Bittensor, a decentralized AI network long discussed mainly within crypto-native circles, has gained wider attention after public remarks from Nvidia CEO Jensen Huang helped push the topic closer to the mainstream. The conversation centered on Covenant-72B, a large language model trained through distributed collaboration rather than conventional centralized infrastructure, highlighting the growing credibility of decentralized approaches to AI development.
A Crypto-Native AI Experiment Reaches a Broader Audience
The renewed attention followed comments by Chamath Palihapitiya on the All-In podcast, where he presented Bittensor’s Covenant-72B as a real-world example of decentralized artificial intelligence moving beyond theory. Bittensor operates as a blockchain-based decentralized network that creates a peer-to-peer market for machine learning models and AI compute, using incentives to coordinate participants.
Palihapitiya framed the idea in accessible terms: instead of relying on a single company’s data center, a large model can be trained by a distributed group of independent contributors who supply excess computing power. In his remarks, he referenced a 4 billion-parameter LLaMA model trained in a fully distributed manner and described the result as a striking technical achievement. The comparison suggested a familiar distributed-computing logic, where many participants each contribute a small share of work to produce a larger outcome.
Jensen Huang Sees Open and Proprietary AI as Complementary
What made the discussion notable was not only Palihapitiya’s enthusiasm, but Huang’s response. Rather than dismissing decentralized AI as a fringe alternative, the Nvidia chief executive argued that the future of AI is unlikely to be defined by a winner-take-all choice between open and proprietary systems. In his view, the relationship is not “A or B,” but “A and B.”
That framing matters because it places decentralized and open AI models inside a broader market structure rather than outside it. According to Huang’s comments, highly polished closed systems such as ChatGPT, Claude, and Gemini are likely to remain central for mainstream users seeking convenience and general-purpose performance. At the same time, open models are essential in sectors where control, customization, and domain-specific expertise are not optional.
Huang’s logic rests on a distinction between models as technology and products as end-user experiences. Most users, he suggested, will continue to rely on mature general-purpose systems instead of building their own stack from scratch. But many industries have requirements that extend beyond convenience. Organizations in those areas need systems that can reflect proprietary knowledge, regulatory constraints, or internal operational logic in ways they can directly manage. In Huang’s words, that capability can only come from open models.
Covenant-72B Pushes the Boundaries of Distributed Training
That argument aligns closely with what Bittensor is trying to prove. Covenant-72B, developed through Bittensor’s Subnet 3, known as Templar, represents one of the largest decentralized training efforts disclosed so far. The project reportedly coordinated more than 70 contributors using ordinary internet connections and no central authority, a setup that directly challenges assumptions that frontier-scale model training must happen inside tightly controlled data-center environments.
On the technical side, the model was built with 72 billion parameters and trained on roughly 1.1 trillion tokens. To make that possible outside traditional centralized infrastructure, the project relied on methods such as compressed communication protocols and distributed data parallelism. These techniques are important because decentralized training is not simply a question of gathering enough hardware. It also requires reducing coordination overhead and making geographically dispersed compute resources behave efficiently enough to support large-scale model development.
The significance of Covenant-72B, based on the source material, is that it appears to move decentralized AI beyond a conceptual or ideological proposition. Performance benchmarks reportedly place it alongside established centralized models, suggesting that distributed training can produce outcomes that are competitive rather than merely experimental. That is a key reason why the project’s visibility has expanded beyond the crypto community.
Market Reaction Lifts TAO
The market responded quickly to the increased attention. After video clips featuring Palihapitiya and Huang circulated on social media, TAO, Bittensor’s token, rose 24%. While token price action does not by itself validate a technology thesis, the move reflects how closely investors are watching the intersection of AI infrastructure, open-source development, and blockchain-based incentive systems.
The reaction also underscores a broader theme in crypto markets: narratives that connect digital assets to fast-growing sectors such as AI can rapidly reprice expectations. In this case, the catalyst was not a new token launch or a traditional crypto milestone, but the perception that a decentralized AI training model had earned recognition from one of the most influential executives in the semiconductor and AI ecosystem.
A Hybrid AI Future May Be Taking Shape
Huang’s comments point toward coexistence rather than disruption as the more plausible long-term story. Centralized proprietary AI systems are still likely to dominate broad consumer usage because they offer simplicity, support, and highly refined user experiences. Open and decentralized models, however, may carve out a durable role in specialized applications, cost-sensitive deployments, and settings where sovereignty over data, model behavior, or infrastructure matters deeply.
That balance has important implications for startups and developers. Huang reportedly described a practical pattern among companies Nvidia is backing: many begin with an open-source-first approach and later add proprietary advantages. The suggestion is that openness may accelerate experimentation, adoption, and ecosystem growth, while proprietary layers can later help define business models and defensibility.
If that view proves correct, the future of AI will not belong exclusively to one architecture, one ideology, or one business model. Instead, it may favor builders who understand when centralized systems are the right tool and when open or decentralized alternatives offer a better fit. Bittensor’s latest milestone does not settle that debate, but it does strengthen the case that decentralized model training is becoming a serious part of it.
Why This Matters Beyond Crypto
The broader relevance of the Bittensor story is that it challenges one of the strongest assumptions in AI: that only massive centralized operators can train meaningful large-scale models. Even if centralized players remain dominant, successful distributed experiments widen the design space for the industry. They suggest a future in which not all important AI development has to be concentrated in a handful of corporate environments.
For crypto observers, the development offers a more substantive AI-blockchain narrative than many short-lived market trends. For AI observers, it raises a practical question: if open and decentralized methods continue to improve, what kinds of models and workloads might eventually migrate away from conventional data-center structures? The answer remains uncertain, but Bittensor’s progress has made that question harder to ignore.