Bittensor's decentralized network completed training of Covenant-72B in March 2026 — a 72-billion-parameter large language model pre-trained across 70+ globally distributed contributors using commodity internet hardware. The model scored 67.1 on the MMLU benchmark, placing it in the performance r...
"That's a pretty crazy technical accomplishment." — Chamath Palihapitiya, Social Capital CEO, on the All-In Podcast, March 2026
Bittensor's decentralized network completed training of Covenant-72B in March 2026 — a 72-billion-parameter large language model pre-trained across 70+ globally distributed contributors using commodity internet hardware. The model scored 67.1 on the MMLU benchmark, placing it in the performance range of Meta's LLaMA-2-70B. An arXiv paper published the same month confirmed it as the largest decentralized LLM pre-training run on record.
TAO, Bittensor's native token, surged 46% month-to-date to approximately $288, driven by the Covenant-72B milestone, an endorsement from Nvidia CEO Jensen Huang on the All-In Podcast, and Grayscale's pending S-1 filing for a spot TAO exchange-traded product. The AI crypto sector's combined market capitalization reached $17.2 billion. Bittensor's subnet ecosystem hit a $550 million combined valuation across its top 10 subnets.
The technical achievement, however, sits atop precarious economics. Total identifiable demand-side revenue across the entire Bittensor network ranges from $3 million to $15 million annually. Against a $2.6 billion market cap, this implies a 175–200x revenue multiple. Against a fully diluted valuation of $5.8 billion, the multiple exceeds 400x — four to ten times higher than comparably valued infrastructure assets in either crypto or traditional markets.
Templar, operating on Bittensor's Subnet 3, completed training of Covenant-72B on March 21, 2026. The model was trained on 1.1 trillion tokens across 70+ contributors without any centralized computing cluster. No single entity coordinated the training run. Participants contributed standard internet-connected hardware and were compensated through Bittensor's token emission mechanism.
The model's 67.1 MMLU score places it in a competitive range with Meta's LLaMA-2-70B, though below frontier models from OpenAI and Anthropic. Analyst @ElCryptoDoc called it "Bittensor's DeepSeek moment," drawing a parallel to the Chinese AI lab's demonstration that effective training could occur outside the hyperscaler paradigm.
The arXiv paper documenting Covenant-72B's methodology was published in March 2026. It confirmed two firsts: the largest parameter count achieved via fully decentralized pre-training, and the highest MMLU score produced without a centralized cluster. Prior decentralized training efforts had been limited to models under 10 billion parameters.
A correction is notable. During the All-In Podcast episode that catalyzed the price move, Chamath Palihapitiya described the model as having 4 billion parameters. The Opentensor Foundation subsequently clarified the model is 72 billion parameters — an 18x discrepancy that did not dampen market enthusiasm.
TAO traded at a low of $143 on February 11, 2026. By March 20, the token broke $300 for the first time since January, a 110% recovery from the February trough. The rally accelerated after the March 19 All-In Podcast episode in which Nvidia CEO Jensen Huang compared Bittensor to "a modern version of folding@home," referencing the distributed computing project that coordinated volunteer hardware for protein folding simulations.
Futures open interest on TAO rose from $131.9 million on March 4 to $361.1 million by March 17 — a 174% increase in two weeks. The RSI reached 76–77, placing the token in overbought territory by standard technical measures.
Three institutional signals converged in Q1 2026:
The broader AI crypto sector rose 42% on March 21, with Bittensor, Render (RNDR), and the Artificial Superintelligence Alliance token (FET) leading gains. The sector's combined market cap reached $17.2 billion, a 4% single-day expansion.
The gap between Bittensor's technical milestones and its revenue fundamentals is the central tension in its investment case.
Following the first halving in December 2025, daily emissions dropped from 7,200 TAO to 3,600 TAO. At $275 per token, this equates to approximately $990,000 in daily emissions, or $361 million annualized. These emissions constitute the primary subsidy mechanism that compensates miners, validators, and subnet operators.
Demand-side revenue tells a different story:
| Subnet | Role | Est. Annual Revenue | Subsidy Ratio | |--------|------|-------------------|---------------| | Chutes (SN64) | Serverless inference | $1.3–2.4M | 22:1 to 40:1 | | Targon (SN4) | Confidential GPU compute | ~$10.4M (unaudited) | N/A | | Templar (SN3) | Model training | $0 | Fully subsidized | | Network total | | $3–15M | 24:1 to 120:1 |
The Chutes subnet, which receives 14.4% of total emissions (~$52 million annualized), processes 101 billion tokens daily and serves 3,000+ customers. At unsubsidized rates, its pricing would be approximately $1.41 per million tokens — 1.6x to 3.5x more expensive than Together.ai ($0.88/M tokens) or DeepSeek V3 ($0.40–0.80/M tokens).
According to Pine Analytics, the network-wide subsidy-to-revenue ratio implies the protocol spends $24 to $120 in emissions for every $1 of external revenue generated. This framework is consistent with the broader industry finding that 85–90% of blockchain ecosystem value flows remain subsidy-driven.
Bittensor's total annual emission budget of approximately $361 million is less than what Microsoft spends on AI infrastructure in a single week. Microsoft, Google, Amazon, and Meta collectively invested over $200 billion in AI capex in 2025.
Bittensor occupies a distinct niche within the decentralized AI compute sector. Unlike Render Network (GPU rendering marketplace), Akash Network (decentralized cloud), or io.net (aggregated GPU supply), Bittensor coordinates the training and serving of AI models rather than raw compute rental.
Market cap comparison (March 2026):
| Project | Market Cap | Focus | 1-Year Performance | |---------|-----------|-------|--------------------| | Bittensor (TAO) | $2.6B–$3.4B | AI model training/inference | +46% MTD | | NEAR Protocol | $1.6B | AI-integrated L1 | Mixed | | Render (RNDR) | ~$2B | GPU rendering | -85% YoY | | FET (ASI Alliance) | $1.85B | AI agents | Mixed | | Akash (AKT) | $363M | Decentralized cloud | -89% YoY |
The performance divergence is stark. While TAO rallied 46% in March, Render and Akash declined 85% and 89% year-over-year respectively. The market is currently pricing Bittensor's model-layer approach at a premium over raw compute provision.
Decentralized compute networks face structural limitations for large-scale training. Centralized data centers offer tightly coupled hardware with low-latency interconnects that decentralized networks cannot replicate. As of 2026, approximately 70% of AI compute demand is driven by inference rather than training, according to industry estimates. This shift favors decentralized networks for serving models post-training, though Covenant-72B demonstrated that training itself can be coordinated across distributed nodes for models up to a certain scale.
Five structural risks weigh against Bittensor's valuation:
1. Revenue multiple compression. At 175–400x revenue, TAO trades at multiples 4–10x higher than the most aggressively valued comparables in either crypto or traditional infrastructure. Traditional AI infrastructure companies trade at 15–25x forward revenue. High-growth SaaS rarely exceeds 50x.
2. Zero switching costs. Bittensor subnets produce open-source models with standard APIs. Users face no lock-in. When subsidies shrink — as they will with the next halving projected for late 2029 — there is no structural barrier preventing customers from migrating to any provider serving the same model weights.
3. Subsidy cliff dynamics. The December 2025 halving cut daily emissions from 7,200 to 3,600 TAO. The next halving will reduce this to 1,800 TAO. Subnet operators must either double their pricing (making them uncompetitive with centralized alternatives), accept lower compensation (risking miner attrition), or generate sufficient organic demand to fill the gap. None of these outcomes is assured.
4. Audit opacity. No subnet publishes independently audited revenue figures. The $10.4 million attributed to Targon is an estimate based on on-chain token buys. Pine Analytics notes the absence of transparent, verified revenue dashboards across the network.
5. Hyperscaler capex asymmetry. Bittensor's $361 million annual emission budget competes against $200+ billion in combined AI infrastructure spending from four companies alone. The hardware allocation priority and purpose-built data centers of centralized providers constitute a resource advantage that token incentives cannot offset at current scales.
Bittensor has produced a legitimate technical proof-of-concept. Training a 72-billion-parameter model across distributed commodity hardware without central coordination is an engineering achievement that merited the attention it received from Huang and Palihapitiya. The 67.1 MMLU score demonstrates functional competence, not just a demonstration of distributed coordination.
The economics, however, remain in early subsidy phase. The network generates between $3 million and $15 million in identifiable external revenue while distributing $361 million in annual token emissions. This 24:1 to 120:1 subsidy ratio places Bittensor squarely within the 85–90% subsidy-dependence range that characterizes most blockchain ecosystems.
The market is pricing TAO at 175–400x revenue — a bet that demand-side adoption will scale by one to two orders of magnitude before the next halving compresses the incentive budget further. Whether decentralized inference demand, which now constitutes 70% of AI compute workloads, can generate that revenue at scale remains the open question. The technical capability is proven. The business model is not.