How to Build a Successful Bittensor Subnet: Lessons from Top Builders

Key Takeaways:

• Founder-proven wisdom: Interviews with Taostats, BitMind, and Score reveal what separates top subnets from the rest
• Success starts local: Test incentive logic rigorously before main-net; debugging live costs TAO + reputation.
• Incentives are everything: Miners optimize what you measure, design simple, exploit-proof mechanisms (expect 50+ iterations).
• Master cold starts: Open-source guides, elite miner recruitment, stellar docs bootstrap from zero participants.
• 3 phases of subnet success: Technical validation → convince TAO holders → build real revenue-generating company.
• Target scalable demand: Solve measurable AI problems (deepfakes, sports vision) Big Tech ignores.
• Team + relentless iteration: Lean experts (3-5 people), constant evolution beats stagnation.

Most Bittensor subnets fail. They launch with ambitious visions of decentralized AI, attract some initial interest, and then quietly fade into irrelevance. No miners. No validators. No reason to exist beyond the founders' optimism.

But a few subnets break through. Score processes football games at $10 per match with 94% accuracy and has secured partnerships with Reading FC and AVIA. BitMind protects users across 100+ countries with 95% deepfake detection accuracy, addressing a problem that cost victims $410M in losses during the first half of 2025 alone. Taostats became the analytics backbone for the entire Bittensor ecosystem.

These builders didn't succeed by accident. We interviewed Mogmachine or Mog (Founder of Taostats, incubator/advisor to multiple subnets including Hippius and Vidaio), Ken Jon Miyachi (Co-founder and CEO of BitMind), and Maxime Sebti (Co-founder and CEO of Score) to understand what separates successful subnets from the majority that never gain traction.

Understanding Subnets: The Basics

A Bittensor subnet operates as a competitive marketplace with three key components:

Miners produce a specific digital commodity: AI inference, deepfake detection, storage, computer vision, or other services. Validators evaluate the quality of miners' work and rank their performance. The market determines value through each subnet's alpha token, which trades against TAO in an automated market maker built into Bittensor's protocol.

When people stake TAO into your subnet, they're buying your alpha token, driving up its price and increasing your emissions from the network. When they unstake and swap back into TAO, your token price drops and your emissions decrease. This is Dynamic TAO. The market, not a committee, decides which subnets deserve resources.

The catch? Everything depends on your incentive mechanism: the rules defining what miners should do and how validators should judge them. Get this wrong, and the market will price your failure accordingly.

What Makes a Subnet Truly Successful?

"I would probably define what success is first," Max Sebti of Score explains. "There's success within Bittensor and success outside of it."

Max breaks down three phases of subnet success:

Phase 1: Earn the technical stamp. "The technical success is to be able to find a very clear target and build an incentive mechanism that actually optimizes for reaching that target," Max says. This is how you earn respect from the OGs (original Bittensor community members). You prove you deserve to be on Bittensor.

Phase 2: Convince TAO holders. "How can you convince the TAO holders that your problem is big enough to help you build a billion-dollar company?" Max asks. Success here means getting your messaging across the Dynamic TAO world so people talk about your subnet when you're not in the room. This is about proving you deserve to stay around.

Phase 3: Build a real company. "Now it's time for you to make it a real company and pay everyone back," Max concludes. This is where you combine working on a big problem with actually generating value that extends beyond the Bittensor ecosystem.

Mogmachine emphasizes the evolution of what matters: "Right now, one of the most important things is demand for the commodity you're producing. You need something that has scalable demand. Bittensor is not just about servicing Bittensor users or a very small niche. You're launching companies now."

The bar keeps rising. "Everyone wants to be a top 10 subnet," Mog notes. "The capital flowing in is now more discerning and intelligent, with a wider spectrum of interest. You have funds coming in that invest across tech sectors, and they want competitive, successful investments. You need to appeal to them, not just guys with a lot of TAO who want to spread it around for the good of the ecosystem."

Ken Jon Miyachi of BitMind adds another critical factor: "You need to figure out how to have a good relationship with your miners and get utility out of the structure of Bittensor. That involves the technical aspect of building a good incentive mechanism, but also your communication and how you articulate your product to miners and users."

The common thread? Technical excellence, clear communication, real-world demand, and the ability to iterate relentlessly.

Start Local or Fail Publicly

Every successful builder emphasized the same lesson: test locally, validate on testnet, then and only then launch on mainnet.

The founders emphasized spending months on local development to get incentive logic right before launching. The cost of debugging on mainnet is brutal, both in TAO and reputation. Once miners figure out they can game your system, good luck attracting serious participants afterward.

The Bittensor subnet template is your foundation. Master how miners and validators communicate through the Axon-Dendrite protocol. Understand how Synapse objects structure data exchange. Simulate attack vectors locally. If your incentive mechanism breaks in controlled testing, it will absolutely break in production, except production failures cost you credibility you can't recover.

Local testing also reveals hidden complexity. What seems like straightforward scoring logic often contains edge cases that miners will exploit. Better to discover these in your development environment than watch your subnet's token price crater because miners found a shortcut you missed.

But here's the uncomfortable truth Mog shares: "When any team tells me they're ready to push their subnet live, they are not ready for what's coming. They think they are ready, but the reality is once you put that code out... they can't be ready."

You can invite miners to testnet, "but they don't come. Maybe some might come to see if they can fuck with your subnet, and then they keep very quiet because they found the exploit and just get ready for you to launch it," Mog explains. "You can run your own miners to see if anything breaks, but you're not trying to break it the same way other people are."

"The solution? 'Make sure the code fundamentally isn't broken and you have a good incentive mechanism design,' Mog explains. 'Unless it's your third or fourth subnet and you don't think you're ready, you're just like, well, I'm gonna push this out because I've got three days in front of me with no engagements and I'm ready to stay up until 3 AM backing miners.'"

Your Incentive Mechanism Is Everything

Yuma Consensus, Bittensor's on-chain algorithm that distributes emissions, only works if your incentive mechanism makes sense. Validators submit weight vectors ranking miner performance. Yuma aggregates these rankings, clips outliers to prevent manipulation, and allocates emissions accordingly. But if your scoring model is broken, Yuma can't fix it.

"Miners will optimize for whatever you measure," Ken explains. "If your incentive mechanism has loopholes, they'll find them in days. We've had to continuously update our adversarial architecture to stay ahead of gaming attempts."

Mog puts it even more directly: "Miners are driven by incentive, not what you want."

Consider two scenarios:

Bad incentive design: A text generation subnet that only measures response speed. Result? Miners return garbage instantly because that's what maximizes their score.

Good incentive design: BitMind's adversarial system where multiple models compete to detect and generate deepfakes, with validators scoring both detection accuracy and adversarial resilience. Miners will find it hard to game this because the system adapts to their strategies.

The Reality of Revisions

Expect to overhaul your incentive mechanism multiple times. Ken's BitMind has done "three major overhauls but probably over 50 or 60 changes and upgrades over time."

Max's Score went through two complete ground-up revamps: "We literally took the repo and threw it in the bin. The first time we weren't even aware of all the exploits. The second time we thought we could work in iterations, patching exploits as they emerged. What we realized is the way we were framing the problem was so narrow. We were just keeping patching exploits."

Their third approach was radically different: "We're going to first think about how can we make it exploit-proof, then see if someone can exploit it. We ran a lot of tests, worked with five different mining teams, and shared different information with each to see how they would try to exploit the subnet."

Score also got creative about pushing exploit risks elsewhere: "Instead of trying to be good at ranking miners on their ability to run inference, we pushed the inference problem to Chutes. Chutes works 24/7 to make sure they're not being exploited on inference. For our private track, we run things on Targon. So we can focus on one thing: the quality of the models we're building."

This is the future of subnet architecture. "More and more subnets are working with this kind of new setup where you can combine different subnets together," Max notes.

Mog emphasizes two types of incentive evolution:

Reactive evolution: "There's always a dynamic in mining groups where one is accusing another of gaming the system, and the other says, 'Just because you can't do as well as me doesn't mean I'm gaming it.' You've got to look at it. If you're incentivizing for something and that becomes the norm, your incentive mechanism needs to evolve to continue incentivizing advancements."

Proactive evolution: "A good subnet doesn't stop growing. You've pivoted, you've evolved, you've gone in different paths. So the incentive mechanisms continually evolve along with the iterations of the subnet. They also have to take into account economic changes in Bittensor, like TAO flow."

Team Structure for Incentive Development

"Every subnet has the subnet guy, the one who's committing the code in the repo for the incentive mechanism design," Mog explains. "However, a successful subnet has experts in other elements. For video, there's an upscaling guy, a compression guy, someone looking into other things the subnet is doing. Where the magic happens is when they all start to work together. The experts understand the subnet and the incentive mechanism, so they work together. Normally they're concentrating on product advancements, and the subnet person is working out how to integrate that within Bittensor itself."

How lean can you run? "Jake (Jacob Steeves, Bittensor co-founder) would argue it only takes one person to make a really good subnet," Mog notes. "But the reality is, at a minimum, you've got your subnet dev, a front-end developer, a designer. If you find a front-end developer who's also a designer, that's a holy grail. Then you need someone skilled at biz dev, marketing, socials. You can probably get away with three, but I would say five is a good number to run a subnet."

The Most Counterintuitive Lesson

What's the biggest surprise about incentive design?

Max's answer is brutally simple: "Trying to be smart is the worst thing. Good incentive mechanisms have to make sense and really have to be limited to optimizing for one continuous target. When you enter the Bittensor ecosystem, there's so many smart people that you want to come up with something that looks smart, sounds smart. You're dropping math, you almost want everything to look like the Dynamic TAO white paper. At some point you're creating doors for people to come and exploit your subnet. The most counterintuitive thing? Less is more."

The Cold Start Problem

Launch day arrives. You've registered your subnet, deployed your code, and written documentation. Now you need participants.

"Month one, you have nobody except maybe some friends doing you a favor," Mog admits. "You need a credible answer to: why should someone allocate compute to your unproven subnet instead of mining something established?"

Bootstrap Strategies That Work

Ken's approach with BitMind was to lower barriers aggressively: "We open-sourced our solution and created miners ourselves to get started. We made a bunch of guides for how to get started on Runpod or Tensor Dock, just making it really easy to get set up to start mining. The miners weren't competing on a baseline lower than what we needed. We set the baseline for folks, and then they could improve on that."

Max notes that the Dynamic TAO era has actually made this easier: "Pre-DTAO you had to go on meetings with everyone. Now it's easier because there's more scrutiny, more people trying to see what's the latest subnet. There's also more structure. A lot of different mining teams have their own unit economics based on their ability to find new subnets early on."

But there's a critical requirement: "If your documentation is not good enough, you'll completely lose attention. People will just jump on another subnet. The best way is to have a good website, good documentation, and probably try to get on different podcasts or at least chat on different community channels. Miners and validators are going to reach out to you if they think they can help."

Attracting Elite Miners

"The biggest thing now that we have incentive built in: once you build economic incentive for folks to come, I think that is still the biggest driver," Ken explains. "But it's almost like active recruitment. You're trying to pitch your idea that it's going to be really successful. I think of them as venture miners. You want them to really believe in your product."

The holy grail? "If your miners mine but they actually hold the token. It's probably not completely feasible because they have expenses and capex to spend for compute, but if you can get them to really believe in what you're building and think it's a good investment, that is the ultimate goal."

Ken's number one tip: "When people do reach out, when they're asking questions in the Discord, just be really supportive and try to help people. That is tough because it can be a huge time sink, but really making an effort to be approachable and getting people started matters."

The Elite Miner Evolution

Mog offers a fascinating observation about how elite miners emerge: "In the early days of subnets, I was the low-quality first miner, the one that just takes your README file, follows the instructions, does what's needed. Until someone comes along that does something beyond that, I get incentive, and then I don't anymore."

He distinguishes between two types of miners based on his Vidaio (video upscaling subnet) experience:

"The initial mining team that was most successful early on came from a different niche of Bittensor, but had Bittensor experience. They had experience optimizing for Bittensor workflows by being a miner and could look at the upscaling requirements and optimize the serving of them. They have a huge relevant skill set: optimizing for Bittensor, serving as a miner, optimizing models."

But then something interesting happened: "Along came a team who specialized in upscaling, but not video. They had an image upscaling background, were not crypto-based, and managed to apply their skills to upscaling video. As soon as they got to grips with the Bittensor side, they completely destroyed the Bittensor-native team because their specific industry skills were much more valuable to the task at hand."

The lesson? "I know very few subnets that have had to make a concerted effort to attract those elite miners, other than building something that attracts the incentive, and then the miners come. If you're going for a niche industry, you may need outreach. But what I've found is that elite miners actually form around your subnet."

Handling Low-Effort Participants

"Any subnet has those initial miners that just want to follow the README," Mog explains. "It's actually a bit of a controversial point because Bittensor is marketed as an open network where anyone can come along and mine. But I can tell you from the subnet trenches, subnet owners get fed up with them pretty quickly. This isn't tech support while you install my software. Mining is competitive."

Documentation quality directly impacts your ability to attract serious miners and validators. Clear setup instructions, transparent incentive logic, realistic compute requirements, and honest communication about risks make the difference between building a community and collecting opportunists hoping for easy emissions.

Mog had an advantage with Hippius: his existing reputation from building Taostats gave the subnet initial credibility. "If you're unknown, your incentive mechanism needs to be obviously sound, your value proposition clear, and your technical competence evident from your code quality."

Early participants are often your harshest critics and best teachers. Listen to their feedback. Fix issues quickly. Demonstrate that you're committed to iteration and improvement rather than defending a rigid vision.

Choosing Your Problem: What to Build

Why These Builders Chose Their Niches

Ken (BitMind, Deepfake Detection): "I felt generative AI was getting so much better, and I was really inspired by all the upcoming elections in 2024 and the effect of generative AI to influence those elections. But more broadly, I just felt it was a very fundamental thing that needed to be solved to keep AI beneficial to humans."

What made deepfakes perfect for Bittensor? "What I really liked about the deepfake problem was the source of truth and the evaluation was extremely simple and elegant. It's either AI-generated or it's not. It's binary and very simple."

Ken also explored computer vision applications before focusing on deepfakes: "I thought the computer vision field had a lot of exciting areas like robotics, manufacturing, really cool stuff. But what I really liked about the deepfake problem was the source of truth and evaluation was extremely simple and elegant. It's either AI-generated or it's not. It's binary and very simple. This really called to me as being an existential problem that needs to be solved, not that many folks are talking about it, there's not that many companies pursuing it, and I just thought it was so fitting for Bittensor."

Max (Score, Computer Vision): "We had a customer running a betting platform, and we thought traditional predictions were the best way to help them. We realized there's way more edge or interesting data to find through vision in sports. But then we learned something really important about Bittensor: you should be more ambitious. We started with a very simple subnet, turned it into something more ambitious, and now it's even more ambitious."

Score still uses that phrase internally: "We keep repeating this to each other. We should always try to see things in a bigger way."

Mog (Taostats, advisor to multiple subnets): "I always wanted instant translation, and Babel is doing that now. Industries which have not successfully been integrated into Bittensor: quantum computing is there now. Medical, the ability to create a commercially viable product or service within the medical industry that can work in either the current Bittensor architecture or in a way that you can manipulate the Bittensor architecture to work for you. Big industries with huge amounts of value floating around that you should be able to tap into."

The Most Underrated Opportunities

Ken: "Robotics would be super interesting. There's so much money being poured into it right now. My assessment was our team didn't have the domain expertise, and it's extremely complicated. The other piece about robotics that's difficult is it's physical AI. You can't just purely benchmark it in software. You really need to test it out in the real world on a robot. You can simulate all you want, but that doesn't transition to the real world as well as you would think."

The unlock for robotics on Bittensor? "Some really elegant and efficient way of getting feedback from the real world in a highly accurate manner would be a really big unlock."

Ken also mentions hardware integration is coming: "I helped Zeus (Subnet 18) spin up their subnet, and they just partnered with WeatherXM. WeatherXM deploys weather stations around the world you can buy, and it reports weather back to their network. Zeus is going to start using that as data sources for their predictions and source of truth. I do think over the course of 2026 they'll be maybe not building hardware but incorporating hardware and different IoT devices or even robotics into subnet feedback loops."

Max: "The common answer would be something around agents."

What Problems Fit Bittensor Best?

Mog offers crucial advice: "A lot of people come to me with a subnet idea and want to make their business a subnet. That's great that your business does that, but now you need to understand what is the workflow that obtains that. You're not just going to ask everyone else to produce that commodity for you and it's going to work. You need to build rules around the value of the process and value evaluation of the production of that asset."

Max adds: "Start by trying to tackle a big problem, and then think about go-to-market strategy. Don't do the opposite."

Understanding Value in Bittensor

"I wouldn't measure value in token price," Mog stated. "The token price is an end reflection of user-perceived value, and at the moment it's user-perceived value based on those that are investing in subnet tokens. It's a very small market. The value should be a clear optimization and betterment of an existing product, service, or workflow. You're making something measurably better and then hopefully continually measurably better."

Max has a different take shaped by Dynamic TAO: "DTAO introduced a very short-term, high-frequency type of mindset. I think value in Bittensor will also come from time. It's all about the time horizon. I would Bittensor is extremely undervalued. Most subnets are extremely undervalued for different reasons, but maybe things are going to change in a few months."

He emphasizes patience: "What Chutes is doing now could be done in Web2 at a hundred times valuation. I just think perceived value is being highly impacted by the short-term horizon brought by DTAO. But if we take a step back, Bittensor could be bigger than OpenAI."

Ken distinguishes between two types of subnets and how they should measure value: "For profit-focused subnets, value is straightforward: you need revenue that exceeds your emissions. You're producing more value than what you're putting in. For research-oriented subnets, value is measured differently. It's about velocity of progress, showing that you're making significant progress toward your goal even before you have revenue."

His personal philosophy? "Velocity always wins."

Trust and Reputation in a Trustless Market

"Communication is probably the most critical thing," Ken emphasizes. "Trust in general takes a long time to build up but can be lost pretty quickly. Your social reputation, if you rug someone or do something malicious to miners or scam people, they remember that and they're not going to want to work with you. The entire idea is that people need high confidence this is going to be good for them in the long term, or at least that it's going to be fair. Open communication leads to trust within the ecosystem."

Lessons They Wish They'd Known Earlier

Mog: "Hire fast and fire fast. I've been quite tolerant and like to give people the benefit of the doubt, but all of the stars of my team are clearly stars very quickly. Those that excel get up to speed very quickly. You need a strong team, not a big team."

Ken: "First, start off mining or doing something where you see the process from the miner perspective. Look at the miners as customers or employees. Understand their perspective because there are some adversarial components to the miner-subnet owner relationship. Being able to empathize with them and understand where they're coming from is really important."

Max: "Start by trying to tackle a big problem and then think about go-to-market strategy. Don't do the opposite."

What Keeps Them Building

Mog: "I need that technical mental stimulation. I feel like I've found something where I add value. As long as I can add value and feel motivated to keep doing it, I'll continue. But if I suddenly feel I'm not adding value and it's sucking from me rather than me being able to put back into the ecosystem, I would probably step back."

Ken: "Data privacy in terms of BitMind, helping our users. The entire premise of starting the company was based on almost an ethical concern. These companies controlling everything are going to determine what we see, what type of information we consume. People view ChatGPT and these chatbots as the source of truth. Just making sure that is open and transparent and not controlled by private interests or economic interests."

Max: "Democratizing computer vision. There are thousands and thousands of companies, entities, organizations that would literally change the world if they could have access to this tech cheaply and in an open-source way. We should all be doing open-source AI and make sure people are able to check everything from the data to the model to the weights. Everything should be open-source."

The Uncomfortable Truths

Mog: "A lot of what you feel about how business should be structured is going to be completely turned on its head by a 15-year-old kid or some spotty teenager who understands the code and the incentive mechanism but has no idea about business, and yet is the future. You are going to have everything you think you know about business challenged. At the same time, things don't move as fast as some of these kids think. There's ideology, and I have to turn around and say, 'That's great that you're ideological, but I hope you're prepared to stick to those ideologies for 10 years while the rest of the industry and the powers of the world catch up.'"

Max: "It's probably a hundred times harder than what you think."

Max quotes Carro from Manifold: "Running a subnet is like eating glass for breakfast."

Ken agrees: "You get consistently tested all the time. It's crypto, so it never stops. You need to always be kind of terminally online, always aware. You need to do a lot of things right: technically, communication-wise, relationship building. Startups are hard. It's not easy."

Is Bittensor an Easy Place to Launch Your Startup?

The unanimous answer? No.

Mog elaborates: "It's not easy at all, but it's incredibly rewarding. There have been many times, even from the early days, when teams would say, 'This would be a lot easier if we just did it without a chain and paid for our own GPUs and got backing, or we just did this on our own chain and not on Bittensor.' You're building within the confines of a predefined infrastructure network that you don't have full control over. You have to evolve and adapt to it."

But there are advantages: "You want to have the biggest access to talent, you want to have the most eyes on you, you want to jump onto an already spinning flywheel. It offers a huge amount of opportunity. And let's be honest, there is no other working network like Bittensor that you can come an build on. There's a lot of VC-backed altcoins saying they're going to do all this but are not doing it yet. What Bittensor is doing and achieving is truly unique in the crypto space at the moment."

Ken agrees but adds context: "What we've seen is it does improve velocity. Our alternative was to raise a bunch of money and hire a bunch of AI engineers to create these models and roll it out as a product. I do think it can be more efficient. But it's definitely not easy. You need to be able to define your problem really well and believe you can evaluate it algorithmically. If you can, it's a very compelling place to start a company. But it's definitely not easy. There's a lot and it's crypto, so it never stops."

Core Principles for Success

The builders we spoke to converged on three non-negotiable rules:

Solve a real problem. "Decentralized [existing thing]" isn't a strategy. Chutes works because decentralized inference is genuinely 85% cheaper than AWS for comparable workloads while maintaining competitive performance. BitMind works because deepfake fraud hit $410M in losses during just the first half of 2025, creating desperate demand for reliable detection. Score is making computer vision accessible to thousands of companies that would "literally change the world" if they had affordable access to the technology. Find a genuine wedge where decentralization creates defensible advantages, not just ideological appeal.

Expect constant iteration. Your initial incentive mechanism will be wrong. Not slightly imperfect, wrong. Plan to update frequently based on observed miner behavior and emerging exploits. The best subnets update regularly, incorporating new hardware capabilities, new models, and adapting to new gaming strategies. Stagnation equals death.

Community is signal, not strategy. A Discord with 10,000 members means nothing if your miners are gaming your system and validators aren't doing real work. Focus on attracting competent participants who care about the actual commodity you're producing, not speculators hoping for token appreciation.

The Reality of Building Subnets

Building a successful subnet requires technical competence, economic design skills, and enough paranoia to assume everyone will try to exploit your system. Most people lack at least one of these. The founders who succeed understand that Bittensor is infrastructure for creating genuine value in a brutally competitive, market-driven environment.

The market will find every weakness in your design. Miners will optimize for your metrics whether those metrics align with your intentions or not. Validators will form alliances if collusion is more profitable than honest evaluation. Your token price will reflect the market's collective judgment of whether your subnet actually deserves to exist.

If you can't articulate why your subnet creates real value, why your incentive mechanism resists manipulation, and why anyone should trust your ability to iterate and improve, save yourself the registration fee and build somewhere else.

For those who can answer these questions, Bittensor offers something rare: a protocol-level mechanism for coordinating decentralized production of digital commodities with built-in economic incentives. The subnets that succeed will define what decentralized AI infrastructure actually looks like. The rest will serve as expensive lessons for the next wave of builders.

Start local. Test relentlessly. Ship carefully. Iterate constantly. And remember: the market is watching your every move.

But also remember what motivates the best builders. It's not just the token price or emissions. It's the opportunity to democratize technologies that could change the world. It's the chance to solve existential problems that Big Tech isn't addressing. It's the technical stimulation of building in the wild with immediate, unforgiving feedback.

The playing field is leveling up. The bar keeps rising. And the subnets that succeed will be the ones that embrace this reality rather than fighting it.

Resources:
If you are interested in building a subnet on bittensor you can get involved through the links below.

• Join the Bittensor community on Discord
• Study the documentation at docs.bittensor.com
• Explore existing subnets at Taostats.io
• Learn from subnet code repositories and incentive mechanisms