Quite a lot was written, proposed and announced on Bittensor today, and almost none of it was switched on. That gap, between what has been built and what is actually running, is the clearest thing about the network right now, so it is where we will start.
Nothing new switched on
Bittensor's core software, the program every machine on the network runs, is updated in numbered releases. Getting one live takes 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. Only then does an upgrade reach the machines that make up the network.
Version 452 was built and given its first signature on 29 August. It joins versions 447, 448 and 450, none of which has the second signature. A release branch for 451 has also been opened, and a separate proposal bumps the version number to match it. So the waiting room now holds four finished upgrades and the beginnings of a fifth, and nothing has come out of it.
This is worth understanding because it changes how you should read every other piece of news about the code. Work that gets merged into Bittensor's codebase does not affect anyone until a signed release carries it out to the network. One of the waiting items, for example, extends how long the chain remembers a linked order from seven days to a hundred and eighty days, which matters because the trading tools built on top sign orders that are meant to last six months. Until a release ships, the mismatch stays. The same applies to the fee corrections and safety fixes we will come to in a moment.
There is an innocent reading and a less comfortable one, and today we cannot tell you which is true. The innocent reading is that batching several releases together and signing them at once is normal caution. The less comfortable one is that the pace at which Bittensor can deliver anything is set by the availability of a very small number of people.
Which leads neatly to the one thing that genuinely did land. A documentation page was merged into the core code that states plainly, in one place for the first time, how the chain itself is run. Bittensor uses Proof of Authority. That means a known, permissioned set of machines takes turns producing the blocks that record what happened and confirming them, rather than anyone with enough stake being able to do that job. Think of it as a bus route run by a fixed roster of approved drivers, rather than one where anybody who puts down a deposit and gets vouched for can take the wheel.
The same page sets out a non-binding aim to move to Nominated Proof of Stake, the second version of that picture, within about a year. Writing something down is not the same as changing it, and non-binding means exactly what it says. But you cannot argue about an arrangement nobody has described, and now it is described. For an ordinary holder the question this raises is simple: how much of the network's day to day operation depends on a small named group, and what is the actual plan and timetable for widening that. The document opens the question rather than answering it.
One more thing is live, from slightly earlier, and it matters more than most readers realise. On a community call on 28 August, Bittensor's founder Const said that Root Reborn and its curated staking baskets had gone live within the previous two hours. Root staking is when you put TAO into the network's shared central pool rather than backing a single project. The yield from that pool gets sold into subnets, the independent projects that make up the network. Until now, deciding where that went meant forming a view on more than a hundred teams. Now each validator, a machine that scores work and votes on who gets paid, can choose the set of subnets it sells that yield into, and that selection is called a basket. You can subscribe to a validator's basket instead of picking yourself, which is roughly like choosing a fund manager rather than choosing shares.
Const was careful about this and so are we: it is not risk free, because you still end up holding subnet tokens that can go down as well as up. What it removes is the research burden, not the risk. He also said the way yield is calculated has been normalised so it works the same way everywhere, and that basket curation tools which teams like Crucible and Mentat Minds had built for themselves now exist at the protocol level in a standard format. He mentioned he had already set weights with a curator called Arbos, and that baskets are not tokens, though they can be made to behave a little like them.
The teeth on this showed up in today's news. Subnet 23, Trishool, said 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 token issuance is destroyed rather than paid to anyone, and a curator read that as a sign of the wrong incentives. Trishool's account is that the burn was misread. Whoever is right, the episode tells you something real: basket curators now have the power to move where a large pool of stake points, and being in or out of a popular basket is going to matter to subnets very quickly.
What is coming next
The biggest proposal of the day, and the one an ordinary holder should understand before it arrives rather than after, is pull request 3135. It would add a built in way to bet for or against a subnet's token for up to thirty days.
Start with what the token is. Every subnet on Bittensor has its own token, called alpha. It is how the network prices that project and how the people working on it get paid. Today, if you have a view about a subnet, you have exactly two moves: buy the token, or do not own it. There is no way to back a negative view, and no way to take a position that ends on a date.
The proposal adds both. You could open a long, which gains if that subnet's token rises, or a short, which gains if it falls, and every position expires after thirty days.
The clever part, and the part worth explaining, is where the money comes from. Each subnet has a pool holding TAO on one side and its own alpha on the other, and the ratio between those two piles is what sets the price. Picture a currency booth with two stacks of cash, where the relative height of the stacks is the exchange rate. When someone opens a position under this proposal, a slice of both stacks, the same proportion of each, is lifted out together. Because both sides shrink equally, the ratio does not move, so the price does not move. Half of what was lifted is then traded through the ordinary swap, and when the position closes the trade is reversed and the pool gets its own liquidity back. Nothing is created and nothing is destroyed. It is a loan from the pool to the trader, repaid at the end.
Who does this help. Anyone who thinks a subnet is overvalued currently has only one honest option, which is to own none of it, and silence is a very weak signal. Letting a sceptical view be expressed with money behind it is how prices in most markets end up meaning something. It also helps a believer take a dated position without committing to hold forever, and it gives someone already holding a way to protect against a fall without selling.
Who it could hurt is just as real. Subnet pools are small. Lifting a slice out and keeping it out for up to thirty days means the pool has less depth while positions are open, and a shallower pool means ordinary buyers and sellers see the price move further against them on the same size of trade. On a small subnet, concentrated betting against it could become self fulfilling: the price falls, confidence falls, the team finds it harder to attract the people who do the actual work, and the fall justifies itself. There is also a softer risk, which is attention. Several subnets have spent this month trying to shift the conversation from their token price to their revenue, and a market in dated bets on token prices pulls hard in the other direction. And there is the plain engineering risk of new code that reaches into the pools where the money sits.
None of this is live. It is a proposal, tagged for a future version of the runtime, and it can be changed, delayed or dropped entirely. But it is the sort of change that alters the texture of the whole network, so it is better met early than late.
A second idea, softer and further from shipping, is being discussed in the community under the name fractal emissions. Emissions are the new TAO the network creates every few seconds and hands out to the projects running on it, split between the people doing the work and the people backing them. New TAO paid to a subnet is shared with holders of that subnet's own token, and creating new tokens dilutes the ones that already exist, in the same way a company issuing new shares makes each existing share a slightly smaller slice. The idea is to let each subnet voluntarily accept a fraction of what it is entitled to: all of it, half, a quarter, an eighth, or less. Taking less would mean less new supply and less dilution.
It is an appealing idea with an obvious trap. Emissions are not decoration, they are the wage bill. A subnet that halves its intake is halving what it can pay the people supplying computers and models, and if that pay drops below what the work costs, the work leaves. There is also a risk that cutting emissions becomes a marketing gesture, something a team announces to look disciplined while quietly losing contributors. It is a discussion, not code, and there is nothing to react to yet beyond keeping the trade-off in mind.
Finally, three unglamorous proposals that decide whether money people are already owed actually arrives. Bittensor lets you give another account a permission slip, called a proxy, so it can act for you without holding your main key. One of those permission sets is the staking proxy, which most people would reasonably assume covers everything to do with staking. It does not. Root rewards, the payouts owed to TAO staked in the shared central pool, only materialise when someone runs a specific claim instruction, and that instruction sits in a category the staking proxy cannot reach. So a setup that looks complete silently stops collecting root rewards. A proposal would fix that by letting the staking proxy make the claim. A second adds a new narrow permission called SetRootWeights, which allows one account to curate which subnets your root stake backs and nothing else, so delegating basket curation no longer means handing over broader powers than the job needs. A third lowers a fee that was being quoted too high for claiming root rewards, because it was calculated as though the network had more subnets than it actually permits. All three are proposed and none is live, which means the silent gap in the staking proxy is still there today.
Money coming in from the outside
The most quoted criticism of Bittensor is that it pays people with newly created tokens rather than with money from customers. Several stories today push back on that, with very different amounts of evidence behind them.
Bitcast, subnet 93, pays creators to promote brands on YouTube and X and charges the brands for it. It says it spent seventy five thousand dollars of that customer revenue buying its own token back off the open market during August, its biggest month so far and more than double what it spent in July. A buyback means taking real income and using it to purchase the token the network runs on, which takes some supply out of circulation and, unlike emissions, does not create anything new. Separately, a well followed community account has put revenue figures around it, claiming August was Bitcast's best earning month and that income has grown four point seven times since February, roughly thirty percent a month. That second part is one person's claim and the team has not published those numbers, so treat the buyback as reported by the team and the growth rate as unconfirmed.
Numinous, subnet 6, made its first buyback and it is tiny: four TAO. What makes it worth mentioning is not the size but the rule behind it. The team says it is ten percent of what its Forecasting API, a paid service other programs call to get predictions, took in across July and August, and it published the transaction so anyone can check. A small verifiable number attached to a stated policy is more useful than a large one nobody can audit.
Widening out, a researcher's index shared on Friday by the ReadyAI team estimates that Bittensor 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, and projects a path beyond a hundred million. That is an outside estimate rather than audited accounts, and the projection is the least reliable part of it. Even so, it is the first number of that shape anyone has put on the table, and it is the number the network's critics and defenders will now argue about.
Elsewhere on the commercial side, Tatsu, subnet 84, spent the week building a single prepaid credit that covers training a model, running one and using its assistant, payable by card or with tokens. The team's own framing is the honest one: taking the money is easy, and the hard part is metering, making sure nobody can quietly consume more computing power than they paid for. And Tao.com, a trading site for TAO and subnet tokens, says the automatic sell instructions it added recently, which sell if a price reaches a chosen level up or down, have been its most adopted feature this quarter. That is the company describing its own product, and we are reporting it as that.
Subnets shipping, and one that lost its seat
Subnet 111, run by DeSci Claims, says its main network launch is done and emissions have been switched on. That switch is the moment a subnet stops being an experiment and starts being paid, so it is the milestone worth marking even before there is a product to look at.
Subnet 65, run by TPN Labs, turned on three things two weeks after reaching the main network: a public benchmark, a public dashboard and a gateway. The subnet pays people to shrink large AI models so they run faster and more cheaply, and seventeen participants have squeezed a model called Qwen3.5-2B down into a two gigabyte limit. Shrinking a model is like compressing a film so it fits on a phone without ruining the picture, and the interesting part here is that the judging of who did it best used to happen out of sight and now happens on a page anyone can read.
A few teams spent the week lowering their front door. Enigma, subnet 63, works on quantum computing problems and used to require anyone taking part to run miner software, a program that must stay connected and needs a crypto wallet configured before it does anything. Submissions can now be made by uploading a file through a web page, which matters because the people most likely to have good answers, researchers, are the least likely to want to run infrastructure. TensorUSD, subnet 113, says its dollar pegged token now appears automatically in the SubWallet app instead of requiring someone to paste in a long technical address, a step that quietly loses newcomers. And Yanez, subnet 54, which works on proving who someone is without collecting their personal details, says its login has started rolling out inside two other products, reporting that only two percent of people signing up in the early stages needed to contact support. That figure is the team's own.
On the raw computing side, Nodexo, subnet 106, added twenty two H100 graphics cards and says fourteen are available to rent immediately with no waiting list. The H100 is the expensive Nvidia chip most serious AI work runs on, and getting hold of them is the main bottleneck for anyone building AI outside the biggest companies. Beam, subnet 105, is connecting to Hugging Face, the site where most freely available AI models and datasets live. Beam is not offering to store them, it is offering to move them, shifting very large files across many machines at once rather than down a single connection. Cacheon, subnet 29, reported a new speed record in its own model contest, 2,274 tokens per second on an unmodified starting model, where a token is a small chunk of text so the number is simply how fast the model writes. Targon, subnet 91, gave free computing credits to Cascade, the first team out of a training programme for new subnet teams. Open Quantum, subnet 48, rewrote its whitepaper after nine months of changes, which is housekeeping rather than news. And SOMA, subnet 114, says an early access product is coming with no date and no description, which is an intention rather than a launch.
Then the other direction. Babelbit, subnet 59, has been deregistered, meaning it lost its slot on the network and no longer receives any of the newly issued TAO. It happened roughly a week after the team announced a dubbing service for translating audio into other languages. Slots on Bittensor are finite and competitive, and a subnet that cannot hold enough backing loses its place to one that can. The lesson is uncomfortable and worth sitting with: shipping a product does not protect you here if the support behind your token thins out.
Related, Zeus, subnet 18, explained a run of selling in its token as one early partner winding down. That partner had paid the fee for Zeus to get its slot and received a share of the newly created tokens the subnet earns each day in return, an arrangement that is common and rarely spelled out to holders. Zeus says the selling connected to it is finished and that it did not agree with the decision. That is the team's account of its own token, and there is no way to verify it from outside.
Finally, a group called Kusanagi has set itself up as a business layer around Bittensor rather than registering a subnet. It runs a validator, rebranded from an existing one, keeps a token trading site as its shop window, and offers services on top. It is a sign of a second ring forming around the network, of businesses that make money from Bittensor without competing for a slot on it.
What the builders said this week
Two long conversations added more than the posts did.
Ben from the 404 team, subnet 17, which turns typed descriptions into 3D objects, explained why they abandoned an approach called Gaussian splats. Splats looked visually stunning, he said, but studios could not actually use them, because they do not fit the file types and production systems game engines already expect. It was a beautiful photograph when the factory needed a blueprint. So the subnet built a model called Titan that produces meshes, the format games actually run on, on top of an open source model called Trellis 2. He says Titan makes a model in about thirty seconds where closed competitors take several minutes. Because the competition rules require outputs that can be used commercially, miners stripped out parts that blocked that and rebuilt them, fixing holes and geometry errors along the way, and added things like normal maps, a trick that paints the appearance of surface detail onto a flat surface so an object looks carved without being heavier for a phone to draw.
On top of that sits Atlas, which packages the raw output into something a game can drop in: simplifying a heavy model while keeping its detail, and setting the pivot point so a weapon sits in the hand and a helmet on the head. A mobile game called Colony, made by Parallel, is using this in early play testing to let players generate their own weapons and armour mid game, with the waiting time hidden behind the walk across the map. Ben says Parallel report that crafting in the player's own style is the single most liked feature in testing. He also says two people used the technology to win the first Chroma Awards with a game containing over twelve thousand generated assets. Those are his claims, but they describe the thing the network is most often accused of lacking, which is somebody outside paying attention.
The second conversation was with Josh from subnet 110, which is trying to build computing capacity on renewable energy in the UK. His argument starts with a problem most people have never heard of. The UK grid is bad at carrying power backwards. A large solar field can sit behind a local substation too small to accept what it makes, and wind turbines are switched off when it is too windy, so the power is simply dumped. His pitch to farmers with turbines, solar or anaerobic digesters is to put small data centres on site and sell computing instead of electricity. His figures, and they are his: selling to the grid earns roughly eight to twelve pence per kilowatt, while over the past year running compute has earned sixty to eighty pence and in some cases a pound. In his worked example, a client who spent five million pounds on turbines and a further million on servers could see a fifteen year payback fall to four or five years.
He also drew a distinction worth borrowing. Most compute you rent in this world is rented from someone who rented it from someone else, so a business that needs a hundred machines reliably for six months cannot get them. What he calls permanent compute means hardware you can count on being there next month, not a pool that may have evaporated. The catch he was open about is verification: the green energy claim is currently checked by him visiting sites in person, which is workable at today's size and is the first thing to watch if this grows.