Full article written by Kai. Connect with him on Substack and Twitter (X).
At the core of every subnet is its incentive mechanism. This is the process used to convince people to contribute, determining the limitations of the subnet competition and the parameters for achieving rewards.
The incentive mechanism is what both attracts and repels miners. If the setup appears both feasible and lucrative for a miner, then they’re likely to take part. If it appears too complex, and the payout seems too hard to reach, then many will be deterred.
However, in the sea of varying incentive mechanisms in the protocol, there’s none more controversial, yet as enticing, as the winner-takes-all setup. To call it polarising is an understatement. Let’s dive into it and cover what it really means for the ecosystem, and how it defines the space.
The hero’s journey
Winner-takes-all is an extreme form of the Bittensor architecture. In all subnets, the way miners are rewarded is determined by Yuma Consensus – this is where validators evaluate their work, and give them a score accordingly. The higher the score, the higher the reward. The magic of Bittensor is that the scores are given more weighting when validators give similar scores to the same miners. In other words, miner-rewards are treated more seriously when validators are in consensus about them.
Winner-takes-all is a distilled, electrified form of this. If a subnet implements this within their incentive mechanism, then only the highest scoring miner will receive a reward. Everybody else will receive nothing.
In the average subnet, you’d receive a reward for your work even if you weren’t the cream of the crop, you just wouldn’t receive as much as the top performer. In winner-takes-all, the best miner receives the whole bounty – in other subnets the best still receives the most, but others get something, too. As long as they’re able to contribute something meaningful, they can make an earning.
Winner-takes-all is a form of hero’s journey – to enter a subnet, perform better than anybody else, and then take the whole pie. It doesn’t mean you’ll always be the best, as competitions often refresh over time, but for the one you compete in, you’ll always be etched in the ledger as royalty.
Not only that, but you can get rich. If we take SN62, Ridges, as an example, back in August 2025, you could theoretically make $58,000 in just ten days.
No subnet will ever pay out as impressively as a winner-takes-all model.
The bounties are eye-watering. Of course, the work required is astronomical – nobody can take home money like that unless they know exactly what they’re doing, innovating and studying in the subnet’s field on a regular basis (or at the very least throwing astronomical money at the problem on their own side).
The incentive for being the best is extremely high…
Social dynamics within winner-takes-all
As a Content and Communications Manager in the Bittensor ecosystem, my role is to observe the status of different communities across the protocol. What I’ve seen in winner-takes-all subnets is striking.
The communities for these projects have a slightly different flavour than the rest of Bittensor. Naturally, when miners are all competing for the same (high value) reward, there will be a specific type of feverish behaviour exhibited, compared to subnets which distribute rewards across more individuals.
There will be a scrambling for the top. Not just because of the scarcity of earnings, but because people who do compete will naturally be dedicating their time, resources, and (in some cases) money to achieving. Some subnets implement a slight variant – a winner-takes-most approach where the top handful of participants get the lion-share – which reduces the pressure, but only by a little.
The best example of this is Macrocosmos’ subnet 9. In the past, SN9 was designed as a winner-takes-all network, where miners fought to submit the most optimised ML pretraining models. Its community was bustling, but it was fractured. Back then, a peek at the Discord server revealed dedicated, highly intelligent individuals who all had critiques, issues, and chips on their shoulders.
The team had attracted some of the best talent in the ecosystem, but the cost was a community lacking warmth. Is that a feature, or is it a bug? It’s hard to say with any conviction. In reality, when you push incentives like that, you’re bound to create a tightly wound environment. Tensions will be high.
This is the hidden cost.
A winner-takes-all subnet must sacrifice any notion of collaborative engagement in return for attracting the best talent. You see this too in the Ridges community. It’s a miner-eat-miner world – there’s no reason to help your peers because the reward isn’t shared.
While the success of the subnet is theoretically great for everybody, it’s only great in practicality if you can win its competitions. This is opposed to subnets where rewards are distributed to a greater range of participants. When SN9 switched from a battle to produce the best pretrained models per miner into a competition where everybody was rewarded for training the same mega-model, its community changed. People were friendlier, and helped each other out more often.
There’s a greater camaraderie when there can be more winners. Not only that, but the goal relates to more stakeholders. It’s not a fight to the death anymore, but a joint march to the goal. The incentives won’t be as high, but the tradeoff is you gain a distributed team. In the decentralized world, community-building is everything – these systems are hard to navigate, and if you want to onboard more people, then it’s best to have those on the inside actively caring about boosting others. It’s the flywheel effect.
Two stories within Bittensor
There’s no polar opposite to winner-takes-all – if there was, it would essentially be a flat line where all miners are rewarded the same, regardless of their output. In Bittensor, that would never work. What we have instead are subnets with incentive mechanisms that have softer boundaries, where success is distributed to a greater number. The ranges will differ based on the subnet, but so long as the greatest share is not shipped purely to the all-star participants then it’s essentially an alternative setup.
Both winner-takes-all, and its commonplace alternatives dial into differing attributes of the protocol. In subnets where there’s one victor, the primary purpose is to bring attention, and draw out top minds who might have never had traditional opportunities to show their expertise. Take an AI development subnet for instance – if it has a winner-takes-all architecture, then the only people who have the time and dedication to fiercely compete are those who don’t already have demanding jobs, yet do have the intellect for it.
In other words, it gives chances to those who never had them before. A top engineer who had the privilege to attend a top university would likely find themselves at a top company. But a highly skilled, highly intelligent engineer who never had those opportunities might be stuck doing something far below their skillset. These winner-takes-all subnets give them the ability to rise to the top, in what feels like a meritocratic way. Whether it really functions like this is a different story, but for the granularity of this discussion it makes sense.
Winner-takes-all becomes a rally across the globe for the best minds to clash. They’re gladiator arenas, able to make dreams come true.
The Historian Frederick Jackson Turner spoke of the “Frontier Thesis”, the idea of a location, specifically America, functioning as a merciless landscape where titles and lineage faded, leaving only you as your raw and stripped back self, forced to survive with your own tools and intellect. It’s a romantic belief, not in spite of its brutal impressions but because of them. You direct your future, removed of your baggage. Everybody is equal. This is, for better or worse, the type of dream winner-takes-all embodies. Leave your prestigious degrees, gold-plated certificates, and peer-reviewed Arxiv papers at the door. People only care for your output.
Alternative reward models work on a different dream: decentralized collaboration. The sum of the subnet is not based on the best performer, but the collective spirit of the miners. They dial into Bittensor’s finely designed ability to align incentives across participants so they can contribute towards unified goals. This is how the current SN9, IOTA, works. It’s how SN25, Mainframe, worked. And it’s how many others operate.
Both are meaningful. Both tap into Bittensor’s strengths. But they take the protocol down different routes. If anything, it’s a sign of the blockchain’s versatility – rarely will you see such differing strategies on the same network. Yet, Bittensor makes it possible.
There’s also times when a subnet might change its structure and ethos, so as to accommodate a new path. With SN9, the team used a winner-take-all architecture to prove that highly intelligent ML engineers could be incentivised to pretrain models. The high reward drew in the initial crowd. It proved the team had a strong command of pretraining, and revealed Bittensor housed the right community.
Once that happened, the team pivoted to a collaborative model, where the world could train in unison. The first phase showcased the talent – the current phase showcases the mission. In this sense, winner-takes-all can operate as an effective bootstrapping tool.
A state of coexistence
This is not a discussion on which is better or worse. Rather, a spotlight on how incentive mechanisms affect communities, userbases, and ideals. These aren’t just methods of convincing people to work, but systems that carry their own themes and baggage. Incentive mechanisms don’t just dictate earnings; they engineer culture, and they define who you become while doing it.
Be the first to comment