Every AI model you have interacted with was built inside a company. A team of engineers worked with a large budget and access to expensive computing infrastructure, producing a result that the corporation owns. That's the model behind ChatGPT, Gemini, and Claude.

Bittensor is building a different model. A marketplace where anyone in the world can contribute to building AI, get paid based on how good their work is, and where no single company owns the output or controls the infrastructure.

It passed my Gate and Key evaluation. Here's why, and here are the risks.

Bittensor is a network divided into specialized sections called subnets. Each subnet is focused on a specific type of AI task. One specializes in text generation. Another in computing provision. Others in data storage, financial forecasting, and fake media detection. There are 128 active subnets as of March 2026.

Within each subnet, two groups interact.

Contributors do the work. They provide computing power, submit AI outputs, and complete the tasks for which the subnet is designed. They earn TAO, Bittensor's native currency, based on their performance.

Validators evaluate the work. They are independent participants with domain expertise in whatever the subnet does. To participate, they stake their own TAO, meaning they lock up a portion of their holdings as a financial commitment to doing the job honestly. Score accurately and contribute to the network's quality to earn more TAO. Score poorly or try to game the system, and they lose what they staked.

What validators are actually evaluating depends on the subnet. In a text-generation subnet, they assess output quality and accuracy. In a financial prediction subnet, the outcome does most of the scoring itself — the prediction was either right or wrong. In a compute subnet, they verify whether a job was completed correctly and on time.

The staking requirement acts as a filter. Someone without real expertise in a subnet's domain is unlikely to stake significant TAO on their ability to score work they don't understand, because they would lose it. "The financial risk of staking creates a practical barrier to participation for people without relevant expertise.

The subnets that produce the most useful output attract the most resources. The ones that don't lose ground to competitors. No management structure decides what gets built or who succeeds.

Training an AI model means running enormous amounts of data through computing systems repeatedly until the model learns to recognize patterns and generate useful outputs. The most capable models today were trained using data centers that cost hundreds of millions of dollars to build and run.

On March 13, 2026, Bittensor's Subnet 3 completed a model called Covenant-72B. It was trained by more than 70 contributors using regular consumer hardware. No corporate data center. No large budget.

The model was tested on MMLU, a standard industry benchmark that measures knowledge and reasoning across 57 subjects. Covenant-72B scored 67.1. Meta's Llama 2 70B, trained with the full resources of one of the world's largest technology companies, scored 68.9.

Grayscale filed to convert its Bittensor Trust into a spot ETF. Grayscale is one of the largest digital asset managers in the world. They don't file for ETF conversion unless they believe regulatory approval is achievable and institutional demand exists. This is infrastructure being built around the asset.

The highest-revenue subnet raised $10.5 million from outside investors. Targon, a subnet serving enterprise clients who need private GPU compute, raised a Series A from institutional capital. Real dollars from conventional investors entering a specific subnet is a different signal than token price movement. It means someone independently underwrote the business case.

Jensen Huang, the CEO of Nvidia, said publicly that decentralized AI training works. On the All-In Podcast, investor Chamath Palihapitiya raised Bittensor's Covenant-72B training run directly with Huang. Huang agreed it was a significant technical accomplishment and said decentralized and traditional AI development are complementary approaches. NVIDIA makes the chips that power most AI training worldwide.

TAO rose 17% the day that the podcast aired.

Bittensor's annualized revenue is reported at $172.8 million. Investors are currently paying roughly $32 for every $1 of annual revenue the network generates.

Here is what that number actually means, and why it requires some care.

Bittensor pays its contributors in TAO tokens. When analysts report the network's revenue, they convert those tokens into dollars using today's token price. That means the revenue figure moves with the token price, not just with actual network activity. If TAO's price drops by half, the reported revenue drops by a similar amount, even if the same amount of work is being done on the network.

Think of it this way. Imagine you get paid in company stock instead of cash. If someone calculates your annual earnings using today's stock price, the number will look completely different depending on whether the stock is having a good day or a bad one. Your actual workload didn't change. The reported number did.

This points to a broader issue. The tools we use to measure business value were built for companies with employees, fixed costs, and cash flow. When contributors get paid in tokens and value is distributed automatically through code, those traditional measures start to break down. Better metrics for networks like Bittensor are still being developed.

The more reliable signal right now is Targon's $10.5 million Series A. That was raised in conventional dollars by outside investors who valued the business on its own merits, independent of the token's price. Targon raised $10.5 million from outside investors in a conventional Series A. That capital came in at a fixed valuation, in real dollars, from people making an independent bet on the business.

Regulatory classification is unresolved. The March 2026 ruling that classified 16 crypto assets as digital commodities did not include TAO. Classification as a security rather than a commodity would materially change who can own it and on what terms. The Grayscale ETF filing suggests the team believes a favorable outcome is achievable. That belief is not the same as the existing classification.

Subnet quality is uneven. The competitive model means that strong and weak subnets coexist. The network's value is concentrated in its best performers, and not all 128 subnets are producing equivalent output. As the network grows, the variance between the best and the rest will likely widen.

Governance is not fully transparent. The validator and contributor model distributes participation through financial incentives, which is well designed. How protocol-level decisions get made and who holds disproportionate influence over core development is less clearly documented. In early networks, governance that looks distributed on paper is often more concentrated in practice.

Bittensor passed my Gate and Key evaluation. It has a real use case, a proven technical milestone, institutional signals in the form of a Grayscale ETF filing and a conventional venture raise, and revenue that is meaningful even after accounting for the token-price caveat.

The two things worth watching before forming a high conviction are the regulatory classification and whether subnet revenue continues to grow independently of token price movements.

Twelve months ago, TAO was trading at higher prices with near-zero revenue. It survived the bear market and built a working infrastructure.

Bittensor is a network you can participate in directly, not just observe or invest in.

If you have domain expertise in a specific area and want to explore becoming a validator, the starting points are Bittensor's documentation at bittensor.com and their Discord, where each subnet has its own community. Targon has its own documentation and is actively building its contributor base, given its institutional backing.

The technical barrier varies by subnet. Some require more coding knowledge than others. The financial barrier is owning TAO to stake.

Next up: Render (RNDR).