Bittensor Ecosystem Token SN Surges 5x in March, What's Behind Richard Heart's One-Liner?

Original Title: "Chamath Palihapitiya Calls Out? SN3 Surges 5x in March, What Exactly Happened?"

Original Author: KarenZ, Foresight News

On March 20, 2026, there was a somewhat unusual conversation in the All-In VC podcast.

Venture capitalist Chamath Palihapitiya passed the mic to NVIDIA CEO Jensen Huang, mentioning a project on Bittensor that had "achieved a pretty crazy technical feat," using distributed compute to train a large-scale language model on the internet in a completely decentralized process with no centralized data center involvement.

Jensen Huang did not shy away. He likened this to a "modern version of Folding@home," the distributed project in the 2000s that had ordinary users contribute idle compute power to collectively tackle protein folding challenges.

Just four days prior, on March 16, Anthropic co-founder Jack Clark, in a research progress report, extensively highlighted and referenced this breakthrough: Bittensor's ecosystem subnet Templar (SN3) completed the distributed training of a 72 billion parameter large model (Covenant 72B), with model performance on par with Meta's LLaMA-2 released in 2023.

Jack Clark titled this section "Challenging AI Political Economy Through Distributed Training" and emphasized in the analysis that this is a technology worth continued tracking — he can envision a future where on-device AI widely adopts models produced through decentralized training, while cloud AI continues to run proprietary large models.

The market's reaction was slightly delayed but very intense: SN3 surged over 440% in the past month, over 340% in the past two weeks, reaching a market cap of $130 million. The subnet's narrative explosion directly translates into buying pressure on TAO. As a result, TAO soared rapidly, briefly reaching $377, doubling in the past month, with a fully diluted valuation of around $7.5 billion.

The question arises: What has SN3 actually done? Why has it been thrust into the spotlight? How will the value narrative of distributed training and decentralized AI evolve?

That 72B Model

To answer this question, we first need to take a close look at the report card handed out by SN3.

On March 10, 2026, the Covenant AI team released a technical report on arXiv, officially announcing the completion of the training of Covenant-72B. This is a large language model with 720 billion parameters, involving over 70 independent node peers (with around 20 nodes synchronizing per round, each node equipped with 8 B200 GPUs), completing the pretraining of the 720 billion parameter model on about 1.1 trillion tokens of text.

Templar provided some benchmarking data, with the LLaMA-2-70B as the point of comparison, a large model released by Meta in 2023. As Jack Clark, co-founder of Anthropic, mentioned, Covenant-72B may be somewhat outdated by 2026. Covenant-72B scored 67.1 on the MMLU, roughly on par with Meta's 2023 LLaMA-2-70B (65.6).

Meanwhile, the cutting-edge models of 2026—whether from the GPT series, Claude, or Gemini—have already undergone training with a parameter size far exceeding 100 billion across hundreds of thousands of GPU blocks. The disparity in inference, code, and mathematical capabilities is not a percentage-based problem but an order-of-magnitude one. This reality gap should not be overshadowed by market sentiment.

However, when translated into the premise of being "trained on open internet-based distributed compute power," the implications are entirely different.

For comparison: INTELLECT-1 (produced by the Prime Intellect team, 100 billion parameters) scored 32.7 on the MMLU for decentralized training; another distributed training project among whitelist participants, Psyche Consilience (400 billion parameters), scored 24.2. With a scale of 72B and an MMLU score of 67.1, Covenant-72B stands out in the decentralized training track.

More importantly, this training was "permissionless." Anyone could access and become a participant node, with no prior approval or whitelist required. Over 70 independent nodes participated in the model updates, contributing compute power from around the globe.

What Huang Renxun Said and Didn't Say

Revisiting the details of that podcast conversation helps correct the external interpretation of this "endorsement."

During the conversation, Chamath Palihapitiya presented Bittensor's technical accomplishments to Huang Renxun, describing it as training a Llama model with distributed compute power, a process that is "completely distributed while preserving state." Huang Renxun's response likened this to a "modern version of Folding@home" and further discussed the necessity of coexisting open-source and proprietary models.

It's worth noting that Huang Renxun did not directly mention Bittensor's token or any investment implications, nor did he delve into decentralized AI training any further.

Understanding the Bittensor Subnet and SN3

To understand the breakthrough of SN3, it is first necessary to understand the operation logic of Bittensor and its subnets. Simply put, Bittensor can be seen as an AI public chain and platform, where each subnet functions as an independent "AI production pipeline," with clearly defined core tasks, incentive mechanisms, and collaborative efforts to form a decentralized AI ecosystem.

Its operation process is clear and decentralized: subnet owners define subnet goals and write incentive models; miners provide computing power in the subnet, complete AI-related tasks (such as inference, training, storage, etc.); validators score the contributions of miners and upload the scores to the Bittensor consensus layer; finally, Bittensor's Yuma consensus algorithm allocates rewards to subnet participants based on the accumulated rewards of each subnet.

Currently, there are 128 subnets on Bittensor, covering various AI tasks such as inference, serverless AI cloud services, images, data annotation, reinforcement learning, storage, computation, and more.

SN3 is one of these subnets. It does not provide an application layer shell, does not lease existing large model APIs, but directly targets one of the most expensive and closed core parts of the entire AI industry chain: large-scale pre-training of models.

SN3 hopes to use the Bittensor network to coordinate distributed training of heterogeneous computing resources, demonstrating through incentive-based distributed large-scale model training that powerful foundational models can be trained without the need for expensive centralized supercomputer clusters. The core appeal lies in "equity" — breaking the resource monopoly of centralized training, allowing ordinary individuals or small to medium-sized organizations to participate in large-scale model training, while leveraging distributed compute power to reduce training costs.

The core driving force behind the development of SN3 is Templar, with the research team behind it being Covenant Labs. The team also operates two other subnets: Basilica (SN39, focusing on compute services) and Grail (SN81, focusing on RL post-training and model evaluation). The three subnets form a vertical integration, covering the entire process of large-scale model training from pre-training to alignment optimization, establishing a comprehensive ecosystem for decentralized large-scale model training.

Specifically, miners contribute computing resources, upload gradient updates (adjustments to model parameters) to the network; validators assess the quality of each miner's contribution, provide on-chain scores based on the improvement in error magnitude. The results determine reward weights, which are automatically allocated, without the need to trust any third party.

The key to the incentive mechanism design is that rewards are directly tied to "how much your contribution has improved the model," rather than merely attendance of computing power. This fundamentally solves the most challenging issue in a decentralized scenario: how to prevent miners from slacking off.

So, how does Covenant-72B address the communication efficiency and incentive compatibility issue?

Having dozens of mutually untrusted nodes with different hardware and varying network quality collaboratively train the same model poses two challenges: communication efficiency, where the standard distributed training approach requires high-bandwidth, low-latency interconnection between nodes; and incentive compatibility, how to prevent malicious nodes from submitting incorrect gradients? How to ensure that each participant trains honestly and does not plagiarize others' results?

SN3 solves these two problems with two core components: SparseLoCo and Gauntlet.

SparseLoCo addresses the communication efficiency issue. Traditional distributed training requires synchronizing full gradients at each step, resulting in a huge amount of data. The approach taken by SparseLoCo is: each node performs 30 steps of localized optimization (AdamW) and then compresses the generated "pseudo-gradient" before uploading it to other nodes. Compression methods include Top-k sparsification (retaining only the most critical gradient components), error feedback (accumulating the discarded part for the next round), and 2-bit quantization. The final compression ratio exceeds 146 times.

In other words, what originally required transferring 100MB of data now only requires less than 1MB.

This allows the system to maintain a computational utilization rate of about 94.5% under bandwidth constraints of a regular Internet connection (uplink 110Mbps, downlink 500Mbps) — with 20 nodes, 8 B200s per node, and each communication round taking only 70 seconds.

Gauntlet addresses the incentive compatibility problem. It runs on the Bittensor blockchain (Subnet 3) and is responsible for validating the quality of the pseudo-gradient submitted by each node. The specific approach is as follows: a small batch of data is used to test "how much the model's loss decreases after using the gradient from this node," and the result is called LossScore. At the same time, the system also checks if the node is training on the data allocated to itself—if a node's loss improvement on random data is better than on its allocated data, it will receive a negative score.

Ultimately, in each training round, only the gradients from the highest-scoring node are selected for aggregation, and the remaining nodes are eliminated for that round. Those eliminated can backfill at any time to keep the system robust. Throughout the entire training process, an average of 16.9 nodes' gradients are included in each round of aggregation, with over 70 unique node IDs having participated cumulatively.

The Value Narrative of Decentralized AI Is Undergoing a Fundamental Shift

From a technical and industry perspective, the Covenant-72B represents several significant aspects:

First, it breaks the assumption that "distributed training is only suitable for small models." Despite being far from state-of-the-art models, it has demonstrated the scalability of this approach.

Second, permissionless participation is truly viable. This has been underestimated. Previous distributed training projects relied on whitelists—only audited participants could contribute computing power. In this SN3 training, anyone with sufficient computing power can join, with the validation mechanism filtering out malicious contributions. This is a concrete step towards "true decentralization."

Third, Bittensor's dTAO mechanism enables market discovery of subnet value. The dTAO allows each subnet to issue its own Alpha token and lets the market decide which subnets receive more TAO emissions through an AMM mechanism. This provides a rough but effective value capture mechanism for subnets like SN3 that have produced tangible results. Of course, this mechanism is also prone to narrative and emotional disruptions, making it challenging for ordinary market participants to independently assess the quality of LLM training outcomes.

Fourth, the political-economic implications of decentralized AI training. Jack Clark raised this issue to the level of "who owns the future of AI" in Import AI. Current state-of-the-art model training is monopolized by a few institutions with large-scale data centers, which is not only a business issue but also a power structure problem. If distributed training can continue to make technological strides, it may possibly establish a truly decentralized development ecosystem for certain model types (such as small-scale state-of-the-art models in specific domains). Of course, this prospect is still far off.

Summary: A Real Milestone and a Host of Real Issues

Huang Renxun said this is like a "modern version of Folding@home." Folding@home made a real contribution in the field of molecular simulation, but it did not threaten the core R&D position of major pharmaceutical companies. This analogy is very accurate.

SN3 successfully ran the protocol, validating the feasibility of distributed training. However, from both a technical and industry perspective, behind this report card it handed in, there are a host of issues that few are willing to discuss seriously:

MMLU itself is a controversial metric in academia, with the risk of leakage of benchmark questions and answers into the training set. Of more concern is the selection of the comparison baseline: the models benchmarked in the paper, LLaMA-2-70B and LLM360 K2, are old models from 2023 to 2024, and the 65 to 70 range in the same interval, when asked about Grok and Bean, are categorized as mid-to-entry-level by Claude but are considered severely lagging. If placed on a dynamically updated leaderboard or a new generation benchmark with anti-pollution design, the conclusions may be more honest.

More critically, the high-quality data that determines the model's upper limit capabilities—dialogue data, code, mathematical derivations, scientific literature—is likely held by major companies, publishing institutions, and academic databases. While compute power has been democratized, the data side remains an oligopoly, a contradiction that has not been discussed.

Regarding security, permissionless participation means you don't know who is behind those 70+ nodes, nor do you know what data they are training on. While Gauntlet can filter out blatantly anomalous gradients, it cannot prevent subtle data poisoning—if a node systematically trains more rounds in a harmful content direction, the resulting gradient changes are subtle enough to pass through loss score screening but accumulate drift in model behavior. The ultimate question is: In high-compliance, security-demanding scenarios such as finance, healthcare, and law, what risks does using a model trained by a few anonymous nodes with incomplete data source traceability bring?

There is also a structural issue worth mentioning: Covenant-72B itself is open source under the Apache 2.0 license and does not use the SN3 token. Holding SN3 tokens shares in the emission benefits of the continued production of new models by this subnet in the future, rather than any direct benefits when the models are used. This value chain relies on ongoing training output and the healthy operation of Bittensor's overall network emission mechanism. If future training stagnates or the quality of new training outcomes falls below expectations, the logic of token valuation will weaken.

List these issues not to negate the importance of Covenant-72B. It proves that something previously thought impossible can be done, and that fact will not disappear. But achieving it, and what it means, are two different things.

The SN3 token has risen 440% in the past month. The distance in between may not be mere hype, as the speed of narrative is always faster than the speed of reality. Whether this distance will ultimately be filled by reality or corrected and digested by the market depends on what the Covenant AI team truly delivers next.

Of note, Grayscale submitted a TAO ETF application in January 2026, signaling institutional capital's entry into this race. Additionally, in December 2025, Bittensor will halve daily TAO emissions, and the structural tightening of the supply side is still brewing.

Original Article Link