Covenant AI quits Bittensor. Here's what just happened.

The team behind Covenant-72B, the largest decentralized LLM pre-training run in history, has publicly quit Bittensor. In a departure letter published this week, Covenant AI accused co-founder Jacob Steeves (Const) of using punitive, unilateral actions to punish a subnet that had grown too independent to control, then hiding behind a governance structure the team calls "decentralization theater."

This is the same team whose work was driving Bittensor's biggest moment of the year.

They built the model that made TAO rally 90%

On March 10, 2026, Bittensor's Subnet 3 (Templar), operated by Covenant AI, completed Covenant-72B, a 72 billion parameter language model trained permissionlessly across more than 70 independent contributors on commodity internet hardware. The model was trained on roughly 1.1 trillion tokens and scored 67.1 on MMLU, putting it ahead of Meta's Llama-2-70B.

The achievement triggered a cascade of market and media reactions. TAO, Bittensor's native token, rallied roughly 90% in March. Nvidia CEO Jensen Huang and investor Chamath Palihapitiya endorsed Bittensor's decentralized approach on the All-In Podcast. Analysts called it "Bittensor's DeepSeek moment." Grayscale and Bitwise had spot TAO ETF applications pending with the SEC. Covenant-72B was the proof point institutional capital was buying.

Weeks later, that same team is walking away.

Four ways Const allegedly squeezed them out

Covenant's departure letter accuses Const personally of taking four actions against their operations:

First, suspending TAO emissions to their subnet, cutting miners off from rewards they were earning through contribution.

Second, stripping Covenant's moderation rights over their own community Discord channels, removing the team's ability to manage the communities they had built.

Third, unilaterally deprecating their subnet infrastructure without any transparent governance process or consensus vote.

Fourth, dumping tokens publicly at moments of operational conflict, which the letter characterizes as direct economic pressure timed as coercion.

Covenant's framing is explicit. These were not governance decisions made through decentralized consensus. They were, in the team's words, "punitive actions taken by one man who never relinquished control of a network he tells the world he no longer runs."

"The triumvirate is a legal shield, not governance"

Beyond the specific incidents, Covenant made a structural argument about how Bittensor is actually governed. The network operates through a triumvirate multisig, three individuals who control network upgrades, presented publicly as distributed governance.

Covenant's letter claims this is a facade. According to the team, Const maintains effective control over the triumvirate, resists meaningful transfer of authority, and deploys changes unilaterally whenever he chooses. The other signers, the letter alleges, function as legal shields, people who can be held accountable while real control stays with one person.

This charge lands differently because of the timing. In February 2026, Const publicly stepped down as CEO of the Opentensor Foundation, framing the move as part of Bittensor's evolution toward full decentralization. At the time, Const himself acknowledged in a public statement that little would change operationally, that he would remain in the same calls, continue writing chain code, and keep building subnets, just without the formal legal authority of a CEO title. Covenant is now pointing at that statement as evidence the step-down was optics, not a genuine transfer of power.

Worst possible timing for the TAO ETF pitch

The departure lands at the worst possible moment for Bittensor's institutional narrative. The network had spent March converting Covenant-72B into a case for decentralized AI training as a real category. Huang's endorsement on the most-listened-to tech podcast in the world extended that case beyond the crypto echo chamber. ETF applications were in flight.

Losing the exact team that produced the flagship achievement, with a letter that attacks the network's core decentralization claim, is the worst possible sequencing for any ETF review. The SEC reads founder departures and governance disputes as material risk. The argument Bittensor has been making to investors, that no single entity controls the network, is now being directly contradicted by the most credible possible source: the team that built the thing the network is marketing.

Where does Covenant go now?

Covenant's letter was careful to frame the departure as a separation from Bittensor specifically, not from decentralized AI as a mission. The team said their research, models, and vision travel with them, and promised announcements of new projects in the near future. They did not name a destination. The open question is whether they are building their own network, forking infrastructure, migrating to a competing decentralized training platform, or something else entirely.

For Bittensor, the immediate question is whether other subnet teams watching this dispute conclude that operating under Const is a sustainable long-term bet, or whether Covenant's exit is the first of several. For decentralized AI as a sector, this is the first real stress test of whether community-owned training infrastructure can survive a founder dispute with its core decentralization claim intact.

The answer will depend less on what Const does next, and more on whether the network Covenant is leaving behind can credibly be called decentralized without the team that built its most famous achievement.

If you find this helpful, feel free to follow us for future updates. ❤

• X : @blockflow_news
• Telegram: t.me/blockflownews