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

Compiled by: Gonka.ai

Foreword: When a decentralized AI team widely recognized as having the strongest technical capabilities publicly withdraws from the network and openly admits to centralized control disguised as decentralization, the entire industry should not simply attribute it to a bilateral dispute. This is a diagnosis—a wake-up call revealing the structural problems of the entire decentralized AI ecosystem. And this diagnosis extends far beyond Bittensor.

I. What exactly happened?

Covenant AI (formerly Templar) has spent over two years building what is arguably the most technologically significant milestone in decentralized AI to date: Covenant-72B—a language model with 72 billion parameters. This model was trained without license by over 70 independent contributors on general-purpose hardware.

This is not a proof-of-concept, but a major breakthrough that can be put into production. This achievement received public recognition from Nvidia's CEO on the "All-In" podcast and was also cited by the co-founder of Anthropic. It previously drove a 90% increase in Bittensor's TAO token, with several subnets approaching a valuation of $1.5 billion.

However, right after this prosperity, the infrastructure that underpinned it all turned into a weapon to attack them.

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

- Suspend the distribution of token rewards to the Covenant subnet;

- Revoke their management rights over the community channel;

- Discarding its subnet infrastructure without consultation;

- Use token sales as a pressure tool during operational conflicts.

Covenant AI succinctly points out that this is "centralized control disguised as decentralized."

Within hours of the announcement, TAO tokens plummeted by over 15%. As of press time, they were still down approximately 9% for the day.

In response, Steeves stated that Bittensor will soon launch a truly independent subnet. However, this response confirms rather than refutes the core issue: while Covenant was fully committed to its development, the "independence" they needed was not yet guaranteed in the architecture.

II. Structural Trap: The Value Hostage Dilemma

To understand why this is far more than an isolated case, we must look at the unique economic structure behind it—the very structure that makes the crisis both dangerous and predictable.

Decentralized AI networks face a fundamental "cold start problem": building the real infrastructure—training tasks, model weights, contributor networks, and community trust—requires a long-term investment of 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 just the builders.

This forms a dynamic mechanism that we call "value-hostage":

The better you build it, the more your network tokens appreciate; the more valuable your tokens, the stronger the influence those in control have over you. And when you achieve your greatest success, that's precisely when you're most vulnerable.

Creating value on other people's networks means that the value you create could ultimately become a weapon against you. The more successful you are, the more you lose.

This is not a governance failure unique to Bittensor, but a structural consequence of all systems: when a minority retains veto power over key permissions (such as token issuance, content moderation, and infrastructure upgrades) while claiming the system is "permissionless," the seeds of problems are sown.

The promise of “decentralization” is the foundation of the system’s operation—builders, miners, validators, and investors all make decisions based on it. Once this premise is proven false or only “conditionally true,” the resulting economic losses will not be limited to the team involved, but will spread to all participants who once trusted the narrative.

Covenant AI points out that Bittensor's governance structure is nominally a multi-signature mechanism with "three-person co-management," but in reality, it is dominated by one person, while the other two are more like legal firewalls than actual decision-makers involved in governance.

We cannot independently verify every single allegation. However, we can see the structural logic behind them:

- A multisignature by a single master is no different from a single private key;

- Governance processes that can be unilaterally bypassed cannot be called governance;

- A token issuance mechanism that can be suspended by one person is essentially a subsidy, not a protocol guarantee.

Third, this is not a problem with Bittensor, but a common problem in the industry.

Some might interpret the Covenant AI incident as a "bittensor case warning." This view is too narrow and even misleading.

The deeper problem lies in the fact that the entire decentralized AI field has long maintained a fragile balance between two mutually exclusive goals—rapid iteration and true decentralization.

Rapid iteration requires someone to make the final decisions; true decentralization means that no one can unilaterally decide everything. Most projects have opted for the latter by default: publicly proclaiming decentralization while holding onto control, hoping that conflicts will never be exposed.

And now, it has been exposed.

In the current decentralized AI ecosystem, a familiar pattern keeps recurring:

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

Encased in a centralized core—the founding team or foundation firmly controls the most critical parameters:

- Token release schedule

- Protocol upgrade permission

- Subnet Admission Mechanism

- Community management rights

This isn't necessarily out of malice. Early networks did require strong coordination, and the purely on-chain governance of complex AI infrastructure is technically unresolved. However, the gap between public promises and actual power creates a kind of "structural debt."

Covenant AI's experience is exactly what this debt will look like when it comes due.

In AI networks, this mechanism is more dangerous than DeFi or Layer 1 because the depth of investment by builders is far greater than ever before.

Training a model with 72 billion parameters is not a two-week sprint project, but a long-term battle that is extremely costly, time-consuming, and reputation-risking. By the time Covenant AI realized the seriousness of the governance issues, they had already completed the most crucial work—which is precisely why they became the target.

This asymmetry is extremely cruel:

- The internet can act at any time;

- The builders, however, cannot "undo" the construction that has already been completed.

Similar events will repeat themselves if the following three conditions are met:

1. Builders need to invest a significant amount of time and resources;

2. Tokens can capture the value of these investments;

3. A minority of people can still unilaterally interfere with governance.

These three conditions are extremely common in current AI networks.

Therefore, we should not expect fewer “next Covenant AIs”—but rather anticipate more.

IV. What does true decentralization require?

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

One approach is to uncritically praise any project that claims to be "decentralized";

Another approach is to completely deny the feasibility of decentralization, viewing it as a scam.

Neither of these options is advisable.

We want to make one thing clear: achieving true decentralization in AI infrastructure is technically extremely difficult. Anyone who tells you it's "easy" has most likely never actually tried.

- Cross-regional node collaborative training

- Verifiability of computational work

- Manipulation-resistant token incentive mechanism

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

However, technical difficulty does not equate to impossibility, nor should it be used as an excuse for misleading propaganda.

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

If the answer is "yes," then no matter how the white paper describes it, and regardless of whether there is a governance vote, this decentralization is superficial.

This question cuts through all the noise. It doesn't care if you have a DAO or a community forum. It only asks one real question: In the event of a conflict, is there a small group of people who can unilaterally suspend your rewards, cut off your access, or coerce you financially?

If so, then the claim of "decentralization," no matter how sincere the initial intention, lacks structural support.

The real solution is to embed decentralized constraints into the protocol layer, rather than relying on the "goodwill" of any individual or group.

This means:

- The token issuance mechanism is determined by on-chain rules and cannot be unilaterally suspended;

- Agreement upgrades require genuine consensus from all parties, not mere formalities;

In the governance structure, actual contributors (builders, computing power providers, and verifiers) have a voice that is commensurate with and protected according to their contributions.

This is more difficult to build, iterates more slowly, and may seem inefficient in the short term. But only in this way can the promise of "permissionless building" be truly credible.

For example, using a Proof-of-Work (PoW) model to allocate computing resources can establish a governance weight directly linked to contribution: one unit of computing power equals one vote. It does not rely on capital holdings or the subjective judgment of the founding team. It anchors power to something that cannot be centrally created or revoked—the verifiable computational work itself.

This cannot solve all governance problems, but it provides a starting point for resistance to censorship and manipulation.

V. What has Covenant AI proven? Where should it go in the future?

We should not allow governance failures to overshadow its technological achievements. The Covenant-72B is a true breakthrough.

Training a 72 billion parameter model across more than 70 independent nodes, general-purpose hardware, and decentralized infrastructure was widely considered impossible before Covenant AI did it.

They proved that decentralized training is technically feasible.

The internet betrayed them afterward, but that cannot erase the fact.

The real question is thus pushed to the next stage: Now that we know the technology is feasible, how do we design a governance structure that can make this model sustainable in the long term?

How can a capable team like Covenant AI, after years of hard work, remain confident that the infrastructure they rely on will not turn into a weapon to attack them?

Covenant AI has stated that they will continue their work beyond Bittensor. This is true in a sense.

Decentralized AI training is not a feature exclusive to Bittensor, but rather a technological capability independent of any single network. Subnets may disappear, but the technology itself will not.

But the industry cannot stop at “absorbing excellent teams that have escaped poor governance.”

We must build a governance structure that eliminates the need for people to flee.

The Covenant AI project was a costly "proof of concept"—

It does not validate decentralized training (that has already been proven).

The question is what happens when governance design lags behind technological ambition.

This lesson is clear, it just needs to be heard.

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