Two things happened this week that pull in opposite directions.
The first is that Bittensor switched on a genuinely new way of staking, one that hands the hardest decision a holder faces to someone else. The second is that a queue of finished upgrades is sitting on the shelf, built and packaged and waiting for a human being to sign them off, and the queue has grown to four. Both stories are about the same underlying question, which is how much of this network still runs on a small number of trusted people.
The new way of staking to Root is switched on
Start from the beginning, because this only makes sense if you know what Root is.
Bittensor is not one project. It is a network of separate teams, called subnets, each doing a different job: one generates 3D models, one forecasts prices, one rents out computers. Each subnet has its own token, called an alpha token, whose value rises and falls with how the network rates that team's work. TAO is the main currency that sits above all of them.
When you stake TAO, you are locking it up to support the network and earning a return for doing so. The awkward part has always been where to point it. Stake into a specific subnet and you are effectively backing that one team out of more than a hundred, most of which you have never heard of and cannot easily judge. Stake into Root instead and you were putting your TAO into the shared pool that sits at the centre, the safe and boring option, which for a long time paid accordingly.
Root Reborn is the rebuild of that shared pool, and the piece that landed this week is the interesting one. On a community call on 28 August, Const, Bittensor's founder, said that Root Reborn and its curated staking products, which he calls beta baskets, had gone live within the previous two hours. A basket is a chosen set of subnets. Each validator, meaning one of the machines that checks the network's work and has usually been watching this ecosystem for years, can now select which subnets it points root staking yield into. Instead of researching a hundred teams yourself, you subscribe to a validator's basket and let them do the picking.
The closest everyday comparison is choosing a fund rather than choosing shares. You are not buying one company, you are buying somebody's judgement about which companies to hold, and you are trusting that they are better at it than you are. Const was careful about this on the call and it is worth repeating in his terms: it is not risk free. You still end up holding alpha tokens, which can go up and can go down, and a basket is a lower effort entry rather than a safer one. He also mentioned he had already set weights with Arbos, one of the curated baskets, and that baskets are not tokens in themselves, though they can be made to behave a bit like them.
There is a guardrail attached, and it is the part most worth understanding. No basket may put more than one sixteenth of its weight, a little over six percent, into any single subnet. The whole point of rebuilding Root was to stop everyone's stake piling into whichever team was fashionable that month, so allowing a curator to dump a fund's entire weight into one favourite would have recreated the problem with an extra middleman. The cap forces spread. It also means that if a curator is spectacularly right about one subnet, you only get a sixteenth of being right.
One honest complication. The change that lets curators select subnets was merged into the code as runtime 449, and 449 never actually shipped to the live network. Its work is now being folded into a larger bundle numbered 500. So we have the founder saying on a call that baskets are live and people can subscribe, and we have the code trail saying the release carrying subnet selection has not gone out. Both are being reported here because both are what we can see. The likeliest reading is that the pieces reached the live chain by different routes and at different times, but we are not going to tell you we know that. If you are considering putting stake into a basket, the sensible thing is to look at whether the basket you want actually exists and holds what it claims to hold, rather than trusting either a call or a changelog.
A related repair did get made. Developers found that if one subnet inside a basket had no usable liquidity, meaning there simply was not enough of its token available to trade in and out of, an attempted claim could fail outright and drag healthy positions down with it. Picture a hamper where one item is out of stock and the shop refuses to hand over the whole hamper. A hotfix was merged so that a dried up holding no longer blocks everyone else's payout. Whether that fix is running on the live network today depends on the signature queue, which brings us to the second story.
Bittensor writes down who actually produces its blocks
Quietly, a documentation page was merged into the core code that states in one place, for the first time, how the chain itself is operated. It uses Proof of Authority. That means a known, permissioned set of machines takes turns producing blocks and confirming them. It is not the case that anyone with enough stake can join in and help run the chain.
This will surprise people, because the word decentralised gets used about Bittensor constantly, and it is decentralised in an important sense: the actual intelligence work, the mining and validating of AI tasks, is open to anyone who can do it well. The block production underneath, the part that decides which transactions happened and in what order, is not. It is a named list.
The same page sets out a non-binding aim to move to Nominated Proof of Stake within about a year, which would mean the machines producing blocks are chosen by how much stake backs them rather than by who is on the list. Non-binding is doing real work in that sentence. It is a stated intention, not a commitment with a date attached.
Why this matters to an ordinary holder is not that something changed, because nothing did. It is that the honest description is now written down and can be pointed at. A network that publishes its own uncomfortable facts is easier to judge than one that leaves you to infer them. What it could lead to, if the intention holds, is a chain that stops depending on a specific set of trusted operators. What it could equally lead to is another year of the same arrangement with a document acknowledging it.
Two quiet repairs that make the plumbing better
Two smaller changes were merged and are worth a paragraph each.
The first involves precompiles, which are built-in shortcuts that let Ethereum-style smart contracts reach network features like staking without writing that logic from scratch. Think of them as buttons wired directly into the machinery. The fix means a shortcut will now only run when the contract being executed is genuinely the shortcut itself, rather than one contract reaching for it while standing somewhere it does not belong. Purely cryptographic shortcuts, which hold no state and cannot be abused this way, are exempt. This is the sort of hardening that never produces a headline unless it is missing.
The second is about new computers joining the network. Normally a machine has to work through the chain's entire history before it can take part, which is slow enough to put people off running one. There is a fast path, called warp sync, that lets it jump most of the way by following a trail of signed handover records, but it needs a known good starting point to trust. That starting point, a historical checkpoint plus the signatures backing it, now ships inside the network's configuration file. More people being able to run a node easily is quietly one of the healthier things that can happen to a network.
What is coming next
None of what follows is live. All of it could still change, be renumbered, or be dropped.
The biggest item is not a feature at all. It is the queue.
Bittensor's core software goes out in numbered bundles called releases, and a release only reaches the live network when two separate approvals are given. An automated build system submits the first half. A small group of trusted humans, known as the triumvirate, has to provide the second. It is a two key system, like a safety deposit box that needs both the bank's key and yours, and it exists so that no single compromised machine or person can push code to a chain holding real money.
Release 448 was built and submitted on 21 August and is still unsigned. Since then 450 and 452 have been built and submitted too. Counting 447, which is also outstanding, four separate upgrades are now stacked up behind the same missing signature. That is the story, and the new information is not what any individual release does. It is that the queue keeps growing while the approvals do not arrive.
What this means for you depends entirely on why it is happening, and nobody outside that group can say for certain. It could be caution: a signer looking hard at a bundle before letting it touch a live chain is doing exactly the job the design intends. It could be a scheduling problem, with the right people simply not in the same place at the same time. What it does mean practically is that work you read about as finished may not be running on the network you are actually using. This week's fix for baskets jammed by a dried up subnet is a good example. It is merged. It is not obviously live. If you hit that problem today, the fix existing does not help you.
The risk in a growing queue is that bundles start to depend on each other. Release 451 was built and never signed, so its contents have been renumbered into 452. Release 449 was never shipped, so its contents are moving into 500. Every time that happens, what eventually goes live is a bigger, less rehearsed lump of change than any of the pieces were individually, and big lumps are where surprises live. The thing to watch over the coming week is simply whether the number of outstanding releases goes down.
The second proposal worth understanding properly is a fix to how baskets are paid out, because it is about somebody being able to take your money.
A basket is a pooled holding, and a share in it is priced as though the whole holding were sold at once. But these pools sit on a curved pricing formula, which means selling a small slice fetches a better price per unit than selling everything would. That gap is not a rounding error, it is a door. Someone could deposit into a basket, immediately claim their share back out, and walk away with slightly more than they put in, and the difference comes out of everyone else still in the pool. It is the financial equivalent of a bucket with a slow leak: nobody notices on any single day, and everyone is poorer by the end of the month. The proposed change closes it.
There is a companion proposal that adds automated tests proving subnet owners do not accidentally burn their own Root rewards, and that tidies up leftover machinery from the migration. Also proposed is a fix to the fee quoted when you claim root rewards, which was being calculated as though a claim had to touch more subnets than the network actually permits, so the number shown to you was higher than the real cost.
Three smaller proposals round it out. One creates a narrow permission slip called SetRootWeights, so you can let someone else curate which subnets your root stake backs without handing them a broader set of powers than that one job needs. One raises how long the chain keeps the paperwork behind standing orders, from seven days to a hundred and eighty, because those orders are signed to last a hundred and eighty days and the record proving they exist was expiring first. And one, purely cosmetic but welcome, rounds displayed wallet balances to three decimal places instead of nine, so the screen most holders look at most often stops being a wall of digits. The exact figure is still kept behind the scenes.
The standardised fund price is also on its way in. At the moment different wallets, explorers and apps each work out what a Root fund share is worth using their own arithmetic, so two dashboards can show different numbers for the same holding and both be technically correct. The change records one agreed reference point on the chain itself. Nothing about what anyone owns changes. It is plumbing, and it is the sort of plumbing that stops a reader panicking because their balance appears to have moved when it has not.
A subnet that builds playable 3D worlds from a sentence
The most interesting subnet material this week came from a long call with Ben of the 404 team, which sits on subnet 17 and generates 3D models.
They have changed direction, twice, and the reason is instructive. They started with Gaussian splats, a technique that produces visually stunning results. Studios could not use them. Splats do not slot into the file types and production pipelines that game engines already expect, so however good they looked, they were not usable in real work. So the team switched to meshes, the standard way games represent a 3D object, and built a model called Titan on top of an open source model called Trellis 2. Ben says Titan produces a 3D model in about thirty seconds, against several minutes for most closed source competitors.
On top of Titan sits Atlas, which does the unglamorous work that turns a generated shape into something a game can actually use: simplifying a heavy model down while painting the detail back on as a surface texture, and setting the pivot point so a sword sits in a hand and a helmet sits on a head rather than floating beside it.
The evidence that this is real is a mobile game called Colony, made by Parallel and currently in early play testing, where players describe a weapon or a piece of armour to a character in the game and it is finished by the time they have walked across the map. Ben says Parallel report that in-game crafting in the player's own style is the single most liked feature in testing. He also says two people used the technology to build a game that won the inaugural Chroma Awards, generating over twelve thousand assets. These are his claims, not verified numbers, but they are specific and checkable, which is more than most subnet announcements offer.
Why this matters beyond gaming is that it is one of the clearest examples on Bittensor of a subnet changing its technical direction because customers said no. That is ordinary product discipline, and it is rarer in this ecosystem than it should be.
Turning wasted farm electricity into AI computers
The other substantial interview was with Josh of subnet 110, which is building computing capacity on renewable energy in the UK.
His starting observation is one most people have never heard. The UK grid is not good at carrying electricity back the other way. A large field of solar panels can sit behind a local substation too small to accept what it generates, so only a fraction gets through. When it is too windy, turbines are switched off. Power that was generated is simply dumped. Josh's pitch to farm owners with wind, solar, hydro or anaerobic digestion assets, which is the process of turning animal waste into gas and then electricity, is to put servers on site and sell computing power instead of throwing the electricity away.
The numbers he gives are the interesting part, and they are his figures rather than independently checked ones. He says UK renewable generators currently earn roughly eight to twelve pence per kilowatt selling back to the grid, which gives something like a twelve to fifteen year payback on the original equipment. He says that over the past year, running compute for clients has generated sixty to eighty pence per kilowatt and in some cases a pound. His worked example is a farm that spent five million pounds on wind turbines, then a further million on servers, seeing that fifteen year payback fall to four or five years including the extra cost.
His argument for why this subnet needs to exist alongside the other compute subnets is worth hearing. He says the existing ones are good at cheap, fragmented, short jobs, but a business that wants a hundred computers reliably for six months cannot get that today, because the person renting to you may themselves be renting from someone else on a temporary basis. What he is selling is permanence: machines that will still be there next month. Green energy is currently verified by him driving to the sites and looking, which is unglamorous and, at least at this scale, honest about its own limits.
Serving new AI models on the day they arrive
Two subnets ran the same play this week. GLM-5.3, an AI model whose files were published openly, was made available on subnet 28 by SayGm on launch day, hosted through a partner that runs it inside hardware designed so the operator cannot read what users send. Subnet 53, Engy, is also serving it, at roughly a hundred tokens a second, which is about seventy five words of output per second, and charging ninety eight cents per million words in and $3.08 per million out. The official service and the reseller OpenRouter charge $1.40 and $4.40 for the same thing.
Undercutting the official price by around a third is the most concrete demonstration of what these subnets are for. Whether that price survives contact with sustained demand is the open question, and it is the question worth asking of every cheap inference claim.
Money coming in from the outside
Numinous, on subnet 6, made its first buy-back: it spent 4 TAO purchasing its own alpha token, funded from ten percent of what its paid forecasting service earned across July and August, and published the transaction so anyone can verify it. Four TAO is a trivial amount of money. The mechanism is not trivial, because it is the loop this entire network is trying to close: customers pay for something useful, and some of that money comes back to the people holding the token.
Separately, and treat this as a claim rather than a fact, a researcher's index shared by the ReadyAI team estimates that Bittensor's subnets are collectively bringing in roughly thirty two to thirty five million dollars a year from paying customers, spread across about twenty five of them, with a projected path past a hundred million. It is an outside estimate with a methodology we have not seen. Real outside revenue is exactly the thing critics say this network lacks, which is precisely why a number this convenient deserves to be treated carefully rather than repeated.
Smaller things worth knowing
Subnet 3, Teutonic, has begun training a model with about 110 billion parameters, the internal settings a model adjusts as it learns and a rough proxy for its size, on volunteered hardware rather than in one company's data centre. The team calls it rough. Training a model that size across strangers' computers is the hard version of the problem and worth watching for whether it finishes rather than for what it announces.
Subnet 106, Nodexo, added twenty two H100 graphics cards, the expensive Nvidia chips most serious AI work runs on, with fourteen available to rent immediately. Subnet 113's dollar-pegged token now appears automatically in SubWallet, removing the step where a newcomer has to paste a long technical address by hand. Subnet 84, Tatsu, built a prepaid credit system covering its AI products, with the team noting that taking the money is easy and the hard part is metering so nobody quietly uses more computing than they paid for. Subnet 114, SOMA, which builds matching algorithms, says another Bittensor project has started using its algorithm inside its own product, which is a small but real sign that the output is good enough for someone else to depend on. Subnet 104, Masx.ai, has named Monday evening as the date for its version three beta, which is a first firm date rather than a working product. And subnet 23, Trishool, says it has been added back into the Arbos root basket after being dropped over a misread of its burn policy.