Bittensor's native token TAO fell as much as 23% on April 10, 2026, erasing approximately $900 million in market capitalization within 12 hours, after Covenant AI — the team behind the network's largest subnet and the 72-billion-parameter Covenant-72B model — announced its departure from the netw...
"It is decentralization theatre." — Sam Dare, Founder, Covenant AI
Bittensor's native token TAO fell as much as 23% on April 10, 2026, erasing approximately $900 million in market capitalization within 12 hours, after Covenant AI — the team behind the network's largest subnet and the 72-billion-parameter Covenant-72B model — announced its departure from the network. Founder Sam Dare accused Bittensor co-founder Jacob Steeves of maintaining unilateral control over the protocol's governance, emissions, and infrastructure deprecation.
The exit marks the most significant internal rupture in Bittensor's history and arrives just weeks after Covenant-72B's March 10 launch drove TAO up 90% on the strength of the decentralized AI narrative. Grayscale Investments, which filed an amended S-1 on April 2 to convert its TAO trust into a spot ETF under ticker GTAO, now faces a governance controversy at the center of its thesis. The broader AI crypto sector — 919 tokens worth $22.6 billion — is watching closely.
On March 10, 2026, Bittensor's Subnet 3 (Templar) completed training of Covenant-72B, a 72-billion-parameter large language model. The model was trained by over 70 independent contributors using commodity GPUs connected through standard home internet connections — no data center required.
Covenant-72B achieved a 67.1 MMLU zero-shot score, according to a March 2026 arXiv paper (arXiv:2603.08163), surpassing Meta's LLaMA-2-70B and the LLM360 K2 benchmark. The model was trained on approximately 1.1 trillion tokens using SparseLoCo, a communication-efficient training algorithm that combines Top-k sparsification, 2-bit quantization, and error feedback to reduce inter-node communication by over 146x compared to standard distributed training approaches (arXiv:2508.15706). Each participant performed 30 local optimization steps before sending a compressed gradient update back to the network.
The result is not frontier-level — GPT-4 class models score significantly higher on MMLU — but it demonstrated that decentralized training over consumer-grade hardware could produce models competitive with centralized efforts from 2023. Nvidia CEO Jensen Huang referenced the achievement on the All-In Podcast, calling it "a modern version of folding@home." The AI token sector rose 40.9% the following day.
TAO rallied approximately 90% in March 2026 alone, briefly exceeding $340 with a market cap near $3.5 billion.
On April 10, 2026, Covenant AI founder Sam Dare posted a statement on X announcing the team's departure from the Bittensor network. The statement accused Bittensor co-founder Jacob Steeves (known as "Const") of maintaining effective unilateral control over the protocol despite marketing a "three-signature multi-sig governance" structure as decentralized.
Specific allegations included:
Separately, critics accused Covenant AI of dumping approximately 37,000 TAO tokens — valued at over $10 million — during its exit, raising counter-allegations of a coordinated "rug pull."
A newly launched site called Tao Papers (taopapers.com) published on-chain forensics and internal documents from multiple whistleblowers. The central claim: Bittensor's triumvirate governance structure is a "counter-signature ritual."
Key findings from the site's analysis:
These allegations, if accurate, describe a protocol where governance decentralization exists at the code level but not in operational practice — a pattern that has precedent in other proof-of-stake networks but rarely surfaces with this level of documented specificity.
Jacob Steeves responded publicly, denying he has the ability to suspend subnet emissions. According to Steeves, he sold some of his "alpha holdings" on Covenant AI's three subnets because they were not running and were operating on "near 100% burn code." He stated this changed emissions through normal market mechanisms — the same way any buy or sell on Bittensor affects the Dynamic TAO system — and that he holds no special privileges beyond those of a standard TAO holder.
The denial directly contradicts Dare's claims. No independent third-party audit of the governance logs has been published as of April 10.
TAO's price action on April 10, 2026:
| Metric | Value | |--------|-------| | Pre-announcement price | ~$332 | | Intraday low | ~$254 | | Maximum drawdown | -23% (some sources report up to -27%) | | Market cap erased | ~$900 million | | Long liquidations | >$9 million | | Covenant AI alleged TAO dump | ~37,000 TAO (~$10 million) | | 24-hour trading volume | $649M–$801M (varies by source) |
By late April 10, TAO was trading in the $270–$329 range, depending on the source and time of query. CoinMarketCap ranked TAO at #27 with a market cap of approximately $3.1–$3.6 billion.
The sell-off was concentrated in TAO; the broader AI crypto sector ($22.6 billion across 919 tokens) showed limited contagion as of end-of-day April 10.
Grayscale Investments filed Amendment No. 1 to its S-1 registration on April 2, 2026, to convert its over-the-counter Bittensor Trust into a spot ETF (ticker: GTAO) listed on NYSE Arca. The initial S-1 was filed December 30, 2025.
Key institutional data points:
The governance controversy creates a timing problem. The SEC evaluates ETF applications partly on whether the underlying asset's market is resistant to manipulation and whether the issuer's governance poses systemic risks. Covenant AI's allegations — that one individual controls network upgrades, emissions, and infrastructure — could complicate Grayscale's approval timeline, though the SEC has not publicly commented on the dispute.
As of early April 2026, before the Covenant AI exit:
| Metric | Value | |--------|-------| | Active subnets | 128 (cap; planned expansion to 256) | | TAO staked | ~70–76.55% of supply | | Subnet Alpha token market cap | ~$1.12 billion (27% of TAO market cap) | | Top 10 subnets combined valuation | ~$712 million | | Daily TAO issuance (post-halving) | 3,600 TAO | | First halving date | December 14, 2025 |
The December 2025 halving cut daily issuance from 7,200 to 3,600 TAO. Under the Dynamic TAO (dTAO) model introduced in February 2025, each subnet operates its own Alpha token and liquidity pool, with markets — rather than validators — determining emissions allocation. This market-driven model is central to Bittensor's value proposition but also means that large holders can influence subnet economics through token sales, as both Dare and Steeves acknowledged from different perspectives.
The Covenant AI exit exposes a structural tension in decentralized AI networks: the teams capable of producing frontier research (Covenant-72B required significant coordination across 70+ nodes) are also the teams with enough leverage to challenge governance structures. When they leave, they take both technical capability and narrative credibility with them.
Three observations for the sector:
1. Governance maturity lags technical achievement. Bittensor produced a 72B-parameter model on commodity hardware — a genuine technical milestone. But the governance infrastructure that allocates resources, resolves disputes, and manages upgrades appears to have remained at a seed-stage centralization level. This pattern is common in crypto protocols, where technical scaling outpaces organizational scaling.
2. Incentive alignment remains unsolved. The dTAO model allows markets to price subnets, but it also allows large token holders (including founders) to economically pressure subnet operators by selling their Alpha tokens. Whether Steeves's sales constituted routine portfolio management or targeted economic coercion depends on timing and intent — data that on-chain forensics can partially, but not fully, resolve.
3. ETF filings amplify governance scrutiny. Traditional finance entry points (Grayscale's GTAO, the Decentralized AI Fund's 43% TAO allocation) bring institutional due-diligence standards that crypto governance structures were not designed to withstand. The Covenant AI dispute is precisely the type of governance risk that SEC examiners and institutional allocators will flag.
The Covenant AI exit is a stress test for decentralized AI's core proposition: that meaningful AI work can be produced, governed, and funded without centralized control. Bittensor demonstrated the technical half of that thesis with Covenant-72B. The governance half remains unproven — and the same team that validated the technology is now the one challenging the organizational structure.
For the network, the immediate question is whether the remaining 127 subnets and their operators view the governance concerns as systemic or as an isolated dispute. For the broader sector, the question is whether institutional capital — already flowing through Grayscale's 43% TAO allocation and pending ETF — will wait for governance clarity or reprice the risk.
No independent governance audit has been completed. Until one is, the factual record consists of two contradictory accounts and a set of on-chain logs that have not been independently verified.