Cú Sốc Covenant AI: 3 Subnets Leave Bittensor and Reveal Cracks in the Decentralized Promise

On the morning of April 10, 2026, a lengthy article from Covenant AI, the operator of a subnet within the Bittensor ecosystem, quickly spread on Twitter. In the article, Covenant AI founder Sam Dare announced that their three subnets—Templar (SN3, pre-training large-scale models), Basilica (SN39, focused on computational services), and Grail (SN81, focused on reinforcement learning fine-tuning and model evaluation)—had all withdrawn from the Bittensor network. This announcement also raised the core question of whether Bittensor’s decentralization is merely a facade or not, becoming a focal point of attention. The market reaction was unmistakable: TAO plummeted over 15% within two hours of the announcement, briefly reaching 281 USDT. Of course, the token declines across all three Covenant AI subnets were even more severe. According to CoinGecko data, at the time of writing, SN3 had dropped more than 57% in 24 hours, while SN39 and SN81 had each fallen over 70%. The entire Bittensor subnet portfolio declined 26.5% in 24 hours. Considering that SN3 had hit a peak growth of 460% in March thanks to Covenant-72B’s achievement, this reversal was particularly brutal. But the withdrawal letter is only the tip of the iceberg. Two Years of Harvest, Then Suddenly Ending Covenant AI’s intelligence is not an unfamiliar entity. In March, they completed the industry’s largest decentralized pre-training project for LLMs in the crypto space to date on Bittensor—Covenant-72B, with over 70 peer nodes, about 20 synchronized nodes per round, each equipped with 72B200 tokens, and completed pre-training a 72-billion-parameter model on a dataset of approximately 1.1 trillion tokens. According to some benchmark evaluations provided by SN3, Covenant-72B’s score of 67.1 on MMLU is comparable to the 65.6 score of Meta’s LLaMA-2-70B announced in 2023. SN3 was praised by venture investor Chamath Palihapitiya, caught the attention of Nvidia CEO Jensen Huang, and was even prominently mentioned in Jack Clark’s AI research progress report. However, Clark’s assessment was quite cautious — “Covenant-72B might be somewhat outdated by 2026.” But in the crypto world, where stories are woven endlessly, this is an extremely rare form of technical confirmation. In their withdrawal statement, Covenant AI founder Sam Dare listed the achievements they carried: research, models, and even the team—all intact. The reason given was a list of accusations expressed in a restrained but clear manner, claiming that Bittensor is a centralized control operation disguised as decentralization. First point: The decentralization is suspicious. This directly targets the “three-signature multi-party governance” structure that Bittensor heavily promoted, claiming it’s nothing more than a “decentralization farce”—on the surface, decision-making rights are distributed among many parties, but in reality, Const (Jacob Steeves, co-founder of Bittensor), refuses any significant transfer of power and can unilaterally make changes without any consensus process. Second point: Cessation of emissions. Covenant states that Const directly halted the emission allocation for Covenant’s related AI subnets. Emissions are the main revenue source for Bittensor’s subnets, distributed to subnet owners, miners, and validators/endorsers. Halting emissions is equivalent to unilaterally cutting off the economic lifeline of an independent group and potential benefits for participants. Third point: Privilege stripping. Covenant’s community channel moderation function was forcibly removed, subnet infrastructure was unilaterally abandoned, and the space for negotiation was reduced to nearly zero. Fourth point: Token dump. More controversially, Covenant accused Const of selling off a large amount of tokens to exert market pressure amid conflicts over Covenant’s operations. Const’s Counterattack: Crisis as an Opportunity Const’s response was stronger than expected, but its direction was unexpected.

Instead of dismissing each of Dare’s accusations, he viewed the entire situation as an opportunity to develop governance. He said: “When dTAO was launched, there was a design idea that was put on hold: allowing alpha token holders of subnets to vote on-chain to elect a team that can truly promote subnet development and control hyperparameters. This mechanism was delayed because it needed to prioritize control for the initial subnet owners, but Const believes now is the time to restart this discussion, so the community can vote to select who will take over these subnets and operate them again.” As for Sam Dare, Const was blunt: “Sam clearly made this poor decision out of malice and greed.” This was almost the only direct description of the incident in his response. But he immediately changed tone, saying he had seen too many moments like this, and usually it’s not the end, but a new opportunity. For Bittensor, this is a new chapter in redefining ownership and subnet governance. Community: Two Factions, But Most Oppose Covenant The community quickly split into two camps. However, the majority opposed Covenant. The core criticism is straightforward: it’s hard not to see this as a “retreat tactic” when a development team announces withdrawal at a peak and reclaims subnet tokens. It’s a common practice for project teams to cash out and hand over control to the community. While Const’s response left many gaps for interpretation, other voices within the ecosystem were much more direct. An operator of another subnet project, Fish, revealed many foundational details never disclosed before: “Const personally built the initial versions of Grail and Templar, then directly handed these two subnets over to Sam to operate; he helped Sam find developers, raised startup funds for TAO, and even transferred ownership keys of Templar for free. Once, on Christmas, Const even gave Sam a substantial bonus of 2000 TAO. According to Fish, the direct cause of this rift was simply Const selling about 5% of his personal stake. Sam was extremely dissatisfied with this, ran away with all investors, spread rumors within the community, and behaved childishly when leaving all groups. And you really say this is Const’s fault?” Fish’s conclusion is simple and blunt: from start to finish, only Sam is at fault.

This information still needs verification, but it significantly shifts the focus of the entire event: from a story of “founder oppression of independent teams” to a personal rift. Some commentators pointed out that without Const’s initial support and emission assistance, Covenant would never have grown to its current scale. Additionally, Siam Kidd, CTO and co-founder of the Bittensor ecosystem hedge fund DSV Fund, directly described the incident: “In short, the owners of subnets SN3, 39, and 81 decided to sell and leave, that’s all.” He mentioned that the previously trained decentralized subnets by Nous Research also left the ecosystem in a similar way, and that Sam Dare’s departure this time seems to have been pre-planned or even “enticed.” Regarding Covenant’s accusations about decentralization, Siam Kidd admitted they are not entirely unfounded — “I also mentioned some of his statements about centralized control several times, and he’s not wrong” — but emphasized that Const is not power-hungry: “Const is not a megalomaniac unwilling to relinquish power. His motives are very sincere and straightforward. Objectively speaking, all actions Const takes serve only one purpose: to help Bittensor, nothing more.” As for the cause, Siam Kidd’s description is quite similar to Fish’s: Sam argued with someone on Discord and started deleting messages. Then, Const revoked Sam’s deletion rights, leading to conflict. Const then sold some of his alpha tokens, “and that’s it. Sam seemed to lose control over this trivial matter and used it as an excuse.” According to Jesus Martinez, an influential (crypto influencer) with 310,000 followers on Twitter, “Covenant AI founder Sam Dare sold 37,000 TAO tokens today, worth about $10 million. Maybe he’s just struggling to survive and looking for an exit.” Of course, we cannot verify whether that address belongs to Sam Dare at this time. Supporters of Covenant often focus on the core value of decentralization. They argue that if a protocol allows founders to punish non-cooperative groups by suspending emissions, it’s merely a technological veneer over a centralized platform. Structural Challenges of the Promise of Decentralization Bittensor’s design attempts to build a decentralized AI compute power marketplace using incentive mechanisms: miners contribute compute power, validators assess quality, specialized subnet groups focus on niche scenarios, and TAO tokens connect all these incentives. The model is theoretically consistent, but the problem is not just at the conceptual level. The issue lies in who decides the “value.” Const resigned as CEO of OpenTensor Foundation in February this year in an event dubbed the “Satoshi moment,” widely welcomed by the community and seen as a signal of active decentralization. However, Covenant’s accusations have created a fissure in this narrative: can someone who has resigned as CEO still exert decisive influence over the network? This is not an issue unique to Bittensor. Most protocols claiming to be “decentralized” face the same early-stage pressure: the technical influence of founders far exceeds what any governance document can restrict. Of course, as mentioned earlier, Const is on the right track to turn this crisis into an opportunity to promote a governance mechanism—“alpha holders vote to elect the team”—and he also announced that Bittensor will consider making subnet ownership dependent on the team’s long-term economic contribution to the project. This means investors could be aware of subnet unlock schedules and could redirect liquidity to support other teams or AI agents managing the system. However, whether this can be quickly implemented under public pressure and truly withstand the next test remains an open question to watch. What Does This Departure Mean? Some issues to continue monitoring: How to compensate for the 9% emission shortfall? The Bittensor Ecosystem Weekly account, @BittensorWeekly, released a special video after the incident, confirming basic information about the three subnets’ departure and the 9% emission loss. In the short term, this may attract other groups to compete for the vacated subnet positions, but doubts remain whether Covenant’s training data, model assets, and contributor network can be quickly replicated after their departure. What are Covenant’s new projects? Covenant’s withdrawal statement clearly states that their mission remains unchanged, and their research, models, and team will continue on a new platform. Where will they be based? Could this trigger a domino effect? The Bittensor ecosystem currently has over 120 active subnets, with Covenant being one of the most widely known. If governance disputes escalate, will other leading subnets act? How do institutional investors view this? Grayscale’s TAO ETF (is currently in registration), with institutional investment and market participation based on assumptions of steady growth within Bittensor. Public disclosure of the governance crisis might cause some institutions to delay their actions. How this crisis ultimately ends, whether Const’s proposed on-chain voting mechanism can be effectively implemented, and who the final token holders of the three subnets will vote for—these questions will determine whether Bittensor can turn this crisis into a true starting point for its decentralized story’s maturity.