Skip to main content

Bittensor's Conviction Test: Can Locked TAO Save Decentralized AI After the Covenant Shock?

· 9 min read
Dora Noda
Dora Noda
Software Engineer

On March 10, 2026, a network of roughly 70 strangers scattered across the open internet finished training a 72-billion-parameter language model that beat LLaMA-2-70B on MMLU. Six weeks later, the same network was trying to stop itself from falling apart.

That whiplash — from a historic technical milestone to a full-blown governance crisis — is the story of Bittensor in 2026. And the fix on the table, a strange new primitive called the Conviction Mechanism, may be the most important governance experiment in crypto-AI this year.

From Covenant-72B to Covenant AI's Exit​

The high point came first. Templar, Bittensor's Subnet 3 (SN3), announced the completion of Covenant-72B — the largest decentralized LLM pre-training run ever attempted. Over 70 independent contributors pooled commodity GPUs over regular home internet connections and processed roughly 1.1 trillion tokens to produce a 72B-parameter model that scored 67.1 on zero-shot MMLU, edging past LLaMA-2-70B and LLM360 K2.

The secret sauce was an algorithm called SparseLoCo. By combining Top-k sparsification, 2-bit quantization, and error feedback, SparseLoCo reduced inter-node communication by roughly 146x. In practical terms: what used to require 280 GB per sync dropped to 2.2 GB per sync, with nodes synchronizing 500x less often. Gradients were compressed to about 0.78% of their original size, and nodes could take 15–250 local steps before talking to anyone else.

That is the punchline decentralized AI has been chasing for years: you do not need an NVLink-stitched data center to train a frontier model. The open internet is enough, if the optimizer is smart enough.

Markets noticed. τemplar, the subnet's alpha token, jumped 194% in seven days, and TAO itself rose 54.8% over two weeks. On-chain analysts called it TAO's "DeepSeek moment."

Then the ceiling caved in.

In mid-April, Covenant AI founder Sam Dare sold roughly 37,000 TAO — about $10.2 million — and exited the ecosystem, publicly criticizing what he framed as creeping centralization. The sale hit an already nervous market. TAO dropped roughly 25%, wiping out an estimated $650 million in market capitalization. Worse, it exposed a structural flaw that had been quietly accumulating: Bittensor subnet owners had few reasons to stay.

The Governance Bug Nobody Wanted to Fix​

Bittensor's original design gave subnet owners significant power and tokens but imposed no meaningful skin-in-the-game requirement after launch. An inactive or disillusioned owner could sit on a subnet indefinitely, or unwind their position with no warning — precisely what Covenant AI did.

For a protocol marketing itself as the coordination layer for decentralized AI, "the founder can leave on Tuesday and crash the token on Wednesday" is not a story that sells to institutions. And the institutions were watching. Grayscale filed an S-1 amendment for a spot TAO ETF on NYSE Arca on April 2, with Bitwise filing a parallel strategy ETF the same day. ETF issuers do not love governance that depends on one person's mood.

The question became concrete: how do you force long-term alignment in a system where tokens are liquid and subnet ownership is transferable?

Co-founder Const's answer, announced April 16, is BIT-0011 — the Conviction Mechanism.

What the Conviction Mechanism Actually Does​

The design is unusual enough to be worth describing carefully.

Under BIT-0011, anyone who wants subnet ownership — or the voting rights that come with it — must time-lock alpha tokens for a chosen period, measured in months to years. In exchange, the locker receives a conviction score that starts at 100% and decays over 30-day intervals. The conviction score, not the raw token balance, is what maps to governance weight and subnet control.

A few consequences fall out of this:

  • Sitting still is not enough. Because conviction decays, a subnet owner cannot lock once and coast. They have to keep renewing commitment, or watch their voting power bleed away to someone more engaged.
  • Ownership is contestable. Under BIT-0011, anyone can challenge for control of a subnet by locking alpha and building conviction. If a challenger's conviction score exceeds the incumbent's, they take over. This is a continuous, market-driven recall mechanism — the opposite of the "founder keeps the keys forever" default.
  • Exit is expensive. Unlocking means losing conviction. Dumping tokens without unlocking is impossible. The incentive surface directly penalizes the behavior that triggered the April crisis.

The mechanism is being rolled out first to mature subnets: SN3 (Templar), SN39, and SN81. That choice is not accidental. These subnets have real economic activity and real work being done, which means the market actually cares who owns them.

Why Start With Templar?​

Putting the Conviction Mechanism on SN3 first is a tell. Templar is the subnet that just demonstrated frontier-scale decentralized training — it is the single most defensible proof that Bittensor is more than a token-emission game. If the Conviction Mechanism can retain serious operators on the subnet that matters most, the model generalizes. If it cannot, starting anywhere else would be cosmetic.

There is also a narrative symmetry. Templar's SparseLoCo is an engineering bet that coordination without constant communication is possible. The Conviction Mechanism is a governance bet that alignment without constant trust is possible — if you make exit costly enough, you do not need everyone to be a believer. Both are attempts to replace centralized control (a data center; a benevolent founder) with economically-enforced coordination.

The Uncomfortable Questions​

The Conviction Mechanism is elegant, but it is not free of friction. A few open questions worth sitting with:

1. Who can actually lock enough to matter? If serious governance weight requires locking millions of dollars of alpha for years, the pool of viable subnet owners narrows to funds and well-capitalized teams. That is arguably fine — you want serious operators — but it is not obviously more "decentralized" than what came before. It is a different form of concentration, just one that rewards time preference rather than early entry.

2. Does decay calibrate correctly? A 30-day decay interval is a design choice, not a physical constant. If decay is too fast, the subnet churns. If it is too slow, the mechanism repeats the original bug on a longer timeline. Expect the first year to involve parameter tuning and probably some painful edge cases.

3. What happens during a challenge? Continuously contestable ownership sounds great until a well-funded adversary decides to build conviction on a subnet critical to someone's business. The mechanism needs robust tie-breaking and probably some form of cool-down, or it risks turning subnet governance into a hostile-takeover playground.

4. Does it actually satisfy ETF due diligence? The deeper question is whether "conviction score determines the owner" reads as sound governance to a U.S. regulator. Time-locked stake is a real commitment signal, but it is still novel enough that allocators will have to develop frameworks to evaluate it. Bittensor is, in effect, running that evaluation in public.

Where This Leaves Decentralized AI​

Zoom out and the 2026 Bittensor story reads like a stress test of the entire decentralized-AI thesis.

The technical side — can we train frontier models without data centers? — got a powerful "yes, with caveats" from Covenant-72B. SparseLoCo shows the bandwidth problem is tractable. The 67.1 MMLU score shows the quality problem is tractable. Seventy miners on home internet is no longer a pitch deck; it is a result.

The economic side — can we align operators without a central authority? — is exactly what BIT-0011 is trying to answer. Covenant AI's exit showed what happens when the answer is "no." The Conviction Mechanism is the first serious attempt to build alignment into the token design itself, rather than relying on norms or founder goodwill.

If both answers stabilize to "yes," the implications travel well beyond Bittensor. Every decentralized compute network — training or inference, AI or otherwise — faces versions of these two problems. Most of them have been quietly hoping the governance problem would not matter until later. Bittensor just showed that "later" arrives faster than anyone wanted.

The Call​

For builders, the interesting move in Q2 2026 is not to take a side on whether BIT-0011 is the right design. It is to watch SN3 closely over the next two quarters. If Templar's operators actually lock meaningful alpha, if conviction-weighted voting produces decisions the subnet can live with, and if no hostile challenger breaks the mechanism in the first 90 days — then the Conviction Mechanism becomes a reusable governance primitive, not just a Bittensor patch.

And if it fails, the failure itself will be one of the most informative data points in the decentralized-AI design space. Either outcome is worth paying attention to.

Bittensor spent March proving that the internet can train a 72B model. It is spending April proving whether the internet can govern one. The second problem, it turns out, was always the harder of the two.

BlockEden.xyz provides enterprise-grade RPC and indexing infrastructure across 30+ chains, including dedicated support for AI and compute-focused networks. Explore our services to build on foundations designed to last through governance shocks, not just bull markets.

Sources​

Share on Twitter