Bittensor's subnet SN3 has achieved a milestone by training a 72-billion-parameter AI model, Covenant-72B, using a fully decentralized computing network. The model's performance is comparable to Meta's LLaMA-2, but it was trained without centralized data centers, relying instead on over 70 independent nodes contributed by the community.
A Decentralized Training Breakthrough
The achievement has been likened to distributed computing projects like Folding@home, but applied to Large Language Models. Industry figures such as NVIDIA CEO Jensen Huang and Anthropic co-founder Jack Clark have acknowledged the significance. Covenant-72B demonstrates that permissionless networks can compete with centralized hyperscalers in training cutting-edge AI, opening up new possibilities for censorship-resistant and accessible AI development.
Market Rally and Token Performance
Following the announcement, SN3 token surged over 440% in the past 30 days, reaching a market capitalization of $130 million. The native token of the Bittensor ecosystem, TAO, more than doubled to a peak of $377. Crypto derivative activity also spiked, reflecting strong investor enthusiasm for decentralized AI narratives. The rally underscores the market's appetite for projects bridging blockchain and artificial intelligence.
Challenges Ahead: Security and Regulation
While the successful training marks a technical triumph, questions around data sovereignty, model integrity, and regulatory compliance remain unresolved. Decentralized training involves relying on untrusted nodes, which could introduce risks of adversarial inputs or intellectual property leakage. In heavily regulated sectors like finance and healthcare, these issues may hinder adoption. Meanwhile, traditional cloud providers continue to invest aggressively in AI infrastructure, posing a competitive threat.
Overall, Bittensor SN3's achievement highlights the promise of decentralized AI, but its long-term viability will depend on strengthening security guarantees and building a sustainable ecosystem of contributors.