Bittensor Tanks $900M After Covenant AI Walks — Drama, Dumps, and Decentralization Theater
The chaotic exit and market tumble
In short: a big developer left, the market freaked out, and roughly $900 million in value evaporated in a few frantic hours. Covenant AI — the team behind one of the largest subnets and a headline-grabbing 72-billion-parameter model — announced it was abandoning the Bittensor network, and the native token suffered a violent drop. Prices slid sharply, forcing margin liquidations and wiping hundreds of millions from related subnet tokens.
The tumble was dramatic because it mixed technical achievements with very human drama. Covenant’s project had been riding a recent rally fueled by institutional interest and technical milestones. Then allegations of backroom control and coordinated token selling spilled into public view, and the market rewarded the soap opera by selling first and asking questions later.
Accusations flew about who really ran the network. Covenant’s founder publicly said the ecosystem was more “decentralization theatre” than genuinely decentralized, alleging that a co-founder effectively overruled subnet teams — suspending token emissions, removing moderation powers, and using token sales as leverage. The accused denied having unilateral power and said any token sales were small relative to holdings.
Complicating things, there were big token moves right before the announcement: large sell-offs of subnet tokens that amplified selling pressure and tanked portfolios of retail followers. Social media snippets and clips showing the involved parties sounding exhausted or transactional made the optics worse, and many community members reacted with anger toward what they saw as a messy exit strategy.
Tech cred, governance fixes, and what’s next
Despite the battlefield vibes, the technical side of the story is impressive. Covenant helped build Subnet 3 (known as Templar), a decentralized training setup that proved you could train large language models with consumer-grade hardware across many contributors. Their Covenant-72B model reportedly processed over a trillion tokens across dozens of independent participants and scored competitively on standard benchmarks — a clear signal that decentralized AI research can produce real results.
So why did this get so ugly? Partly it’s the gap between ideals and incentives. Permissionless networks let anyone build, but they also let anyone bail — and tokenized projects create obvious exit opportunities. In response to the chaos, Bittensor leaders floated changes aimed at preventing future sudden dumps: a lock-based subnet ownership model that ties a project’s valuation to developers’ long-term commitment, gives markets advance notice if founders unlock tokens, and makes it easier for investors to shift their stake to different managers.
Whether those fixes will calm things or just add another set of rules remains to be seen. Institutional interest hasn’t vanished — organizations are still building across multiple subnets and the network plans to increase active subnets later this year — but the episode is a reminder that decentralized systems are as much human messy as they are technical brilliant.
Bottom line: you’ve got an impressive decentralized AI experiment with real tech milestones, a bitter governance fight that cost a fortune in market value, and a community scrambling to figure out how to lock in the good parts while avoiding repeat disasters. Popcorn recommended; bring patience and skepticism.
