Changing the Table Mid-Hand.

A business is dying in public on Bittensor right now, and it shouldn't be.

Babelbit runs subnet 59: low-latency speech-to-speech translation. The team isn't three anons and a Discord. Matthew Karas has spent 25 years in speech technology and was launch CTO for ITV Player; his co-founder Tom Horner built systems that were not allowed to fail, from flight simulators to trading floors at Citibank and Millennium. I know them both.

Last week Matthew posted a demo of their new real-time dubbing model, built, in his words, "in my free time between worrying about dereg and other Bittensor complexities". Tom posted a farewell in advance: short of a miracle they would be the next to go, after "going from standing start to pushing the frontier of tech in this space and out performing the solutions by Google and ElevenLabs in just a handful of months."

As I write, Babelbit sits first in the pruning queue, waiting for the axe to be swung. It is not being removed for failing. Bittensor's deregistration mechanism doesn't measure failure, only price.

TAO trades at about $190 today (4th August 2026), roughly two thirds off its high. Meanwhile Dropbox is piloting one subnet's product, Intel has co-authored a paper with another, OpenAI with a third.

Why does the market refuse to believe? I've spent 18 months invested across this network, and I think I know: Nobody can tell you what the rules will be next Tuesday.

In April, Sam Dare stomped out of Bittensor calling its governance "decentralisation theatre". As a description of Bittensor in April, he was wrong. As a prediction, he has been made right, not by anything he did, but by the reaction. In the four months since his exit, the network's leadership has demonstrated exactly the concentrated discretion he accused it of.

Sam Dare wasn't a whistleblower. He was an accidental prophet, and the network spent the summer fulfilling him.

Look at the official release index. July 2026 alone carries six named upgrades. Six. In a month.

The emission distribution mechanism, the thing that decides who gets paid on this network, was rewritten three times in eight weeks: net TAO flow in late May, price-plus-burn in late June, and the Emission Gate in late July, under which a subnet below roughly rank 32 sees its passive yield fall, in OTF's own words, "to approximately nothing".

It isn't just the pace. It's the notice. In October 2025 the foundation tweeted advance warning of a burn-cost change "to ensure no community member is caught off guard". The June 2026 change that zeroed emissions to 57 subnets shipped the same day it was announced. Yuma put it in the politest possible corporate English: the ecosystem's "lack of process... leaves business builders limited notice or ability to properly plan, assess risk, and execute."

And then there is the confession. The Emission Gate's own release notes describe the old registration economics as "one of the network's quietest failures", the new doctrine: "a slot should cost nothing."

Fine. Good, even. Except that in the first half of this year, 28 teams paid nearly 20,000 TAO combined - about $5M - to register under those economics. The network sold tickets at boom prices. Then it announced they had always been mispriced. No refunds.

Emissions were never meant to be a dividend. They were the funding model. A subnet sells alpha to cover salaries and pay GPU bills. Selling alpha was never a betrayal of the network. It was the point of it.

Then Covenant left, and the remedy was conviction: lock your alpha in public, prove you are not the next founder out of the door. The worry behind it was real enough, but the fix asks a builder to prove his commitment by destroying his ability to leave, using the money he was supposed to be running his company on.

Plenty of teams complied. Babelbit's did.

Eighteen days after conviction went fully live, the Emission Gate landed and moved the money to the top 32.

You can argue the gate on its merits. You cannot argue its timing to a team that threw away the key eighteen days earlier because the network encouraged them to. They were asked for commitment, they gave it, and then the thing they had committed to was quietly repriced. Everyone below the line is now holding a promise they cannot unmake, on a token they cannot sell, inside a business the network has stopped funding.

There are 128 subnet slots. When someone registers a new subnet, the lowest-ranked existing one without immunity is deregistered, its slot reissued, its accumulated position erased. And the ranking is not revenue, not users, not shipped code. It is the exponential moving average of the token price. Capital punishment administered by a column on Taostats.

The roll-call from this year makes the point better than any argument. Vericore: fully doxxed team, good subnet, gone. Bitrecs: much the same, gone. Bitads: had revenue, gone. Leoma and Vocence: both good teams, both deregistered, both now back on the network. Djinn: real revenue, a fully doxxed team, deregistered anyway.

The facts don't translate well for Babelbit.

I should be clear about my own interests here: I'm not an investor in Babelbit, but I was in Djinn, and I am in other subnets that sit in that same pruning queue. I am not writing this from the press box.

I'm not crying over any individual project. Some deregistrations are deserved; a mechanism that never deleted anything would fill up with squatters. The point is systemic. These were people adding value to the network, or at the very least genuinely trying to, and the network took the value away. Walking over to the people who turned up to play, taking their football away, and telling them they're not wanted any more cannot be the best thing for the network. It certainly isn't the best advert for joining it.

Six months ago this community felt like one team. A hundred-odd startups, all pulling the same direction, all lifted by the same tide. The Emission Gate ended that, not accidentally but by construction: with a fixed pool and a cliff at rank 32, one subnet's rise is now, mechanically, another subnet's dereg risk.

The community noticed immediately. "The 440 changes disincentivize collaborating." The foundation's own release notes do the arithmetic that explains why: a subnet at rank 36 that grows its demand by 10% gains roughly 26% more emission, so growth near the cliff is the best-paid move on the network, which is another way of saying that the most profitable thing you can do is take someone else's place.

Public criticism of Const is rare, and I don't think that's contentment. Subnet owners say things to me privately that they will not post publicly, and I understand why. When one man effectively controls your emissions, criticism of that man starts to look less like discourse and more like a resignation letter.

Let's start with what Const's defenders say, because it's more revealing than anything his critics manage. "Bittensor does not deserve decentralisation yet," runs one summary of his position. "Const is forging the machine until it works, then throwing the keys away for good." The defence concedes the whole case. One man holds the keys, decides when the network has earned its own founding premise, and reforges the machine weekly until then.

He does not dispute it. In April he wrote, "I do not have the ability to suspend emissions... I don't have any privilege beyond what normal TAO holders have." In June an update shipped that zeroed emissions to 57 subnets, with a weekly eligibility review, and his roadmap the day before conceded that Bittensor "at its core economic incentive layer, is currently not decentralized". The privilege was not written inside the rules, it was holding the pen that wrote them.

The keys are sometimes literal. Subnet 24, Quasar, is run by a couple of teenagers Const rated so highly he bought them out of their original investment agreement with me and Siam, bailed them out of two sizeable hacks, then took personal custody of their owner key. Plenty of alpha investors held that token precisely because he held the key. DSV Fund signed in July to invest $280,000 so the team could pay for their training run.

How lightly that key is held became clear last week. Const said he didn't have the cold key and then two days later he corrected himself.

The Quasar alpha didn't drop because Const didn't hold the keys, but it did shoot up when he confirmed that he did, because alpha investors know his involvement or blessing will send a project to the moon even if it later turns out the whole production was shot in a TV studio.

The market is not pricing the work. It is pricing the man who still holds the keys, perhaps more for his influence than his judgement.

Meanwhile at the grey-haired end of the network, Philip Maymin and Iosif Gershteyn at Djinn have an embarrassing wealth of experience and expertise. Their subnet got the chop anyway. Matthew, Tom and Babelbit are next in line, nervously watching the executioner resharpen his guillotine. Matthew's verdict on the machine about to delete him carries no bitterness at all. "I have a fear that at the moment it is not rewarding subnets in proportion to technical innovation," he wrote this week.

The network is losing another builder.

On this network, a founder's blessing beats revenue and a lifetime of shipped work. That is not effective incentive design.

The deepest confession of the summer came from the same place as the rest: the foundation's own release notes. Root Reborn exists to fix a mechanism under which root stakers' dividends were auto-sold, creating, in its words, "a drain equal to 74% of the roughly 1,321 TAO of fresh emission entering all subnet pools". The network's core plumbing was dumping three quarters of the fresh capital entering its subnet economies. Every day. By design.

To be fair, and this matters later: v441 is also the fix. The confession and the correction arrived in the same document.

In April, vaN ττ published "The Taoflow Angel", a warm and honest account of Bittensor as it then stood: angel investing made accessible, investors "never structurally trapped", survival for "the ones that attract capital and deliver results", and a warning that concentrated betting was "gambling with extra steps". He wrote truthfully about the rules as they were. Within four months the network falsified all of it. On 29 July a tracker noticed that the bar deciding whether a subnet earns anything at all had moved from 61% to 75% in 48 hours. No runtime upgrade, no release notes, no announcement. The gate is written into the code where everyone can read it. Where the bar sits is a setting, and settings have owners. His prospectus wasn't dishonest. It just painted a picture of an investment landscape that no longer exists.

DSV has paid the first tranche of our Quasar commitment and is holding the rest, because we cannot underwrite the ground the team is standing on. My partner Siam Kidd put it bluntly: "right now, dTAO is basically uninvestable. Goal posts moving so often based on various OG spontaneous whims make this just gambling, not investing now... my simple wish is that dTAO just stayed in one state/form for 6 months. Just 6 months." Every institutional conversation he has ends the same way: looks cool, come back when it's more mature.

It isn't only the capital that's thinking twice about Bittensor. I know a serious AI project, strong team, currently raising from VCs, that was going to use the money to take a subnet slot and build inside Bittensor. I spoke to them a few days ago. The project is going ahead. The raise is going ahead. Bittensor has been cut, not because they doubt the technology, but because it had become a variable they cannot control.

New registrations haven't stopped, of course, and slots are cheaper than they've been all year. Every one of them swings the axe on a team below the line. The ones pausing are the teams with something to lose.

Warren Buffett said that risk comes from not knowing what you're doing. On Bittensor right now nobody can really know what they are doing, because that requires the rules to hold still long enough to learn them. A casino at least keeps the same rules all night. Bittensor changes the table mid-hand and calls it an upgrade.

First, Bittensor is the largest collection of independent, incentive-funded AI startups anywhere. In July OpenAI published a field report on scientific computing with agentic AI. One of its case studies is a genome-generation engine built by a Bittensor subnet, MinosAI, listed in the author affiliations alongside NVIDIA. Lium rents GPUs to paying customers at a rate its own reporting puts in the millions annually. Macrocosmos has pretrained a 100-billion-parameter model across single GPUs scattered around the open internet. None of that is vapour and the market simply isn't pricing it properly.

Second, the changes I've spent all this time criticising are mostly aimed at real problems. Slots were mispriced. Emissions were being mechanically dumped. Grifters did farm passive yield, and deregistration has to exist, or the network fills up with them. My objection was never that Bittensor deletes things. It is what it selects on, and how little warning it gives.

Const's own doctrine is that if the game is toxic, the fault is the mechanism designer's and nobody else's. By that standard he has spent the summer blaming himself in public, in code, at a pace no committee could match. The very discretion that terrifies me is also why the 74% drain was fixed in weeks rather than years. His clock runs in decades. Mine, invested alongside, can afford to run in years.

So here is the ask, and it is about process, not power. None of it is original to me. Yuma put it to the ecosystem in June and Siam echoed it in his update yesterday.

Give notice before changes. Set a cadence businesses and investors can plan around. Give the proposal to widen the gate from 32 to 64 a proper hearing, because the pendulum that needed to swing at the extractors has swung into teams doing real work. And then hold still for six months. Six. Long enough for builders to build on the rules and for allocators to underwrite them.

That is not a revolution. It is a pause between revolutions, and it would do more for the price of TAO than another six upgrades.

Because the thing about everyone agreeing that Bittensor is uninvestable is that everyone agreeing is precisely when it stops being true. Buffett, in 1979, when nobody would touch equities either: "The future is never clear; you pay a very high price in the stock market for a cheery consensus. Uncertainty actually is the friend of the buyer of long-term values."

There is no cheery consensus on Bittensor right now. Which is, if you can hold your nerve, exactly the point.