Covenant AI Exits Bittensor Amid Governance Dispute

Original | Odaily Planet Daily (@OdailyChina)

Author | Azuma (@azuma_eth)

Do you remember the story of NVIDIA CEO Jensen Huang liking Bittensor (TAO)?

On March 20, while appearing on Chamath Palihapitiya’s All-In podcast, Huang was asked whether he was bullish on decentralized AI systems or computing networks. Palihapitiya cited Bittensor as an example (with a hint of bias), noting that a subnetwork team on Bittensor had successfully trained a Llama model with 4 billion parameters (actually 72 billion parameters) using distributed computing power. Upon hearing this, Huang remarked, “That’s an impressive technical achievement.”

Benefiting from this positive development, TAO surged against the market last month, reaching a high of over $370, and Bittensor has been regarded by the cryptocurrency industry as "the village's only hope."

Then, just half a month later, the situation took a sharp downturn due to an unexpected announcement—by the morning of April 10, TAO had fallen below $290, declining sharply for three consecutive days, and Bittensor found itself in what may be its largest public controversy since its inception.

Before explaining the details of the event, we need to first understand Bittensor’s subnet architecture.

Bittensor is a decentralized machine learning network centered on token incentives, using subnet mechanisms to enable different teams to build various AI task markets, with miners and validators jointly participating in computation and evaluation to distribute TAO rewards.

The "certain subnet team" mentioned earlier by Palihapitiya is actually called Covenant AI (formerly Templar), and the model liked by Huang is named Covenant-72B—a 72-billion-parameter model collaboratively trained in a permissionless manner by over 70 independent contributors on general-purpose hardware, and the largest decentralized large model pretraining project in history.

In simple terms, Bittensor serves as the underlying infrastructure for projects like Covenant AI, providing incentives, governance, and network rules—rather than directly developing specific AI models or applications. Meanwhile, subnets like Covenant AI act more like “application-layer builders,” offering concrete AI tasks and model capabilities on top of this foundational network.

On the morning of April 10, Covenant AI founder Sam Dare unexpectedly released statement (given TAO’s sustained decline, the underlying tensions may have been brewing longer), stating that Covenant AI had decided to exit the Bittensor network due to Bittensor and its founder Jacob Steeves (alias Const) violating decentralized principles.

Covenant AI stated that the team’s core belief is that “training cutting-edge AI models should not be controlled by any single entity,” but when a single actor can pause subnet emissions, override subnet owners’ authority over their own community spaces, publicly abandon projects without process, and use token sales as a coercive mechanism to force compliance, this is not decentralization—it is centralized control disguised as decentralization.

Covenant AI continues to allege that every participant in the Bittensor ecosystem—miners, validators, and investors—should understand that this power exists and has been exercised by Const. Const did not wield this power for the health of the network, but rather to reassert control over a team that had become “too independent” and difficult to manage—a subnet owner capable of building its own community, making independent decisions, and operating without permission, as this threatened Const’s authority across the entire ecosystem. Specifically, although Bittensor claims to operate under a “three-person custody” structure—where three individuals manage network upgrades via a multisig and present this as decentralized governance—the reality is otherwise. Const retains absolute control and resists any genuine transfer of power—the authority in the Bittensor ecosystem has never left one person’s hands.

Covenant AI also noted that, over the past few weeks, Const has taken a series of actions conflicting with Bittensor’s stated principles, including suspending emissions for the Covenant AI subnet, revoking the team’s administrative rights over their own community channels, unilaterally abandoning subnet infrastructure, and applying economic pressure through large-scale public token sales during periods of operational conflict.

Therefore, Covenant AI has decided to exit the Bittensor network. The team concluded that decentralized, permissionless AI training is not unique to Bittensor, but rather a technological capability Covenant AI remains committed to advancing. Covenant AI’s research, team, models, and vision will continue to move forward, with several highly exciting projects already in progress and set to be announced to the public soon.

Due to the success of Covenant-72B (SubNet-3), combined with Covenant AI’s operation of two other key subnets—Basilica (SubNet-39, positioned as an AI model evaluation and inference subnet) and Grail (SubNet-81, positioned as a more complex task-driven AI subnet)—the team holds a pivotal position within the Bittensor ecosystem. It may be precisely because of Covenant AI’s enhanced community presence, resources, and influence that the power struggle with Const emerged.

As the conflict between the two parties became public, the Bittensor ecosystem quickly descended into a storm of public opinion.

On the product level, with Covenant AI's departure, the community has begun to question the future development and value of the Bittensor network. As one of the most technically compelling and practically accomplished teams within the current Bittensor ecosystem, Covenant AI’s exit means this critical capability chain has been directly removed, creating uncertainty around Bittensor’s technical progress in AI model training and overall ecosystem vitality. As a result, the market has become more cautious in its assessment of its long-term value.

Reputation impact: Bittensor’s decentralized narrative is facing its biggest challenge since its inception. Covenant AI’s allegations directly target Bittensor’s core narrative—the “decentralized AI network.” For Bittensor, which relies on this decentralized story to attract developers and computing power participants, this governance controversy has implications far beyond short-term price fluctuations and could undermine the confidence of ecosystem participants.

At the brand level, Covenant AI has used this incident to overshadow Bittensor in the community’s perception. Before this statement, the market generally associated the “Old Huang like” incident with praise for Bittensor, and few realized that Covenant AI was the true protagonist—or even knew the team existed. As the event unfolded, Covenant AI’s visibility continued to grow, while Bittensor became perceived by the community as the one losing ground.

As of publication, Bittensor’s official social media channels have not made any statement, while Const responded vaguely on his personal account: “This will drive Bittensor to produce its first truly ‘headless’ (presumably not reliant on a single team) and truly commoditized subnets... Thank you to Covenant AI for making Bittensor more decentralized.”

Below Const's response, a large number of Bittensor community users (particularly TAO holders) are urging Const to provide a more detailed reply to the allegations made by Covenant AI, but Const has not yet responded further.

Odaily Planet Daily will continue to follow this matter—please stay tuned.