Covenant AI's Exit from Bittensor: An Alarm Bell Exposing the Industry's "Pseudo-Decentralization"

Author: David & Daniil Liberman | Co-founders of the Gonka Protocol

Compiled by: Gonka.ai 

Preface: When a decentralized AI team widely recognized for its technical prowess publicly exits a network, explicitly calling it centralized control disguised as decentralization, the entire industry should not simply dismiss it as a bilateral dispute. This is a diagnosis—a wake-up call revealing the structural issues plaguing the entire decentralized AI ecosystem. And this diagnosis extends far beyond Bittensor.

1. What Exactly Happened?

Covenant AI (formerly Templar) spent over two years building what is arguably the most technically significant milestone in decentralized AI to date: Covenant-72B—a language model with 72 billion parameters. This model was trained permissionlessly by over 70 independent contributors on commodity hardware.

This is not a proof-of-concept; it is a production-ready breakthrough. This achievement received public recognition from NVIDIA's CEO on the "All-In" podcast and was cited by an Anthropic co-founder. It once drove Bittensor's TAO token up by 90%, with multiple subnet valuations approaching $1.5 billion.

However, following this prosperity, the very infrastructure that supported it turned into a weapon used against them.

According to Covenant AI's public statement, Bittensor founder Jacob Steeves (alias Const) unilaterally took several actions:

• Suspended token rewards for the Covenant subnet;
• Revoked their administrative rights to community channels;
• Decommissioned their subnet infrastructure without consultation;
• Used token sales as leverage during operational conflicts.

Covenant AI pointedly stated: This is "centralized control, disguised as decentralization."

Within hours of the announcement, the TAO token plummeted over 15%. As of writing, it was still down about 9% for the day.

In response, Steeves stated that Bittensor is about to launch truly independently operated subnets. But this response precisely confirms, rather than refutes, the core issue: when Covenant was fully committed to building, the "independence" they needed was not yet architecturally guaranteed.

2. The Structural Trap: The Value-Hostage Dilemma

To understand why this incident goes far beyond an isolated case, we must see the unique economic structure behind it—the very structure that makes such crises both dangerous and predictable.

Decentralized AI networks face a fundamental "cold start problem": building real infrastructure—training tasks, model weights, contributor networks, community trust—requires long-term investment spanning months or even years. This investment comes at the cost of time, capital, and reputation. But the tokens that ultimately capture this value belong to the entire network, not the builders themselves.

This creates a dynamic we term the "value-hostage" mechanism:

The better you build, the more the network's token appreciates; the more the token appreciates, the greater the leverage those in control have over you. And the moment your achievement is most glorious is also when you are most vulnerable.

Creating value on someone else's network means the value you create can ultimately be weaponized against you. The more successful you are, the more you stand to lose.

This is not merely a governance failure unique to Bittensor, but a structural consequence of any system where a few retain veto power over critical permissions (like token distribution, content moderation, infrastructure upgrades) while proclaiming the system is "permissionless." The problem is already embedded.

The promise of "decentralization" is the foundational premise upon which the system operates—builders, miners, validators, and investors all make decisions based on it. Once this premise is proven false or only "conditionally true," the resulting economic damage extends not just to the involved team, but to all participants who trusted that narrative.

Covenant AI pointed out that Bittensor's governance structure is nominally a "three-party co-management" multisig mechanism, but is in reality dominated by one individual, with the other two acting more as legal firewalls than genuine decision-making participants in governance.

We cannot independently verify every allegation. But we can see the underlying structural logic:

• A multisig dominated by a single individual is no different from a single private key;
• Governance processes that can be unilaterally bypassed cannot be called governance;
• A token distribution mechanism that can be paused by one person is, in essence, a subsidy, not a protocol guarantee.

3. This is Not a Bittensor Problem, It's an Industry-Wide Disease

Some may interpret the Covenant AI incident as a "cautionary tale about Bittensor." This view is too narrow and even misleading.

The deeper issue is this: the entire decentralized AI field has long maintained a fragile balance between two difficult-to-reconcile goals—rapid iteration and genuine decentralization.

Rapid iteration requires decisive leadership; genuine decentralization means no single party can unilaterally decide everything. Most projects default to the former: publicly preaching decentralization while privately holding tight control, hoping the contradiction is never exposed.

Now, it has been exposed.

A familiar pattern repeatedly emerges in the current decentralized AI ecosystem:  

A decentralized shell—token distribution, community forums, governance proposals;  

Wrapped around a centralized core—the founding team or foundation firmly controls the most critical parameters:  

• Token release schedule  
• Protocol upgrade authority  
• Subnet admission mechanisms  
• Community management rights  

This is not necessarily born of malice. Early networks indeed require strong coordination, and pure on-chain governance for complex AI infrastructure remains technically unsolved. But the gap between promotional promises and actual power creates a form of "structural debt."

Covenant AI's experience is what that debt looks like when it comes due.

In AI networks, this mechanism is more profoundly dangerous than in DeFi or Layer1, because the depth of builder investment far exceeds previous paradigms.

Training a 72-billion-parameter model is not a two-week sprint; it's a costly, time-consuming, reputationally risky long-term campaign. By the time Covenant AI realized the severity of the governance issues, they had already completed the core work—which is precisely why they became a target.

This asymmetry is brutally cruel:  

• The network can act at any time;  
• The builder cannot "withdraw" the work already completed.

As long as the following three conditions are met, similar incidents will recur:

1. Builders must invest significant time and resources;
2. Tokens can capture the value of this investment;
3. A small group can still unilaterally intervene in governance.

And these three conditions are extremely common in current AI networks.  

Therefore, we should not expect "the next Covenant AI" to be fewer—we should anticipate there will be more.

4. What Does Genuine Decentralization Require?

Faced with such events, the industry often falls into two extreme reactions:  

One is uncritically praising any project that calls itself "decentralized";  

The other is completely dismissing the feasibility of decentralization, viewing it as a scam.

Both are misguided.

We want to be clear: achieving genuine decentralization in AI infrastructure is technically extremely difficult. Anyone who tells you it's "easy" has likely never seriously attempted it.  

• Cross-geography node collaborative training  
• Verifiability of computational work  
• Manipulation-resistant token incentive mechanisms  

These remain unsolved or partially solved challenges. An honest industry should acknowledge this.

But technical difficulty does not equate to impossibility, nor can it serve as an excuse for misleading宣传.

We propose a simple yet sharp standard: Can the infrastructure you rely on be used against you?

If the answer is "yes," then regardless of how the whitepaper describes it, regardless of whether there are governance votes, that decentralization is superficial.

This question cuts through all the noise. It doesn't care if you have a DAO, nor does it look at your community forums. It asks one practical question: When conflict arises, can a small group of people unilaterally pause your rewards, cut off your access, or use economic means to coerce you?

If yes, then the claim of "decentralization," however sincere the original intent, lacks structural support.

The real way forward is to embed the constraints of decentralization into the protocol layer, not rely on the "goodwill" of any individual or group.

This means:

• Token distribution mechanisms are determined by on-chain rules and cannot be unilaterally paused;
• Protocol upgrades require genuine multi-party consensus, not formalism;
• In governance structures, actual contributors (builders, compute providers, validators) have protected voice proportional to their contribution.

This is harder to build, slower to iterate, and may seem inefficient in the short term. But only in this way can the promise of "permissionless building" become truly credible.

For example, adopting a Proof-of-Work (PoW) model to allocate compute resources can establish a governance weight directly tied to contribution: one unit of work, one vote. It does not rely on capital holdings nor the subjective judgment of a founding team. It anchors power to something that cannot be centrally manufactured or revoked—the verifiable computational work itself.

This doesn't solve all governance problems, but it provides an anti-censorship, anti-manipulation starting point.

5. What Did Covenant AI Prove? And Where Should We Go From Here?

We should not let governance failures overshadow their technical achievement. Covenant-72B is a real breakthrough.

Completing the training of a 72-billion-parameter model across over 70 independent nodes, on commodity hardware, without centralized infrastructure—before Covenant AI did it, this was widely considered impossible.  

They proved: decentralized training is technically feasible.

The network later betrayed them, but that does not erase this fact.

The real question is therefore pushed to the next stage: Now that we know the technology is feasible, how do we design governance structures that allow this model to be sustained long-term?

How can capable teams like Covenant AI, after investing years of effort, remain confident that the infrastructure they rely on will not be turned into a weapon against them?

Covenant AI stated they will continue their work outside of Bittensor. This is correct in a sense.  

Decentralized AI training is not a feature exclusive to Bittensor; it is a technical capability independent of any single network. Subnets can die, but the technology will not.

But the industry cannot stop at "absorbing great teams fleeing bad governance."  

We must build governance structures that don't require fleeing from.

The Covenant AI incident is a costly "proof-of-concept"—  It proves not decentralized training (that was already demonstrated),  but what happens when governance design lags behind technological ambition. This lesson is clear for anyone willing to truly listen.

The question is: Is the entire industry ready to face it?

About the Authors:  

David and Daniil Liberman are the co-founders of Gonka, a decentralized AI compute network, currently the largest by GPU count. Both previously served as Product Directors at Snap Inc., with years of deep experience in the AI field.