Bittensor’s Covenant-72B Shatters Centralized AI Monopoly Despite Network Governance Crisis

The decentralized AI (DeAI) sector reached a historic milestone on May 1, 2026, as Bittensor (TAO) successfully completed the training of “Covenant-72B,” the first 72-billion-parameter large language model to be developed entirely across a permissionless peer-to-peer network. While the technical achievement has sparked a 5.63% rally in TAO prices, it has also triggered a high-stakes governance crisis as the architects of the model announce their exit from the network, citing “centralization theatre.”

By Jennifer Kim | May 1, 2026

TL;DR

• Technical Breakthrough — Bittensor’s Subnet 3 has finalized Covenant-72B, achieving parity with centralized models like Meta’s Llama 2 70B using only decentralized compute.
• Governance Rift — The Covenant AI team is exiting Bittensor, alleging that founder Jacob Steeves maintains excessive control over subnet emissions.
• Market Momentum — TAO has surged to $263.39, bolstered by Grayscale’s Bittensor Trust achieving SEC-reporting status and a high staking rate of 68%.

The promise of decentralized artificial intelligence has long been viewed as a theoretical alternative to the “closed-door” development favored by industry giants like OpenAI and Google. However, the release of the Covenant-72B model this week marks the first time that permissionless infrastructure has produced a model capable of competing on the global stage. Despite the technical triumph, the Bittensor (TAO) ecosystem is navigating a “identity crisis” as it balances the need for coordination with its foundational ethos of absolute decentralization.

Covenant-72B: The MMLU Benchmark That Changed Everything

The Covenant-72B model was trained across a network of over 70 independent nodes using a breakthrough protocol known as SparseLoCo. This protocol allowed participants—many using commodity GPU hardware—to contribute to the training run with 146x less communication overhead than traditional data center setups. This “sparse” approach to training proves that the massive, high-speed interconnects used by centralized AI firms are no longer a mandatory barrier to entry for large-scale model development.

In terms of performance, Covenant-72B scored an impressive 67.1 on the Massive Multitask Language Understanding (MMLU) benchmark. This score puts it in direct competition with Meta’s Llama 2 70B and significantly ahead of previous decentralized attempts. Nvidia CEO Jensen Huang reportedly characterized the project as the “modern version of folding@home,” a nod to the collaborative, distributed nature of the compute power involved. For the first time, a $100 million centralized budget was not required to produce a top-tier AI asset.

The Governance Dilemma: Decentralization vs. Control

However, the celebration has been tempered by a public fracture in the network’s leadership. Sam Dare, the lead architect of Covenant AI, announced on May 1 that his team would be migrating their future development away from Bittensor. Dare alleged that the network’s founder, Jacob Steeves (known as “Const”), maintains a level of “disguised centralization” that allows for the unilateral manipulation of TAO emissions to specific subnets. This “triumvirate” model of governance, according to Dare, stifles innovation by punishing teams that disagree with the core foundation’s vision.

The fallout from this announcement caused a brief volatility spike, though TAO prices quickly recovered as the network began a transition toward “headless” subnets. These are automated, purely mathematical marketplaces designed to operate without human intervention or subjective oversight. The goal of this transition is to remove the “human element” from governance entirely, ensuring that the incentive mechanism (the distribution of TAO tokens) is based solely on verifiable technical performance rather than interpersonal politics.

The Shift to “Headless” Subnets and Agentic Commerce

As the governance debate rages, the broader DeAI category is pivoting toward what analysts call “Agentic Commerce.” On May 1, Rezolve Ai successfully listed its SQD token on Revolut, marking a major milestone for AI agents that act as autonomous economic actors. These agents are increasingly using Layer 2 payment rails to settle transactions, buy compute power, and even hire other AI agents without human approval. Projections for the remainder of 2026 suggest that AI-to-AI transactions could soon surpass human-driven volume on specialized altcoin networks.

This trend is supported by new standards in “Proof of Inference.” With AI models now handling critical tasks in law and medicine, the industry is moving toward protocols like NANOZK, which use Zero-Knowledge Proofs (ZKP) to cryptographically guarantee that an AI’s output is untampered. This ensures that when a user interacts with a model on a network like Bittensor or Fetch.ai, they can verify that the response came from the exact model they requested, rather than a cheaper, less accurate substitute.

Institutional Stance: Grayscale’s SEC Reporting and TAO Staking

Institutional interest in the sector remains robust despite the internal drama. Grayscale’s Bittensor Trust has officially achieved SEC-reporting status, providing a regulated vehicle for traditional funds to gain exposure to the decentralized AI narrative. This institutional bridge has been a primary driver of the recent price strength, with TAO currently trading at $263.39, a 5.63% gain on the day. Other AI-focused assets are also seeing green, with Fetch.ai (FET) up 2.38% and Akash Network (AKT) rising 2.31%.

Supply dynamics also favor the bulls. Currently, 68% of the circulating TAO supply is locked in staking, reflecting deep long-term conviction among the network’s largest holders. As Bittensor attempts to resolve its governance hurdles through the implementation of “headless” subnets, the combination of scarce supply and institutional inflows suggests that the market is looking past the “Covenant crisis” toward the broader potential of ownerless AI infrastructure.

By the Numbers

• $263.39 — The current market price of Bittensor (TAO), rebounding from a mid-April governance dip.
• 72 Billion — The parameter count of the Covenant-72B model, a world-first for decentralized training.
• 67.1 — The MMLU benchmark score achieved by the decentralized model, matching industry leader Llama 2.
• 68% — The percentage of total TAO supply currently staked, indicating high investor loyalty.

Why This Matters

The success of Covenant-72B proves that the technical barriers to decentralized AI are falling faster than many anticipated. While governance disputes like the exit of the Covenant AI team represent the “growing pains” of a maturing ecosystem, the shift toward “headless” subnets and Agentic Commerce represents a fundamental evolution in digital economics. For investors, the takeaway is clear: Decentralized AI is moving from a sub-sector of crypto into a critical layer of the global AI supply chain, offering a verifiable, permissionless alternative to the centralized models that currently dominate the market.

The cryptocurrency market remains highly volatile. This article is for informational purposes only and does not constitute financial advice.