Author: Max.S
The capital market's faith in "decentralized AI" (DeAI) is facing an unprecedented stress test.
Recently, Bittensor ($TAO), the undisputed leader in the decentralized AI field, has suffered a devastating internal upheaval. Covenant AI, one of the top development teams in the Bittensor ecosystem and the one that recently successfully trained a 72-byte language model, suddenly announced its complete withdrawal from the Bittensor network via social media. In its departure statement, Covenant AI directly criticized Bittensor founder Jacob Steeves, accusing him of having "absolute and dictatorial" control over the network, arbitrarily cutting off token rewards for subnets, and stating that the so-called decentralized AI is nothing more than a carefully orchestrated "drama."
Affected by this black swan event, the price of $TAO tokens experienced a panic sell-off in the secondary market, with a single-day drop of 15% to 25%, instantly wiping out hundreds of millions of dollars in market value. While the crypto community was "watching the drama" of the public break between the top team and the founder, it also began to seriously examine a deep-seated industry question: In the AI field, which is extremely dependent on computing power and complex engineering, is the "decentralization" driven by token economics a utopia that reshapes production relations, or just a glamorous cloak to cover up centralized power?
To understand the destructive power of this incident, we must first recognize the significance of Covenant AI within the Bittensor ecosystem.
In Bittensor's multi-subnet architecture, most subnets are still at the low-level API call, model fine-tuning, or simple task routing stage. Teams with the real ability to train from scratch or train large-scale parameter models are extremely rare. Covenant AI is a "hardcore" representative in this ecosystem. Shortly before announcing its exit, the team delivered a milestone achievement to the community: successfully training an open-source large model with 72 billion parameters (72B) in a decentralized network environment.
At current computing power costs, training a 72B model requires mobilizing a massive GPU cluster (typically equivalent to thousands of H100 GPUs running continuously for weeks) and incurring extremely high hardware and electricity costs. Covenant AI's willingness to bear such huge upfront sunk costs stems from Bittensor's "Emissions" mechanism—as long as its provided models and computing power achieve high scores in subnet evaluations, it can continuously receive $TAO tokens as substantial rewards. This is precisely the most attractive flywheel effect in the DeAI narrative.
However, the flywheel came to an abrupt end at its climax. According to Covenant AI, after investing heavily in training and launching its 72B model, founder Jacob Steeves and his stakeholders, by controlling the validators, directly cut off the token rewards flowing to the Covenant AI subnet without any warning or transparent governance process.
For miners and developers, cutting off Emissions is tantamount to "pulling the plug." The ROI of huge computing power expenditures instantly becomes zero. This extremely unpredictable systemic risk directly triggered Covenant AI's angry departure.
The word "charade" used by Covenant AI in its exit statement precisely struck Bittensor's most vulnerable nerve: network control.
Bittensor's underlying design relies on the Yuma consensus, which is based on "validators" evaluating the contributions of "miners" and determining how the newly issued $TAO tokens are distributed. Theoretically, this is a decentralized game system based on staking and algorithms. However, Covenant AI's accusations reveal a harsh reality: while computing power is decentralized, power and capital are highly concentrated.
In the current Bittensor root network, the top validator nodes that dominate token distribution have staked tokens highly concentrated in addresses associated with early investors, the foundation, and founder Jacob Steeves. This means that the founder is not only the rule-maker but also the ultimate referee.
Covenant AI points out that when a subnet's output does not align with Jacob's personal wishes or may threaten the interests of other "protégé" subnets, Jacob can easily use his massive staking weight to alter the distribution of Yuma consensus. This "one-man show" intervention renders decentralization at the smart contract level meaningless. Developers spend millions of dollars on computing power, yet their fate ultimately depends on the subjective will or behind-the-scenes manipulation of a single founder.
Objectively speaking, Jacob and his supporters might defend themselves by citing reasons such as "maintaining the overall quality of the network" and "preventing specific subnets from exploiting loopholes in the rules to generate cryptocurrency." However, in the absence of a transparent DAO governance mechanism and on-chain hearings and appeal channels, this centralized intervention, which acts as a "righteous act," severely undermines the network's core value as a "trusted and neutral infrastructure."
The 15-25% plunge in $TAO in a single day was not merely a stampede caused by panic among retail investors, but rather a repricing of Bittensor's "governance risk discount" by institutional funds.
Bittensor's massive market capitalization and high valuation premium stem from its being seen as the only realistic example of "decentralized OpenAI." This grand narrative is based on the system's inherent predictability: as long as you contribute computing power and high-quality models, the protocol automatically guarantees your rewards through code.
The Covenant AI incident shattered this expectation. Top financial professionals and institutional investors abhor "unpredictable single-node failures," and in this case, that failure point was Jacob Steeves' power.
If even the top teams capable of training 72B models can instantly lose everything due to founder interference, then for other computing power providers and AI research institutions holding tokens and waiting to see what happens, deploying heavy assets on Bittensor is undoubtedly a game of Russian roulette where the table could be overturned at any moment. When high-quality supply-side resources (miners and developers) refuse to enter due to fear of centralized tyranny, the application scenarios and intrinsic value of the $TAO token become unsustainable. The frantic exodus of funds is a preemptive vote against this deteriorating fundamental situation.
The departure of Covenant AI is not just a public relations crisis for Bittensor, but also an inevitable growing pain for the entire decentralized AI field as it enters deeper waters. It cruelly reveals to the industry the "impossible triangle" of the DeAI field: model quality and scale, decentralized trust and neutrality, and incentive alignment to prevent malicious behavior.
Centralized scale vs. decentralized mechanism: Training cutting-edge AI (such as large models of 72 bytes or more) is a typical capital-intensive, centralized project that requires highly collaborative GPU clusters. This is inherently physically different from the permissionless, distributed nodes advocated by Web3.
Anti-spam measures vs. Trusted neutrality: To prevent low-quality nodes from fraudulently obtaining tokens through mutual traffic boosting (a Sybil attack), the network must introduce subjective "quality assessment." However, given that AI assessment standards are not yet fully objective and mathematically quantifiable, entrusting this assessment power to a small number of validators can easily degenerate into centralized rent-seeking.
Bittensor attempted to bridge the gap between the two using token economics, but the Covenant incident proved that the supporting pillars of this bridge (governance mechanisms) remain extremely fragile.
Covenant AI's departure punctured the romantic bubble of Bittensor's "absolute decentralization." For $TAO, this may be a painful moment of disenchantment, but for the entire DeAI industry, it serves as a necessary wake-up call.
