Bittensor, a decentralized AI network long associated with crypto-native circles, has moved further into the mainstream conversation after receiving high-profile attention from Chamath Palihapitiya and Nvidia CEO Jensen Huang. The discussion has renewed interest in whether large-scale AI training can happen outside traditional, centralized data-center structures.
A crypto-native AI project breaks into the broader tech debate
During an episode of the All-In podcast, investor Chamath Palihapitiya pointed to Bittensor’s Covenant-72B as a concrete example of decentralized AI progressing beyond theory. He described the broader idea in simple terms: a large language model trained without relying on a single centralized infrastructure stack, but instead powered by a distributed set of independent contributors offering spare compute.
Palihapitiya referenced an earlier achievement involving a 4 billion-parameter LLaMA model trained in a fully distributed way. He framed the accomplishment as technically remarkable, comparing it to classic distributed-computing efforts that pooled idle hardware from many participants around the world. That analogy matters because it helps explain why Bittensor’s approach stands out: it tries to turn fragmented, otherwise underused compute resources into something capable of supporting advanced AI development.
Bittensor operates as a blockchain-based decentralized network designed to create a peer-to-peer marketplace for machine-learning models and AI compute. In that framework, participants are incentivized to contribute resources and capabilities, while the network coordinates how value is exchanged. For supporters, this represents more than a crypto experiment; it is an alternative architecture for building and training AI systems.
Huang argues the future is not open versus proprietary
Jensen Huang did not dismiss the idea of decentralized AI. Instead, he framed the issue in broader market terms, arguing that open and proprietary AI are not mutually exclusive paths. In his words, the relationship is not “A or B,” but “A and B.”
That distinction is important. The current AI landscape is often described as a split between polished, closed systems such as ChatGPT, Claude, and Gemini, and a rising ecosystem of open or decentralized models that allow developers, enterprises, and communities to modify systems for specialized uses. Huang’s comments suggest that both sides of that divide are likely to remain relevant.
He emphasized that models are a technology rather than a finished product, implying that most mainstream users will continue to depend on highly refined, general-purpose systems instead of building their own from scratch. At the same time, he identified sectors where customization and control are essential rather than optional. In those environments, domain knowledge, internal workflows, and governance requirements need to be embedded into the model itself, and Huang said that capability can only come from open models.
That point aligns closely with Bittensor’s value proposition. If some industries need AI systems they can shape, inspect, and control more directly, then decentralized and open training methods may become more relevant, not less.
Covenant-72B highlights the scale decentralized training may reach
According to the report, Covenant-72B was developed through Bittensor’s Subnet 3, also known as Templar. It represents one of the largest decentralized training efforts reported so far. The project coordinated more than 70 contributors over standard internet connections, without a central authority overseeing the training process in the traditional sense.
On the technical side, the model is said to contain 72 billion parameters and to have been trained on approximately 1.1 trillion tokens. To make such a process viable outside conventional data-center environments, the effort relied on innovations including compressed communication protocols and distributed data parallelism. These design choices are central to the claim that large models may not need to be confined to tightly controlled centralized infrastructure.
The article also notes that benchmark performance places Covenant-72B alongside established centralized models. While the report does not provide full benchmark tables in the excerpt, the implication is clear: this is being presented not as a curiosity, but as evidence that decentralized training can produce competitive results at meaningful scale.
Why the market reacted
The market appeared to take notice quickly. Following the circulation of the Palihapitiya and Huang video on social media, Bittensor’s token, TAO, rose 24%. The move suggests investors interpreted the public discussion as a validation event for the decentralized AI narrative, especially because it involved Huang, whose views carry weight across both AI infrastructure and semiconductor markets.
Still, the price reaction may be the least important part of the story. The bigger development is that decentralized AI training is being discussed by top-tier technology figures not as an eccentric side path, but as part of the broader architecture of the AI industry. Public acknowledgment from Nvidia’s chief executive does not automatically guarantee adoption, but it does lend credibility to the idea that distributed training could occupy a real niche in the market.
A hybrid AI landscape is taking shape
The broader takeaway from Huang’s comments is coexistence rather than replacement. Proprietary systems are likely to remain dominant for mass-market usage because they are polished, easy to deploy, and optimized for general consumers and businesses. Open and decentralized models, however, may continue gaining traction in specialized, cost-sensitive, or sovereignty-driven applications.
That vision also fits the strategy Huang described for startups. He said many of the companies receiving investment today are pursuing an “open-source first, then proprietary” path. In practical terms, that suggests openness may increasingly serve as the starting point for innovation, experimentation, and adoption, even if some firms later build proprietary layers, products, or services on top.
For Bittensor, this is where the significance of Covenant-72B extends beyond technical bragging rights. The project offers a live example of how distributed contributors and incentive structures can be organized to train large models without relying entirely on centralized operators. Even if this approach does not replace hyperscale AI labs, it may broaden the set of viable ways AI systems are built.
What this means for crypto and AI
For the crypto sector, Bittensor’s visibility matters because it gives one of the industry’s long-standing narratives—decentralized infrastructure—a concrete AI use case that can be discussed in mainstream technical terms. Instead of focusing only on token incentives or ideology, the conversation shifts toward performance, scalability, and infrastructure design.
For the AI sector, the development raises a more practical question: how much of model training truly requires the traditional centralized stack, and how much could eventually be distributed if networking, coordination, and incentive mechanisms improve? Covenant-72B does not settle that debate, but it adds evidence that the answer may be more flexible than previously assumed.
In that sense, the real significance of the moment is not that decentralized AI has already won, but that it is increasingly difficult to ignore. With high-profile endorsements bringing projects like Bittensor into the spotlight, the future of AI may look less like a single dominant model of infrastructure and more like a layered ecosystem in which centralized and decentralized approaches serve different needs.