Hello,

Picture a mining town in the nineteenth century. The company pays you in paper it printed itself, and that paper works in exactly one place: the company store.

Now picture a network that pays the people who run its machines the same way. Every few minutes it prints new tokens and hands them over. A token cannot pay an electricity bill, so the owner sells it, and every sale pushes the token's price down a little.

That is how the cheapest AI in the open world gets sold for less than it costs to make. The customer pays a small part. Everyone holding the token pays the rest, quietly, in the price, and most of them never asked it a single question.

Then a proposal landed. Stop paying the bills in something anyone can sell. Pay them in a credit that can only be spent inside the network. I catch myself thinking it is clever, and then I remember what the company town learned about paper that cannot leave: it keeps the store busy, and it never brings in a customer.

So here is the question I keep: when a network only trades with itself, is that activity, or is it demand?

Listen:

Apple Podcasts: https://podcasts.apple.com/kg/podcast/plain-strata/id6783455764?i=1000793882995

YouTube: https://youtu.be/GPM7ui1zNDs

The full piece, no need to click through:

Twenty-seven cents buys you a million words of input to a large AI model on the cheapest serving market in the open world. The same million words cost sixty cents from a well-known American provider. Last week somebody published the division nobody had bothered to do: at the volumes that cheap market actually serves, those million words cost about a dollar forty-one to produce. The customer paid twenty-seven cents. The other dollar and fourteen cents was paid by people who have never sent it a single request.

They paid it without writing a cheque. They paid it by holding a token that the network prints every few minutes and hands to whoever is doing the work, and that the worker then sells to cover an electricity bill. The selling is the payment. Nobody signs anything.

That is the story this week, and it has two halves that arrived four days apart. On 3 October the arithmetic of the subsidy was laid out in public with real numbers. On 28 September, at a developer summit in Montreal, the foundation that builds Bittensor, the largest of these networks, proposed the thing that would change it: a credit that sub-networks mint by destroying their own tokens, and that can only be spent buying services from each other. One half is the bill. The other half is an attempt to pay it internally. The spine of the week is that a network discovering it is subsidising its own customers has exactly two moves, raise prices or stop needing outside money, and it just proposed the second one.

Three things, in one breath each.

There is a ledger, which is a shared spreadsheet nobody is allowed to lie to. It does not hold AI models. It holds records of who did work and who got paid.

Around that ledger sit sub-networks, each one a crowd of machines in different buildings owned by different people, all running the same kind of job: serving a model, renting out graphics cards, storing files. Each sub-network has its own token, called an alpha token, and the price of that token is set by a pool anyone can trade against.

And the ledger emits. Emission comes from the Latin emittere, ex plus mittere, to send out: the chain sends out new tokens every few minutes, and scores submitted by stakeholders decide which sub-network they land in. That is where a worker's pay comes from. It is not revenue. It is new supply.

Now the whole thing is visible. A machine owner receives new tokens, sells them into the pool to pay for power and hardware, and the price of the token walks down a little with each sale. The customer's twenty-seven cents covers a fraction of the real cost. Every token holder covers the rest, quietly, in the price.

The analysis making the rounds this weekend comes from Pine Analytics and concerns the network's largest serving sub-network, called Chutes. In the window they studied it was taking a little over fourteen percent of all daily emission, roughly five hundred and eighteen tokens a day, which at the price then current was something near fifty-two million dollars a year. Customer revenue over the same period: between one and a third and two and a half million dollars. The ratio of subsidy to sales lands somewhere between twenty-two to one and forty to one. Their conclusion is the sentence worth carrying: this is a transfer payment from token holders to inference users, not a business model.

Hold that number loosely and the direction firmly. The emission share is from March, and by September the same sub-network was down to under six percent, because the network changed how emission is allocated in the meantime. The revenue figure is an estimate by an outside tracker, not an audited account. And the sub-network's own published throughput, one hundred and sixty billion tokens a day, is about twenty-four times what an independent routing service measured passing through it. Three of those four numbers are contested. What is not contested is the shape: the price charged is below the cost of production, and emission is covering the gap.

The honest comparison is the rental sub-network called Lium, which rents graphics cards rather than serving models. On its own reported figures it takes in about one and a half million dollars a month and receives an annual subsidy somewhere between eighteen and twenty-five million. That is a ratio closer to four to one. The line in the analysis is the fairest thing written about this all week: it is the closest any sub-network has come to paying for its own emission, and it is still not paying for it.

Zoom out and do one piece of arithmetic on published figures, carefully labelled as arithmetic. The whole network prints about three thousand six hundred tokens a day since its first halving last December. Two dozen sub-networks collect somewhere between twenty-eight and thirty-five million dollars a year from outside customers, including some real ones, a large accountancy's French arm, a file-storage company, a listed property trust. At the token price used in that same analysis, the printing is on the order of ten times the collecting. Not a verdict, an order of magnitude, and nobody outside these ecosystems has audited the books.

At the Exploit Summit in Montreal on 28 September, the foundation behind the network put forward a design called Gamma. The naming walks the Greek alphabet: the root token first, then alpha for each sub-network's own token, and now gamma for a third kind of thing that is not really a token at all.

Here is the mechanism, and it is simple enough to hold in one hand. A sub-network takes some of its own alpha tokens and destroys them. In return the chain mints Gamma, priced against a basket pegged to stable value, so roughly five dollars of destroyed alpha becomes five Gamma, and those five Gamma stay five forever. Gamma does not float. It also does not go backwards: there is no path from Gamma to alpha, and the only kind of address that can accept Gamma in payment is a sub-network owner's key. You cannot cash it out. You cannot sell it to a stranger. You can only spend it with another sub-network, on inference, on compute, on storage.

Seven infrastructure sub-networks are named as the first issuers. It is slated for early next year, the mechanics are not finalised, and the governance of the credit itself is openly unresolved.

Read what that does to the flow. Today a serving sub-network that needs rented graphics cards has two ways to get them: raise outside money, or sell its own token and buy them with the proceeds. Under Gamma it has a third. It destroys some of its token, receives credit at a fixed value, and hands the credit to the sub-network that owns the cards. The outflow becomes an inflow. The sale that used to push the price down becomes a burn that removes supply instead. Fourteen of the two dozen revenue-earning sub-networks already spend their customer money buying their own tokens back, by hand, through exchanges. Gamma is that same reflex moved into the chain and made a settlement rail.

This has a name outside computing, and it is old. Company scrip. In the mining and timber towns of the nineteenth century, a company paid its workers in paper it printed itself, redeemable at the company store and nowhere else. The word scrip is a shortening of subscription receipt, a slip acknowledging you were owed something. Britain called the practice the truck system, from the French troquer, to barter, and passed a series of Truck Acts to stamp it out.

The reason it was outlawed is the reason to look at this carefully, and the reason it kept being invented is the reason to take it seriously. Scrip genuinely works for the issuer: it costs a fraction of real money, it cannot leak out of the local economy, and it keeps the store busy. The modern respectable version is on every airline's balance sheet, where loyalty miles are a liability the issuer prices in a currency the issuer controls. What scrip never does is create a customer. A company store with scrip-paying workers and no outside buyers is a closed loop that still has to buy its coal with money.

So the pattern cuts both ways, and the diagnostic is one question this show has asked before. To pay the people who hold it, does this system have to sell the thing that makes it valuable? Gamma is a serious answer to that question. It is not an answer to a different one.

The honest implication: Gamma changes who bears the loss and when, and it does not change whether there is one. A sub-network that destroys tokens to buy compute is still funding its operations out of its own supply rather than out of sales. The holder is affected differently, by a burn rather than a sale, and a burn is friendlier, since removing supply is the opposite move to adding it in a thin pool. But a credit that cannot leave the system cannot pay an electricity bill. Somewhere at the bottom of the stack, somebody is still handing real money to a utility, and that money comes from outside or it comes from token holders.

The second implication is more interesting and cuts the other way. Pricing below cost is not automatically a scandal. Every network that ever needed to exist before it could be useful has done it, and emission is the cleanest instrument anyone has built for paying strangers to show up early. The question is whether the subsidy is buying something that compounds, a customer who stays when the price doubles, or whether it is buying volume that evaporates the moment it is priced honestly. One of the two sub-networks named in the analysis as commercially healthy sells security testing and the other sells computer vision. Neither of them is in the cheap-tokens business.

The open question is small and answerable, and it is not about Gamma's design. It is whether any issuer publishes what a unit of its service costs to produce at the Gamma price. A credit priced at a fixed value, spent on a service priced below cost, is a settlement rail running over an unpriced liability. That is checkable, by one number, by anybody who chooses to publish it.

The two voices are AI. The research and writing are mine.

Decentralized AI, layer by layer.

Dastan,