Hidden Meta, Revisited

I published a piece on December 1, 2022, pulling apart how Bittensor pays people to improve AI models.

ChatGPT had shipped the day before.

I wasn't reacting to it. In hindsight that's the more interesting fact – the question of how you pay for intelligence was already what I was chasing, and the launch that made everyone else care hadn't happened yet.

That piece is still up, unedited, at Hidden Meta: Quantifiable Real-World Value Accrual.

Three and a half years is long enough to grade it.

The 2022 Claim

Three claims, written out plainly at the time. Incentivize model improvement, at reduced cost, at increased velocity.

Contributors improve a model. The contributed intelligence gets measured by a loss function. They get paid in TAO.

My conclusion then was that velocity was proven and cost was close to zero.

The line I'd most like back: “Reduced cost? Pretty close to $0 since Bittensor doesn't pay anything to generate $TAO.” That's wrong. It took a halving to make it obvious.

The Cost of Issuing a Token

Why not? Issuing a token isn't free.

The cost is dilution. Everyone already holding pays it, which makes it easy to miss, because it never shows up as an expense line.

The first halving, on December 14, 2025, cut daily issuance from roughly 7,200 TAO to about 3,600. That's when the cost started to show up.

Cut the subsidy in half while more subnets keep showing up to compete for it, and there's a lot less to go around for each one.

Incentive designs look elegant while the reward pool grows. You learn what they're made of when it stops.

Velocity and Compensation

I claimed tokens solve a real problem for early organizations. Cash requires someone to front the risk. Equity stays illiquid for years.

That's aged well.

Bittensor now splits rewards 41% to validators, 41% to miners and 18% to the subnet owner, paid automatically in that subnet's own token. No approvers or distributors in the path.

Plenty of organizations have copied the structure since, without ever touching a blockchain.

Measuring Quality

I wrote that it was difficult for me to validate the quality of the accrued intelligence, and that I was reasoning past that limit.

That turned out to be the most important line in the piece.

In February 2025, dTAO gave every subnet its own token and liquidity pool. As of June 2026, each subnet's share of emissions is proportional to its moving-average token price.

So the network never answered the quality question. It handed the decision to the market instead. Whatever a subnet's token price says, that's where the emissions go.

Reflexivity

A subnet with a good narrative and a promotional owner attracts TAO. That lifts its alpha price, which raises emissions, which funds more promotion.

Every step in that loop works the same way whether the output is excellent or worthless.

The mechanism is honest about what it measures, which is attention. People keep reading the emissions leaderboard as though it measures quality.

Attention and quality are two different things.

Root Reborn

The upgrade shipped over the summer. Validators have been migrating since, with xTAO completing its own upgrade on August 17.

Root dividends used to leave the system. They got claimed and sold for TAO on a schedule.

Now each validator directs those dividends into a curated basket of subnet alpha that compounds instead. Estimates put the reduction in automatic annual sell pressure at up to a third.

At least one validator has publicly argued the design carries substantial risks – and that seems like the right instinct to me.

It asks validators to pick winners. Running a validator well and picking winners well are completely different skills.

The ETF Filings

Grayscale filed an S-1 on December 30, 2025 to convert its Bittensor Trust into a spot ETF on NYSE Arca, ticker GTAO.

Bitwise filed an N-1A for a TAO strategy fund that would hold roughly 60% TAO directly, with the balance in other exchange-traded products and derivatives.

The SEC decision window has been tracked for this month.

I want to be precise here, because this gets blurred constantly. A filed registration statement is still a filing, and a trust is still a trust. As I write this, nothing here trades as an ETF.

Production Cost and the Alpha Tokens

The stronger half of the 2022 piece argued that protocols accruing real value with a known production cost get a price floor. The way mining costs support gold.

I still think that's right about TAO. GPU time and developer talent are real costs.

It doesn't transfer to the alpha tokens.

An alpha token's price is set by how much TAO flows in. TAO inflow has no production cost attached to it at all.

So the floor I described is real, and it sits under TAO. It doesn't sit under the alpha tokens, which is where most of the money has ended up.

Wrap-up

The open question I left in 2022 was who ultimately buys the token from the developers being paid in it. For three years the answer was other crypto participants.

An approved ETF would be the first structurally different buyer. It would also be the first buyer with no view whatsoever on whether the intelligence being accrued is any good.

That's either the validation of the thesis or a complete restatement of the problem. I don't know which.

None of this requires tokens, by the way.

If you want to know which of your AI workloads deserves funding, you need something that produces a number continuously. A quarterly review produces an opinion. Internal compute markets produce the number, and they dodge the reflexivity, because the bidders are spending their own budget.

Almost nobody runs one.

Grading your own old work is uncomfortable and worth doing. I was early, and I was partly wrong. The specific way I was wrong is more useful to me now than the part I got right.