NVIDIA CEO Jensen Huang has publicly praised decentralized AI network Bittensor in a recent appearance on the All-In podcast, drawing fresh attention to the project and its broader ecosystem. After Bittensor was brought up during the discussion, Huang said Subnet 3 had trained a large language model in a fully decentralized manner, with participants contributing idle compute resources and coordinating the training process statefully. He described the result as a remarkable technical achievement.
Huang’s description of the model size was not fully aligned with the reported details. In the podcast, he referred to a 4 billion-parameter Llama model. However, the model completed by Bittensor Subnet 3 was reported as Covenant-72B, a model with 72 billion parameters. It is described as the largest decentralized pretraining effort on the network so far, with an MMLU score of 67.1.
Markets read the comments as support for decentralized AI
In the same episode, Huang also pushed back on the common framing of open-source versus closed-source AI as a binary choice. In his words, the future is not “A or B,” but “A and B,” suggesting that decentralized infrastructure and proprietary models can coexist rather than compete in a zero-sum way.
That view was quickly reflected in market action. Following the release of the video, TAO rose about 20% within 24 hours. At the time of reporting, the token was trading at around $300, while daily trading volume climbed to $471 million, marking a recent high.
Bittensor ecosystem momentum builds further
Bittensor had already been gaining momentum before Huang’s remarks. The report noted that YZi Labs, linked to former Binance CEO Changpeng Zhao, has invested in Tensorplex, a project within the TAO ecosystem. A related subnet token reportedly surged by more than 2x at one point. Against that backdrop, Huang’s public endorsement has added further visibility to Bittensor and strengthened market interest in the viability of decentralized AI training.