The team that built the largest decentralized AI model in history has walked out of the network that made it possible and they didn't leave quietly.
Sam Dare, founder of Covenant AI, on Friday published a public exit letter accusing Bittensor (TAO) co-founder Jacob Steeves of running a network that only appears to be decentralized.
Covenant AI then sold 37,000 TAO tokens, worth approximately $10.2 million, on the open market. TAO's price fell more than 20%, with over $10 million in long liquidations triggered, according to CoinMarketCap data.
"It is decentralization theatre," Dare wrote, as reported by Unchained Crypto.
The accusation was that Steeves had unilaterally suspended emissions to Covenant AI's subnets, cutting off the team's income and orchestrated token sales as financial pressure during an internal dispute.
The exit landed just 31 days after Covenant AI's defining achievement, which was training a 72-billion-parameter AI model distributed across more than 70 independent computers worldwide, with no central data centre and no institutional backing.
What Bittensor Is And Why The Industry Was Watching
Founded in 2019 by former Google engineer Jacob Steeves, Bittensor is a blockchain network designed to build an open market for artificial intelligence.
Its logic mirrors Bitcoin's: instead of rewarding miners for solving cryptographic puzzles, it pays contributors for producing and improving AI models. Validators judge the quality of that work. Both are compensated in TAO, the network's native token.
The network runs through "subnets," specialized mini-markets each focused on a distinct AI task, ranging from text generation to deepfake detection. As of April 2026, Bittensor operates 128 active subnets.
Grayscale has filed with the U.S. Securities and Exchange Commission to convert its Bittensor holdings into a spot ETF. The network's market capitalization sits at approximately $3.5 billion.
The appeal is straightforward: today, nearly all frontier AI is developed and controlled by a small number of corporations, OpenAI, Google DeepMind, Anthropic, Meta.
Bittensor's argument is that concentrating AI in this way creates monopoly risks and censorship vulnerabilities. By building AI development on a public blockchain, the network aims to turn AI into an open commodity.
The Milestone That Made Bittensor Famous
On March 10, 2026, Covenant AI, operating through Bittensor's Subnet 3, completed training of Covenant-72B, a 72-billion-parameter language model. The training was distributed across more than 70 independent nodes worldwide, with no whitelist, where anyone with sufficient hardware could participate. Weights and checkpoints were released publicly under an Apache licence.
On the MMLU benchmark, Covenant-72B scored 67.1 in zero-shot testing, outperforming Meta's LLaMA-2-70B and LLM360 K2.
Nvidia CEO Jensen Huang praised the achievement publicly. Investor Chamath Palihapitiya called it a landmark. TAO rallied approximately 90% through March.
Analysts compared it to DeepSeek's breakthrough moment for open-source AI.
The result was the first credible proof that large language model training could happen outside a corporate data centre- distributed, permissionless, and open.
The Governance Crisis Crypto's AI Sector Cannot Ignore
Dare's exit letter went beyond a personal grievance. It included a specific structural allegation that Bittensor's "triumvirate" governance model, officially three co-equal members of the Opentensor Foundation is effectively controlled by one person.
A site called Tao Papers, launched the same day as the exit letter, published on-chain forensics claiming that of 41 Bittensor network upgrades between 2023 and 2026, 38 were proposed, first-signed, and deployed from infrastructure controlled by Steeves.
The other two multisig signers co-signed within minutes, without documented public discussion.
How Bittensor's Governance Is structured
Officially, a "triumvirate" of three Opentensor Foundation members proposes upgrades. A "senate" of the 12 highest-staked validators then votes on those proposals. Critics, including Covenant AI, argue real decision-making authority has never meaningfully left the hands of the co-founder.
PANews described the problem as systemic, not isolated. "When a minority retains veto power over key permissions while claiming the system is 'permissionless,' the seeds of problems are sown." The analysis noted that once the decentralization premise is proven false, economic losses spread to every participant who built on that assumption.
For institutional investors, Grayscale's pending TAO ETF filing remains active as of April 12, 2026, the governance question is a material risk.
The investment case for any decentralized AI network depends on the premise that no single actor can unilaterally alter the rules. Evidence of concentrated control undermines that case directly.
Covenant AI has confirmed it will continue building decentralized AI infrastructure outside of Bittensor.
The technical proof that a 72-billion-parameter model can be trained across commodity hardware with no central coordinator, stands. The governance question does not.
Read Next: The Bill That Could Legally Unlock Trillions In Crypto Has Three Weeks Left To Survive