What was once a niche experiment confined to cryptocurrency circles has just received a powerful nod from Nvidia CEO Jensen Huang. On a recent episode of the All-In podcast, venture capitalist Chamath Palihapitiya joined Huang to discuss Bittensor's Covenant-72B, a large language model (LLM) trained entirely through a decentralized network of independent contributors. The conversation signals that decentralized AI training may be moving closer to the mainstream.
Palihapitiya: A Technical Marvel of Distributed Training
Palihapitiya described Covenant-72B as a groundbreaking achievement. The model features 72 billion parameters and was trained on approximately 1.1 trillion tokens without any centralized cloud infrastructure. Instead, over 70 individual contributors pooled their excess computing power via standard internet connections. Palihapitiya likened the effort to "a cryptocurrency version of SETI@home," where each person contributes a small piece to create a powerful whole. He called it "a pretty mind-blowing technical feat."
Jensen Huang: Open Source and Proprietary Are Both Essential
Rather than dismissing decentralized AI, Huang embraced it as part of a broader ecosystem. "These two things are not A or B; they are A and B," Huang stated. He argued that most users will continue to rely on polished, closed-source systems like ChatGPT, Claude, or Gemini for general use. However, in specialized sectors such as healthcare, finance, and legal, organizations need open models they can fully control and customize. Huang revealed that many startups Nvidia is backing follow a "first open-source, then proprietary" trajectory, underscoring the synergy between the two approaches.
Technical Breakthrough and Market Reaction
Developed under Bittensor's Subnet 3 (Templar), Covenant-72B leverages advanced techniques such as compressed communication protocols and distributed data parallelism. These innovations enable efficient training outside traditional data centers. Benchmark results show that the model's performance is on par with equivalently sized centralized models, proving that decentralized training can yield competitive results.
The news sent Bittensor's native token TAO soaring by 24% within 24 hours, reflecting growing investor confidence in decentralized AI infrastructure. Bittensor operates as a blockchain-powered decentralized marketplace where machine learning models and computing power are exchanged and incentivized peer-to-peer.
Future Outlook: A Hybrid AI Landscape
Huang emphasized that the future of AI will not belong to a single architecture or philosophy. Proprietary systems will dominate general consumption, while open, decentralized models will carve out niches in cost-sensitive, sovereign, or highly specialized applications. For startups, Huang recommended a pragmatic path: start open to build community and then layer proprietary advantages. Bittensor's achievement is a testament that blockchain and AI can converge to create scalable, decentralized infrastructure. With Nvidia's public endorsement, decentralized AI may finally gain the credibility needed to attract broader adoption.