Bittensor Covenant-72B and Grayscale's 43% TAO Bet: Decentralized AI Meets Institutional Capital

When Seventy Miners Match a Frontier Lab, Capital Notices

On March 10, 2026, a network of over 70 independent contributors running commodity GPUs on home internet connections completed the largest decentralized large language model pre-training run in history. The result — Covenant-72B, a 72-billion-parameter model trained on approximately 1.1 trillion tokens — scored 67.1 on the MMLU benchmark (zero-shot), surpassing both Meta’s LLaMA-2-70B and LLM360’s K2. Less than a month later, Grayscale raised Bittensor’s weighting in its Decentralized AI Fund from roughly 31% to over 43%, the single largest allocation shift in the fund’s history. The sequence — technical proof, then institutional reallocation — marks a turning point in how capital markets evaluate decentralized AI infrastructure.

What Covenant-72B Actually Proved

The significance of Covenant-72B is not the model itself. A 67.1 MMLU score is competitive but not state-of-the-art. What matters is how the model was produced.

Training a 72-billion-parameter model typically requires coordinated access to thousands of high-end GPUs in a single data center, with fast interconnects and centralized orchestration. Bittensor’s Subnet 3 (Templar) achieved a comparable result using distributed commodity hardware, with contributors joining and leaving freely throughout the training run. According to Phemex, the team reduced communication overhead by a factor of 146 using a technique called SparseLoCo, combined with 2-bit quantization to minimize bandwidth requirements.

The model weights and all checkpoints are publicly available on Hugging Face under an Apache license. This openness is itself part of the thesis: decentralized training produces not just a model but a public good, one whose value accrues to the network rather than a single corporate balance sheet.

For institutional investors, the implication is structural. If frontier-scale training can be distributed across permissionless participants rather than concentrated in hyperscaler data centers, then the competitive moat around centralized AI labs is narrower than assumed. That does not mean decentralized training will replace OpenAI or Anthropic. It means the marginal cost of producing competitive open models may fall faster than centralized incumbents expect.

The economic argument sharpens when set against centralized AI spending. Hyperscalers have committed hundreds of billions of dollars in capital expenditure to AI infrastructure. Covenant-72B was produced by volunteers on consumer hardware. The output quality gap remains real, but the cost gap is equally real — and the cost gap favors the decentralized approach in ways that compound over time as commodity GPU performance improves.

Grayscale’s Rebalance: Reading the Signal

Grayscale’s Decentralized AI Fund underwent its latest quarterly rebalance in early April 2026. The changes, reported by ainvest, were stark:

• Bittensor (TAO): increased from approximately 31% to over 43%
• NEAR Protocol: reduced to 24.43% from 26.54%
• Render: slight increase to 15.77%
• Filecoin: cut to 9.86% from 13.77%
• The Graph: reduced to 4.15%
• Story: cut to 2.73%

No new assets were added or removed. The rebalance was a pure conviction trade within existing holdings.

To appreciate the magnitude, consider the trajectory. In Grayscale’s Q4 2025 rebalance, announced January 7, 2026, TAO held 29.88% of the AI Fund. NEAR led at 27.31%, and six assets were more evenly distributed. Three months later, TAO commands nearly half the fund, while every other asset except Render saw its weight trimmed.

The timing is deliberate. Between the two rebalances, three things happened: Covenant-72B shipped, Nvidia CEO Jensen Huang endorsed decentralized training at GTC 2026 by describing Bittensor as “a modern version of folding@home,” and TAO’s market capitalization swelled past $3 billion. Grayscale’s rebalance formulas are quantitative, but the inputs that drive those formulas — price momentum, network activity, liquidity — all shifted sharply in TAO’s favor.

This raises a structural question about momentum-weighted fund design. A rebalance that increases weighting in the asset that has already appreciated the most can amplify returns during uptrends but also amplify drawdowns during reversals. For an asset as volatile as TAO — which traded as low as $143 in February 2026, per CryptoTimes — a 43% concentration in a six-asset fund represents meaningful single-name risk.

The ETF Filing: From Trust to Regulated Product

Grayscale is not stopping at fund allocation. On April 2, 2026, the firm filed Amendment No. 1 to its S-1 registration statement, seeking to convert its Bittensor Trust into a spot ETF listed on NYSE Arca under the ticker GTAO. Custody would be split between Coinbase and BitGo, with pricing tied to the CoinDesk Bittensor Benchmark Rate.

If approved, GTAO would be the first U.S.-listed ETF for an AI-focused crypto asset, according to The Motley Fool. This distinction matters beyond symbolic value. ETFs open access to retirement accounts, model portfolios, and institutional mandates that cannot hold over-the-counter trust shares. The conversion playbook is familiar — Grayscale successfully converted its Bitcoin and Ethereum trusts into ETFs — but applying it to a sub-$5 billion AI token represents a meaningful expansion of what institutional crypto infrastructure covers.

Barry Silbert, Grayscale’s founder, described decentralized AI as “developing quickly” and positioned the firm as offering “regulated, custody-free exposure to open-source intelligence,” per CryptoTimes.

The ETF conversion also reveals something about Grayscale’s competitive strategy. In a market where BlackRock and Fidelity dominate Bitcoin and Ethereum ETFs, Grayscale’s edge lies in moving first on niche assets where the firm’s crypto-native expertise provides an informational advantage. TAO, with its complex subnet economics and AI-specific value proposition, is exactly the kind of asset where first-mover status translates into durable market share.

Why the Bitcoin Parallel Resonates with Allocators

Bittensor’s tokenomics mirror Bitcoin’s in ways that institutional analysts find legible. Both networks cap total supply at 21 million tokens. Both use halving events — Bittensor completed its first in December 2025, cutting daily emissions from 7,200 to 3,600 TAO, per CryptoTimes. Both require miners to earn new supply through work.

The critical difference, as The Motley Fool noted, is that Bitcoin miners solve “intentionally arbitrary cryptographic puzzles,” while Bittensor miners contribute computing power to training AI models. The useful-work framing addresses a persistent ESG objection to proof-of-work networks: the energy is producing something beyond network security.

This structural familiarity lowers the analytical barrier for portfolio managers who already hold Bitcoin exposure. TAO fits existing mental models — scarce supply, halvings, miner economics — while adding a growth narrative tied to AI demand. Whether that framing survives contact with TAO’s actual volatility (the token remains more than 59% below its all-time high of approximately $757 to $768, depending on the data source, according to Phemex and CryptoTimes) is an open question.

The halving narrative also carries a timing component. Bitcoin’s post-halving supply compression historically correlates with price appreciation over the following twelve to eighteen months. TAO’s first halving in December 2025 places the current period within that window, giving momentum-oriented allocators a catalyst to point to — even if TAO’s market is far less liquid and far less studied than Bitcoin’s.

The Subnet Economy: Where Fundamentals Meet Speculation

Bittensor’s value proposition ultimately depends on whether its subnet ecosystem generates real economic activity. The network currently operates more than 128 active subnets, per The Motley Fool, with plans to expand to 256, according to Phemex. Each subnet specializes in a different AI task — inference, training, data labeling, image generation — and competes for TAO emissions based on performance.

The most tangible example of subnet product-market fit is Chutes, which offers AI inference services at costs that The Motley Fool described as “as much as 90% less than the dominant centralized cloud providers.” If even a fraction of subnets achieve comparable traction, the network transitions from a token-incentivized experiment to genuine infrastructure.

The February 2025 dTAO upgrade added subnet-specific alpha tokens, creating a secondary market layer where capital can flow directly to individual subnets rather than only to the root network. Combined ecosystem token market value reached approximately $1.5 billion, per Phemex. The Templar subnet token itself surged 194% in seven days following the Covenant-72B announcement.

This sub-economy introduces both opportunity and risk. Opportunity, because it allows price discovery at the subnet level — investors can express views on specific AI capabilities rather than betting on the entire network. Risk, because subnet tokens add leverage and complexity to an already volatile asset class. A broader AI sector market capitalization of $17.2 billion, per CryptoTimes, suggests that TAO’s roughly $3 billion market cap already represents a substantial share of the investable AI-crypto universe — leaving less room for error if the subnet economy fails to mature.

What Covenant-72B Does Not Prove

Intellectual honesty requires acknowledging the gaps. A 67.1 MMLU score is respectable but does not compete with the latest frontier models from leading labs. The model was trained on general internet data, not the curated, high-quality datasets that differentiate commercial offerings. The training run demonstrated coordination at scale, but it has not yet demonstrated the ability to iterate rapidly — shipping one model is different from shipping continuous improvements.

More fundamentally, decentralized training faces physics constraints that centralized labs do not. Communication latency across home internet connections is orders of magnitude higher than within a data center. SparseLoCo’s 146-fold reduction in communication overhead is impressive engineering, but it is a workaround for a structural disadvantage, not an elimination of it. Scaling beyond 72 billion parameters with the same architecture will require further algorithmic breakthroughs.

The institutional thesis also carries execution risk. Grayscale’s ETF conversion depends on SEC approval, which is not guaranteed. TAO’s circulating supply of under 10.79 million tokens (out of a 21 million hard cap), per CryptoTimes, means that large institutional flows could create significant price impact in both directions. And the broader AI-crypto narrative has attracted capital to dozens of projects, many of which will not survive a down cycle.

There is also a governance question. Bittensor’s subnet selection and emission allocation ultimately depend on validator consensus — a smaller and less decentralized set of actors than the miner population. If a handful of large validators can steer emissions toward favored subnets, the “decentralized” label becomes less accurate, and the centralization risk that Bittensor claims to solve in AI may reappear within its own governance layer.

The Broader Signal: AI-Crypto Convergence Gets a Benchmark

Before Covenant-72B, the AI-crypto investment thesis was largely aspirational. Projects promised decentralized compute, training, and inference, but few had produced artifacts that external observers could evaluate against centralized alternatives. Covenant-72B provides that benchmark. It is not the best model in the world, but it is the best model produced without a data center, and that distinction is what institutional capital is pricing.

The market’s response was immediate. TAO rose approximately 54.8% over two weeks following the March 10 announcement, per Phemex. Futures open interest nearly tripled, surging from $131.9 million on March 4 to $361.1 million by March 17, according to CryptoTimes. Chamath Palihapitiya called the achievement a “pretty crazy technical accomplishment” on the All-In Podcast, per CryptoTimes. The European market had already signaled institutional interest: Deutsche Digital Assets and Safello launched a staked TAO exchange-traded product on the SIX Swiss Exchange in October 2025, per CryptoTimes.

Grayscale’s 43% allocation and ETF filing translate that market enthusiasm into a durable institutional product. The question is no longer whether decentralized AI training works. The question is whether it works well enough, fast enough, and at sufficient scale to justify the valuations now being assigned to it.

Key Takeaways

• Covenant-72B demonstrated that a 72-billion-parameter LLM can be trained by distributed contributors on commodity hardware, scoring 67.1 on MMLU — competitive with centralized alternatives like LLaMA-2-70B, per Phemex.
• Grayscale raised TAO’s weighting to over 43% in its Decentralized AI Fund, up from 29.88% just three months prior, making it the fund’s dominant holding by a wide margin, per ainvest and GlobeNewsWire.
• A potential GTAO ETF on NYSE Arca would be the first U.S.-listed ETF for an AI-focused crypto asset, opening TAO to retirement accounts and institutional mandates, per CryptoTimes and The Motley Fool.
• Bittensor’s Bitcoin-like tokenomics (21M cap, halving, miner work) provide institutional analysts with familiar valuation frameworks, though TAO remains highly volatile and well below its all-time high.
• The subnet economy is the execution test: more than 128 subnets are live, but only a handful show product-market fit. Whether Bittensor generates real economic activity beyond token speculation will determine if the institutional thesis holds.