The amount is tiny and the team did not pretend otherwise. What matters is the mechanism and the receipt: income from customers, converted into buying pressure, shown in public.
On the same theme, a researcher's index shared on Friday by the ReadyAI team estimates that Bittensor's subnets collectively bring in roughly 32 to 35 million dollars a year from paying customers, spread across about 25 of them, with a projected path past 100 million. Treat that as a claim rather than a fact. It is an outside estimate, we cannot see its method, and real revenue is precisely the number critics say the network lacks, which means a figure like this will get repeated widely regardless of how solid it is.
Two compute stories fill in the picture. Nodexo, subnet 106, added 22 H100 graphics cards, the expensive Nvidia chips that most serious AI work runs on, and says 14 of them are free to rent right now with no waiting list. And on a podcast this week, Josh from subnet 110 explained a different angle: putting small data centres on UK farms that generate more renewable power than the grid can take. He said local substations are often too small for a big solar field, and turbines get switched off when it is too windy, so power gets thrown away. His figures, and they are his: selling that power back to the grid earns roughly 8 to 12 pence a unit, while running compute on it has earned 60 to 80 pence and sometimes a pound. In his worked example a farm that spent five million pounds on turbines and a further million on servers could see a fifteen year payback fall to four or five. Those are the numbers of a person selling a plan, not audited results, but the underlying observation about wasted generation is easy to check and rarely discussed.
The most vivid product story came from subnet 17, the 3D generation team known as 404. Ben from that team explained why they abandoned Gaussian splats, a technique that produced stunning images but that studios could not use, because it does not fit the file types and production pipelines that game engines already expect. They now build meshes, the ordinary building blocks of game art, using a model called Titan that miners built on top of an open source model called Trellis 2. He says Titan produces a model in about 30 seconds against several minutes for most closed competitors. A second layer called Atlas turns raw output into something a game can actually use, including the unglamorous work of putting the pivot point in the right place so a weapon sits in a hand rather than through it. A mobile game called Colony, made by Parallel, is using it in early play testing to let players describe an item and have it made while they walk across the map, and Ben says the studio reports that this is the single most liked feature in testing. He also said two people built a game with 404 technology that won the inaugural Chroma Awards, generating over 12,000 assets. That is one of the clearer cases of Bittensor output being used by someone outside Bittensor who does not care where it came from.
Elsewhere the pattern was teams removing friction. Enigma, subnet 63, will now accept work through a web page, so a quantum researcher with an answer no longer has to run miner software and configure a wallet first. TensorUSD's dollar-pegged token now appears in the SubWallet app automatically, instead of requiring a newcomer to paste in a long technical address. Yanez, subnet 54, says its private sign-in is now live inside two other projects and that only 2 percent of new users needed to contact support. SOMA's matching algorithm is now running inside another Bittensor project's app, SayGm, which is a small but real sign that one team's output is good enough for another to depend on. Tatsu spent the week building a single prepaid credit that covers all its AI features, and made the sensible point that taking the money is easy while making sure nobody quietly uses more computing than they paid for is the hard part.
On the research side, subnet 3's Teutonic began training a model of about 110 billion parameters on strangers' machines rather than in one company's data centre, which is the hardest version of the problem Bittensor exists to solve. Cacheon, subnet 29, reported a new speed record of 2,274 tokens per second in its optimisation contest. Beam is wiring itself into Hugging Face, the site where most freely available AI models live, taking on the job of moving very large files quickly rather than storing them. Open Quantum republished a much expanded whitepaper after nine months of drift between the document and the product.
Finally, the comings and goings. Subnet 111, DeSci Claims, says its launch is complete and emissions have been switched on, which is the moment a subnet stops being a test and starts being paid. In the other direction, subnet 59, Babelbit, was deregistered, meaning it lost its slot and its share of new TAO, roughly a week after announcing a dubbing tool. Slots are competitive and the network pays on its own scoring, so shipping a product does not save you. Zeus, subnet 18, told holders that the recent selling of its token came from a single early backer who had paid its registration fee in exchange for a share of daily tokens and is now winding down, and says that selling is finished. We cannot independently verify that. Targon gave free computing credits to Cascade, the first team out of BitstarterAI's training programme, to help it launch. A group called Kusanagi set itself up as a business layer around Bittensor rather than registering a subnet, running a validator and the TaoSwap site. Masx.ai named Monday at 9pm central European time for its version three beta. And Tao.com says the automatic sell orders it added recently are the most adopted feature it has shipped this quarter, which is a quiet reminder of why the 180 day order fix stuck in the queue is not a trivial one.
The thing to keep an eye on next week is whether that queue clears.