Leading AI platform Bittensor: Technology is evolving, are users fleeing?

Author: @BlazingKevin_ , Blockbooster Researcher

The integration of Web3 and AI is moving beyond the early stages. Market scrutiny of AI-focused crypto projects is shifting from early “concept hype” to “fundamentals and technological implementation.” In this transition, projects demonstrating strong resilience and technological breakthroughs are reshaping the market valuation system.

1. Bittensor firmly maintains its leading position

The current total market cap of the AI cryptocurrency sector is approximately $17.46 billion, with 24-hour trading volume close to $1.94 billion. Among these, Bittensor (TAO) holds the top spot with a market cap of about $3.43 billion. It accounts for nearly 19.6% of the entire AI crypto market share, establishing an absolute leading position.

A straightforward comparison with core competitors reveals its ecological niche:

Core competitive barriers

Bittensor’s core competitive advantage is its unique “Proof of Intelligence” (PoI) network. It moves beyond simply providing computing power. The network introduces complex incentive mechanisms that directly reward high-quality AI model outputs. This positioning is unique among competitors and very difficult to replicate easily.

2. Real “value creation” capability verification and valuation logic reshaping

Setting aside grand technological visions, the key to a Web3 protocol’s ability to traverse bull and bear markets is its genuine business expansion and revenue-generating capacity.

In the crypto market, Bittensor demonstrates rare real value creation ability. According to data from Q1 2026, the Bittensor network earned approximately $43 million from genuine AI clients (not fake transactions driven by token incentives). This figure has already surpassed the annual revenue of many traditional Web3 protocols.

Core valuation indicators (as of March 29, 2026):

Traditional centralized AI infrastructure companies typically enjoy a forward revenue valuation of 15-25x in private markets. Bittensor benefits from high liquidity premiums, network effects, and scarcity narratives. Its current P/S ratio of about 20x is within a reasonable or even undervalued range. The total market cap of subnet tokens within its ecosystem has reached $1.47 billion. This ecosystem structure feeds back into the mainnet TAO’s value capture.

3. SN3’s breakthrough

Financial data establishes the protocol’s valuation floor. Technological breakthroughs in decentralized training have thoroughly opened up its valuation potential.

The core driver behind TAO’s countercyclical rise is not mere speculation. The underlying technology has achieved a historic breakthrough. Its valuation logic is shifting fundamentally from “narrative-driven” to “product-driven.”

3.1 Covenant-72B validates the feasibility of decentralized training

On March 10, 2026, Bittensor’s subnet Templar )SN3( and the Covenant Labs team behind it published a technical report on arXiv. The team announced the successful pretraining of Covenant-72B, the largest fully decentralized, permissionless dense architecture model trained to date.

This model has 72 billion parameters, trained on 1.1 trillion tokens. Its MMLU score reached 67.1, comparable in basic performance to Meta’s LLaMA-2-70B. The model broke through the communication bandwidth bottleneck in decentralized training. The introduction of the SparseLoCo algorithm played a key role. Nodes only need to transmit 1%-3% of core gradients and perform 2-bit quantization, achieving over 146x data compression (reducing 100MB data to below 1MB). Under normal internet bandwidth, computational utilization remains high at 94.5%. This milestone proves that globally distributed, heterogeneous compute resources can produce cutting-edge models with commercial competitiveness. The technical solution eliminates reliance on expensive InfiniBand dedicated lines and centralized supercomputing clusters.

The success of Covenant-72B quickly caused shockwaves in the traditional AI community:

• Jack Clark, co-founder of Anthropic, highly praised it: On March 16, he extensively cited this breakthrough in his research report, calling it “challenging AI political economy through distributed training.” He pointed out that this is a technology worth continuous tracking and foresees device-side AI adopting such decentralized training models widely in the future.

• Jensen Huang’s “Folding@home” analogy: On March 20, in the All-In VC podcast, Chamath introduced Bittensor’s technological achievement to NVIDIA CEO Jensen Huang. Huang responded positively, comparing it to “a modern version of Folding@home,” and affirmed the necessity of coexistence between open-source and distributed models.

3.2 Two core components of SN3: solving communication efficiency and incentive compatibility

Dozens of untrusting, hardware-diverse, network-quality-variable nodes collaborate to train the same 72B model. SN3 relies on two core components to address bandwidth and malicious behavior:

SparseLoCo (solving communication efficiency): Traditional distributed training synchronizes full gradients every step, resulting in huge data volume. SparseLoCo allows each node to run 30 internal optimization steps (AdamW) locally. The node then compresses and uploads the “pseudo-gradient.” It uses top-k sparsification (retaining only 1%-3% of core gradient components), error feedback, and 2-bit quantization. This process achieves over 146x data compression (reducing 100MB to below 1MB). Under normal internet conditions (upstream 110 Mbps, downstream 500 Mbps), computational utilization remains high at 94.5%. Each training round takes only 70 seconds.

Gauntlet (solving incentive compatibility): This component runs on the Subnet 3 blockchain. It verifies the quality of each node’s submitted pseudo-gradient. The system tests how much the model loss decreases after applying the node’s gradient (LossScore). It also checks whether nodes are training on assigned data (to prevent cheating). Each round, only the highest-scoring node’s gradient is aggregated. This mechanism fundamentally solves the problem of “how to prevent miners from slacking” in decentralized scenarios.

4. Subnet ecosystem and the super leverage of dTAO mechanism

In 2025, Bittensor launched the dynamic TAO (dTAO) mechanism. This mechanism played a key “amplifier” role in the recent rise. dTAO allows each subnet to issue independent Alpha tokens. Subnets establish liquidity pools with TAO via automated market maker (AMM) mechanisms.

4.1 Leverage effect of subnet tokens

Under the dTAO mechanism, the price of subnet tokens is directly determined by the amount of TAO staked in the subnet pool. As TAO appreciates, the underlying reserve value of all subnets rises accordingly. This passive increase in subnet token prices creates a positive feedback loop, attracting more speculative and staking funds to lock TAO into subnets.

Key subnet tokens’ 30-day price increases:

As shown, driven directly by the Covenant-72B success, the Templar )SN3( token surged over 440% in a single month, reaching a market cap of $130 million. This subnet-level wealth effect is evident. The total market cap of subnet tokens reached $1.47 billion by the end of March, with daily trading volume surpassing $118 million. This effect acts as a “super leverage,” transmitting strong buy pressure back to the TAO main token.

4.2 Vertical ecosystem integration

While Covenant Labs operates SN3, it also plans SN39 (Basilica, focused on compute services) and SN81 (Grail, focused on reinforcement learning fine-tuning and evaluation). This vertical integration covers the entire process from pretraining to alignment optimization. The layout demonstrates that Bittensor’s ecosystem has formed a complete decentralized AI industry chain loop.

5. Token distribution

Based on on-chain data from taostats and CoinMarketCap as of March 29, 2026, Bittensor’s network health can be deeply assessed across several dimensions:

Overall on-chain assessment:

Bittensor’s on-chain data exhibits characteristics of an extremely healthy economy. High staking rates lock liquidity; real income supports fundamentals; the dTAO mechanism stimulates subnet innovation. Continuous supply-side tightening (including halving and high staking) combined with demand growth (institutional entry and AI narratives strengthening) creates a highly advantageous price dynamic model.

6. Valuation concerns

It is important to note that on-chain transparency mainly reflects supply-side data, while demand-side (actual AI service call volume) off-chain remains a significant blind spot:

Risk 1: Excessive token subsidies mask true business costs. Currently, most subnets rely heavily on TAO inflation subsidies for low-cost services. For example, the leading inference subnet Chutes )SN64( has a subsidy-to-revenue ratio of 22-40:1. Excluding token subsidies, actual service pricing is far above that of centralized competitors, with service premiums 1.6 to 3.5 times higher than platforms like Together.ai. Ongoing halving cycles will eventually expose the fragility of this business model.

Risk 2: Lack of a strong business moat leads to rapid user churn. Bittensor mainly offers open-source models and standardized APIs, fundamentally different from cloud giants like AWS. The ecosystem lacks proprietary platforms, deep enterprise integrations, or data flywheels that create “lock-in effects.” Developer migration costs are very low. Once token subsidies decline, price-sensitive B2B users will quickly leave. Cheaper centralized compute platforms can easily absorb this outflow.

Risk 3: Valuation dislocation after data squeezing. Regarding the $43 million quarterly revenue, some cautious institutional analyses suggest a different valuation model. Excluding related-party transactions and subsidies, and focusing on verified external fiat revenue, the network’s annual revenue could drop to between $3 million and $15 million. Using this “squeezed” real revenue base, the network’s P/S ratio could soar to 175-400x, posing a significant bubble risk and potential valuation collapse.