Is Bittensor the holy grail of decentralized AI or a castle in the air built on subsidies?

Author: Clow,Plain Language Blockchain

In March 2026, Nvidia CEO Jensen Huang said something on the "All-In" podcast that immediately sent a crypto project trending.

He said Bittensor is a "modern version of Folding@home."

How much is that statement worth?

The price of TAO surged from $243 before the podcast to $365, a monthly increase of over 100%. Open interest in futures contracts ballooned to $639 million, compared to less than $132 million at the beginning of March. Grayscale had already submitted its S-1 application for a TAO spot ETF to the SEC at the end of December 2025, with Bitwise following suit on the same day. But what truly unsettled those in the industry wasn't the price, but something that happened on March 10th. Bittensor's subnet 3 (Templar) announced that it had trained a 72 billion-parameter language model using over 70 ordinary computers scattered around the world, connected via an average 500 Mb/s home broadband connection. This model, called Covenant-72B, achieved a score of 67.1 on the MMLU benchmark—approaching or even surpassing Meta's LLaMA-2-70B, trained using a cluster of supercomputers. A group of individual users' graphics cards trained a large-scale model that rivals Silicon Valley giants. Even more bizarrely, no one knows where these nodes are located. They might be in a rented apartment in Tokyo, in a hacker's basement in Berlin, or even in your next-door neighbor's house. All contribution records and reward distributions are written on the Bittensor blockchain—transparent, immutable, and permissionless. This isn't just a story of "cheap computing power." It's a war about "who has the right to train AI." What is it doing—128 smart factories operating simultaneously? The success of Covenant-72B isn't based on brute force, but on an algorithm called SparseLoCo. In short, this algorithm compresses the training data that needs to be transmitted between nodes by 146 times, achieving a compression rate of over 97%, while maintaining almost no loss of model accuracy. This means you no longer need a multi-billion dollar supercomputing center like Nvidia's; a 500 Mbps home broadband connection is sufficient. But Bittensor does much more than just train a single model. It currently has 128 concurrently running subnets, each performing a different task. You can think of the entire network as a giant industrial park: Factory 3 (Templar) is responsible for training large models; Factory 64 (Chutes) provides inference services, reportedly with over 400,000 users and processing 5 million requests daily, at only 15% of AWS's price; Factory 4 (Targon) handles confidential GPU computing, and Intel engineers even co-authored a technical white paper on trusted execution environments with it—not just a business collaboration with a logo, but two Intel engineers who actually put their names on it; Factory 51 (Lium) is the GPU P2P market, reportedly managing a cluster of over 500 H100 computing power. The head effect is very obvious—the top ten subnets take 56% of the network's emissions rewards. But the economic mechanism itself is even more noteworthy. After November 2025, Bittensor launched the Taoflow model: the amount of TAO rewards each subnet receives is no longer determined by a vote of 64 validators, but directly by the flow of market funds. Rewards go to the subnets where funds flow. An 86.8-day exponential moving average automatically adjusts, and subnets with continuous net outflows will be "starved." Even more ingenious is the dTAO mechanism—each subnet now has its own Alpha token. Investors stake TAO to obtain Alpha from specific subnets, essentially betting on a specific vertical direction. According to v1 research data, as of March, the total market capitalization of all subnet Alpha tokens was approximately $1.12 billion, accounting for about 27% of TAO's own market capitalization. In simpler terms: Bittensor is not an AI product, but a smart production protocol that allows idle GPUs worldwide to automatically team up, compete, and be eliminated. Why it matters—not "cheap," but "sovereignty." On the surface, Bittensor's selling point is that "decentralized computing power is cheaper than centralized power." But this claim doesn't hold up to scrutiny—we'll explain why later. The core of what Jensen Huang said in his podcast wasn't actually price. He said the future of AI isn't about OpenAI dominating, but rather "A and B coexisting": proprietary models and open-source/distributed models will run in parallel for a long time. The reason is simple—fields like healthcare, defense, and manufacturing require complete control over AI models, and this control can only come from open-source or distributed architectures. At the GTC conference, Jensen Huang proposed the concept of a "token factory"—in the future, not only text will be tokenized, but protein structures, robot movements, and physical simulations will also be tokenized. Every scientific field will need its own "intelligence generator." Bittensor's 128 subnets are precisely the prototype of this vision. But a deeper reason is: what if US regulations suddenly tighten, what if a country blocks OpenAI's API, what if your industry is prohibited from using certain closed-source models—what will you do? Bittensor provides an AI infrastructure that is "censorship-resistant." Model weights are open source, nodes are globally distributed, and no single point can be shut down. When Intel engineers chose to co-publish the white paper with Targon, and when early Uber investor Jason Calacanis predicted TAO's 200x growth potential and specifically established the Stillcore Capital fund to bet on TAO and its subnet tokens, they weren't attracted by its "cheapness." They were attracted by the fact that this might be the only AI network that can still function under extreme circumstances. What Bittensor is really selling isn't computing power, it's sovereignty. Doesn't anyone see a problem with this? Now let's look at the other side of the coin. TAO currently has a market capitalization of approximately $3.9 billion. But its most successful subnet, Chutes, only generates $1.3 million to $2.4 million in annualized external revenue. Meanwhile, Chutes receives $52 million worth of TAO subsidies annually from the protocol. Let's calculate the subsidy ratio: 22:1 to 40:1. For every $1 a user pays, the network issues $22 to $40 worth of TAO through inflation to subsidize miners. If these subsidies were removed, Chutes' claim of "15% of the AWS price" would instantly reverse to "1.6 to 3.5 times the AWS price." The network's externally confirmed annual revenue is between $3 million and $15 million, corresponding to a revenue multiple of 175 to 400 times. This figure for traditional high-growth SaaS companies typically doesn't exceed 50 times. So why is the market buying in? Three things: First, TAO's total supply cap of 21 million tokens and its halving mechanism make it as scarce as Bitcoin; second, the expectation of a Grayscale ETF, which, if approved, will open up channels for institutional funds; and third, the narrative itself of being "the only decentralized AI protocol capable of running large models." On-chain data also tells a bullish story. Exchanges saw a net outflow of $5.77 million on March 21-22—large investors were accumulating tokens. According to OnchainLens monitoring, the associated address of well-known trader jez established a 5x leveraged long position in TAO on March 25, with a notional value exceeding $2.7 million. The MACD turned positive in March, and short sellers are being squeezed. However, before the second halving at the end of 2026, if the external revenue of subnets does not increase significantly, the exit of miners could lead to a substantial reduction in hashrate. Bittensor's high beta also means that if the Federal Reserve continues its hawkish stance or the situation in the Middle East spirals out of control, institutional funds will withdraw from these overvalued assets immediately. In summary, Bittensor is either building a decentralized AI "intelligent internet" or constructing a beautiful castle in the air using inflation. The difference between these two outcomes may only be a halving cycle. Jensen Huang said that future factories will produce tokens. But factories will also go bankrupt—especially those factories that can't sell their products and rely entirely on subsidies. Of TAO's $365, how much is the price of intelligence, and how much is the premium for faith? Nobody knows for sure right now.