The Chain of Thought Decentralized AI Canon is my running attempt to chart where AI and crypto collide, overlap, and fuse into something new.

Each Canon zooms in on a different layer of this economy (trust, data, cognition, coordination, robotics) and then pans out to show how they connect into a single system of intelligence and incentive.

The Canons shift as the terrain does. New essays, data, and debates reshape the map every time we learn something worth revising.

Subscribe if you want to follow this live experiment in mapping the intelligent economy.

We’ve been spending more time in the Bittensor ecosystem lately. At this point, it’s literally foolish for anyone to ignore.

Bittensor is the most compelling place to watch Crypto x AI unfold in the near term.

Especially because it’s a time of major change for the ecosystem. Winds of change always bring huge opportunities for the observant.

Bittensor was built when no one cared. It launched in 2021, before Crypto x AI was a thing. That gave it room to grow without distraction. No speculative hype. No VC incentives shaping the roadmap. Just a small group of people aligned around a shared goal.

The TAO token launched fairly. No early allocations or fundraising rounds. That structure matters. Early contributors not only participated: they owned it. Some reinvested. Some built subnets. Some just did their part by shilling on Twitter.

In short, it’s a cult.

There’s weird stuff, sure. But also real conviction. There’s actual activity on the subnets. This is where attention is concentrating, and that alone creates second-order effects.

Of all the crypto-native AI projects we’ve seen, Bittensor is the most structurally aligned with Web3’s first principles: fair launch, open participation, and live incentives for coordination.

Some call TAO the Bitcoin of AI.

Others say it’s the internet of AI, like this investor deck slide above from Fundstrat Capital (a TradFi asset manager and research firm).

The idea is that it acts as a base layer, an index across a growing set of subnets working on everything from training to inference.

There’s some truth in that. But if we’re being frank, there’s also a lot of narrative dressing.

We’re not here to repeat slogans. What we’re trying to do at Chain of Thought is offer a clearer look at how the system actually works, so you can make your own judgment and position accordingly.

The way Bittensor operates has fundamentally changed in 2025.

With the launch of dTAO, validator gatekeeping is gone. Instead, capital now flows to subnets based on real, market-driven demand. The old playbook no longer applies.

Under the previous system, emissions were routed through validator votes on Subnet 0—the “Root.” Validators were supposed to evaluate AI outputs from different subnets and allocate TAO based on quality. In theory, this was a meritocracy. In practice, it broke down fast.

A few problems kept coming up:

The system was slowing down innovation, concentrating influence, and eroding trust. A reset was inevitable.

Without getting into the math, here’s the core shift:

dTAO hands the decision over to the market. Subnets now have to earn emissions by proving demand.

Each subnet issues its own token, called Alpha. Miners, validators, and operators are paid in Alpha instead of TAO.

Every Alpha token is paired with TAO in a liquidity pool, similar to a Uniswap AMM. If you want to earn emissions, you stake TAO into that pool to buy Alpha: essentially swapping out your TAO. A high Alpha price signals investor demand. A low one signals weakness.

TAO emissions are distributed based on those prices. A high Alpha price means more emissions. Low price means less. To avoid short-term price manipulation, emissions adjust gradually using a rolling average of Alpha prices.

This removes the need for centralized validator decisions. Subnets now have to convince TAO holders that their token holds value.

Some critics argue that dTAO adds unnecessary complexity, and we can see their view. But at its core, it solves a crucial problem: aligning incentives. If subnet owners or miners sell their alpha tokens, their emissions decrease.

563 had a good thread on how different participants should react to dTAO, which is still very relevant today:

Builders, miners, validators, and stakers are now all pushing toward the same goal—creating subnets that deliver real utility, attract real users, and sustain real economic value.

As an investor, owning TAO still matters, but it’s no longer enough. You have to go deeper, into the subnets.

Evaluating fundamentals is now essential. Alpha tokens give you leveraged exposure to subnet success. But that leverage cuts both ways. Inflation is high. Liquidity is thin. Volatility is baked in. Some subnets like Chutes (SN64), Targon (SN4), and TAO Hash (SN14) are already at $1B+ fully diluted valuations.

So, how do you navigate this? With patience and sharp analysis.

Time spent understanding subnets > trying to time the buy

Ultimately, the “real Alpha” lies in spotting undervalued subnets before the market catches on. Look for subnets with credible teams and signs of early traction.

We’ve included a list of tools we use at the end of the article to help you with that search.

We track several metrics to get a clear sense of how the dTAO ecosystem is evolving. All of this data is public and updated live at tao.app (really good data website)

The most useful signal is the Sum of Alpha prices. This represents the combined fully diluted value of every subnet’s Alpha token, relative to TAO’s market cap. Right now, that ratio sits around 1.8. It’s ranged between 1.2 and 2.3 over the last few months. The higher the number, the more speculative interest is flowing into subnets.

But that number alone doesn’t tell the full story. We also look at the ratio of TAO injected versus TAO reserves. The green line tracks the TAO being injected into subnet pools as per the dTAO emission mechanism. The purple line shows the actual TAO that users have staked (or unstaked) into those pools + the injections.

When the purple line climbs above the green, it means users are allocating more capital to subnets. That’s a risk-on signal. More confidence, more activity, more subnet exposure. When the purple line falls below the green, it suggests the opposite. Capital is rotating back to Root.

Clearly, we are in a risk-off regime for the subnets at the moment (purple < green). This will probably change soon.

Snapshot data shows that 61% of all TAO is still staked on Root, while only 7.7% is allocated to subnets. This balance is expected to shift. Root yields are gradually decreasing due to the declining Root Proportion, which controls how rewards are split between Root and subnet staking.

Over time, this forces more TAO to migrate outward in search of yield.

At a TAO price of $415, that shift represents more than $468 million moving into Alpha token markets. Of course, Alpha inflation will dilute the impact. Supply overhang remains a factor.

But the bigger point is that this capital inflow won’t be evenly distributed. Some subnets will attract far more capital than others, and not necessarily in proportion to where capital is allocated today.

That’s the opportunity. Most investors are still reacting to where capital was. The edge comes from spotting where it’s going next.

This transition is already drawing in professional capital. Newer investment firms like Unsupervised Capital are positioning early to capture this gap.

If we look at the emission distribution among subnets, the top 10 subnets out of 100+ subnets receive >55% of the total emissions, with 4 subnets receiving >5% of emissions each: Chutes, TAO Hash, Gradients, and Targon.

For more on dTAO and its actual mechanics, we wrote a full-length guide here.

From our conversations with subnet owners and firsthand experience inside the ecosystem, a few key lessons stand out:

The dTAO upgrade handed every subnet team a token and an audience (read: degens).

Neither came with a manual.

As far as we could tell, most teams weren’t ready. Until now, subnets on Bittensor were effectively subsidized. TAO emissions covered costs. There was no need to manage a token, engage with speculative holders, or explain performance to a live market.

That changed fast.

Now, subnet teams are on the hook. They have to fund operations, often with operating costs in the 5-6 figure range per month. The obvious move is to sell some of their Alpha tokens, which are often the only liquid asset they hold. But doing so comes with risk. Even responsible selling can trigger backlash (”team is dumping”). It looks like a rug, even when it’s just covering GPU bills or paying engineers.

It’s not wrong to sell tokens. That’s what they’re for. The difference lies in communication.

Teams that explain their moves clearly build trust: why they’re selling, how it connects to the roadmap, and KPIs

Subnets that handle this well treat the token as part of the product, not an afterthought. They design for long-term alignment. They reward contributors. They stay transparent. They make their economics legible.

Because in the post-dTAO world, emissions aren’t guaranteed. You have to earn them by convincing TAO holders to stake into your pool. And that starts with a token that means something. Clear, differentiated, and able to signal future value.

For example, Hippius (SN75) is a decentralized cloud and storage network. It bridges its Alpha token into its own chain, where it functions as a native currency that can be staked and used for rewards.

If you’re building a subnet in 2025, there are no shortcuts. Reputation and validator connections won’t sustain you anymore. The market is the arbiter now. And the most visible signal it uses is your Alpha price.

Engineering early demand is how you get off the ground.

Github commits, uptime metrics, Discord engagement, and model benchmarks are the signals that matter. If you’re not shipping or not communicating, the market will move on. Because here, price is the pitch. Your Alpha price is a live referendum on your credibility.

Without enough TAO staked early, emissions stay low. Without emissions, growth stalls. Without growth, attention fades. That feedback loop cuts both ways.

So what should you be looking for?

Early on, I thought of Bittensor subnets like AI startups. I’ve since realised that I was wrong.

Instead, it’s much better to think of them as communities

You can’t run a subnet like a company. The people mining your subnet are not just employees. They’re part of the network’s core surface area: users, contributors, and stakeholders rolled into one. If they stop believing in what you’re building, they leave. And when they leave, the subnet collapses.

So relationship-building is survival. You have to understand their needs, earn trust, and keep them engaged through rough patches.

A lot of subnets won’t make it. That’s expected. The system is designed to churn. Capital and compute will flow toward the most resilient ones. The rest will quietly disappear. That’s a feature, not a bug. It’s how Bittensor evolves and gets better.

As for anyone planning to clone Bittensor, good luck!

Bittensor worked because of a special mix of timing, culture, structure, and a lot of luck. You can tweak the mechanics, but that won’t recreate the context.

One of its strengths is how loosely coupled the system is. Subnets focused on vision models, trading bots, or private LLMs can evolve in completely different directions. That level of specialization happens faster than any unified ecosystem could manage.

Culture matters too. Many teams we spoke to in the network, like Macrocosmos, care less about short-term price action and more about building lasting infrastructure.

What’s especially promising now is the rise of subnets focused on foundational pieces—systems that other subnets can build on.

In effect, these subnets function like modular Lego bricks, highly composable and infinitely remixable. SN3, focused on decentralized training, and SN19, specializing in inference, are prime examples.

That modularity opens the door to a long tail of AI-native applications. Not everything needs to be built from scratch.

So when someone says they’re building the next Bittensor, I usually assume they don’t understand what made the original work. If something does surpass it, it won’t be a copy. It’ll come from an angle we haven’t seen yet.

At some point, Bittensor subnets are going to have their “Virtuals moment.” A wave of speculative mania is almost inevitable.

The trigger could be something simple, like bridging of Alpha tokens to other chains, where they land on the radar of degen traders. Projects like taofi and voidai are already hinting at that direction. I don’t really know.

But I’m convinced it probably happens within the next 6 to 12 months.

Despite the cost of over $100K just to register a subnet, plus months of engineering, the number of subnets keeps rising.

As the ecosystem matures, more advanced financial tooling is likely to show up. Concentrated liquidity. Permissionless Alpha pools. Maybe even derivatives tied to model performance. The mechanics will get more sophisticated, and the stakes will rise.

If Bittensor succeeds, dTAO will be remembered as the spark that ignited the decentralized AI renaissance.

Cheers,

Teng Yan

Thanks to 563 from Blocmates for reviewing and providing valuable feedback for this essay.

Share your take or a quick summary on X. Tag @cot_research (or me) and we’ll repost it. P.S. Liked this? Hit Subscribe to get the next Big Idea before anyone else.

These are the ones that have stood out so far. Some we’ve covered already. Others we’re still digging into.

This list isn’t comprehensive. There are definitely strong projects we haven’t gotten to yet. We’ll keep filling in the gaps in the coming weeks.

Templar (Subnet 3) – A decentralized AI training subnet that pools heterogeneous compute resources via blockchain incentives to train large models collaboratively. It just completed the training of a 1.2B parameter model.

Chutes (Subnet 64) – A GPU-powered inference subnet by Rayon Labs offering serverless AI compute. Developers can deploy and scale models on-demand (“deploy, run, and scale any AI model in seconds” ). The network processes millions of tokens daily

Gradients (Subnet 56) – A no-code AI training subnet that lets anyone train image or text models on Bittensor with just a few clicks. It is currently in public beta with an intuitive UI and API for on-demand model fine-tuning, making AI training accessible to non-technical users.

Masa (Subnet 42) – A decentralized data-scraping subnet that provides real-time data feeds from sources like Twitter (X), Telegram, and Discord for AI applications. It has already ingested hundreds of millions of data points and powers services such as real-time AI agents and analytics.

Dippy Roleplay (Subnet 11) – A consumer-focused conversational AI subnet aimed at open-source roleplay chat models . It is backed by the Dippy AI companion app, which boasts a multi-million-user base (over 1–4 million users reported) and top rankings in app stores.

Tensorplex Dojo (Subnet 52) – A human-feedback data subnet designed to crowdsource labeled datasets for AI training. Contributors of any skill level can earn TAO by providing or validating human-generated data (e.g., annotations, feedback), forming a decentralized data pipeline for model development.

Score Vision (Subnet 44) – A decentralized computer-vision subnet focused on sports video analysis (especially soccer). It processes match footage in a distributed way (using VLLMs for frame sampling) to extract features for live commentary, betting, scouting. The subnet has already analyzed hundreds of thousands of minutes of game video

BitMind (Subnet 34) – A deepfake-detection subnet providing open AI tools to identify synthetic media. BitMind offers a free AI Detector (web/app/extension) so users can instantly flag AI-generated images or videos; the subnet backend rewards miners for improving deepfake detection.

Protein Folding (Subnet 25) – A scientific computing subnet by Macrocosmos for distributed protein structure simulation. It uses industry-standard software (GROMACS) on Bittensor to find low-energy protein conformations. Since its June 2024 launch, it has completed hundreds of thousands of folding jobs.

nineteen ai (Subnet 19) – A decentralized inference subnet from Rayon Labs optimized for serving large AI models. It provides on-demand text/image inference through decentralized UIs and APIs, providing high throughput in serving open-source models at lower cost.

404-GEN (Subnet 17) – A decentralized 3D asset generation subnet that democratizes text-to-3D creation. Its mission is to “empower everyone to create immersive 3D worlds, games, and AR/VR/XR experiences” from text prompts. The subnet offers a one-click 3D model generator (and Unity integration).

TAOHash (Subnet 14) - rewards miners for directing their PoW hashrate (initially Bitcoin’s) to validators, creates a decentralized market where hashrate is exchanged for Alpha emissions.

This essay is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investments.