Where will TAO's valuation go when subsidies retreat?
Written by: Pine Analytics
Compiled by: Saoirse, Foresight News
TAO is currently priced at approximately $275, with a market capitalization of $2.6 billion and a fully diluted valuation of $5.8 billion. The project has received endorsements from Grayscale (an ETF listing application was submitted to the NYSE in December 2025) and has also been publicly recognized by NVIDIA CEO Jensen Huang. Additionally, the narrative around token supply is highly attractive: the total supply is capped at 21 million coins, utilizing a Bitcoin-like halving mechanism. After the first halving in December 2025, the daily issuance will drop from 7,200 coins to 3,600 coins. Within a year, the number of subnets increased from 32 to 128, and Templar's Covenant-72B training has also proven that decentralized computing power can produce large language models with competitive benchmarks.
This report does not deny the above facts. What we want to explore is: can the economic model of this network generate real external revenue that supports the current valuation scale, and how competitive is it when competing with centralized service providers and self-hosted computing power?
Bittensor (TAO) token issuance distribution ratio
How network value circulates
Bittensor has four types of participants:
Subnet owners build specialized AI markets and receive 18% of the TAO issuance rewards from the subnet;
Miners execute AI tasks (inference, training, data processing) and receive 41%, totaling about 1476 TAO daily, with an annualized value of about 148 million dollars;
Validators score the output of miners, receiving 41%;
Stakers put TAO into subnet liquidity pools in exchange for subnet-specific tokens.
Under the Taoflow model, a subnet's reward share is determined by the net inflow of TAO staked; if the net inflow is negative, there will be no rewards. The top ten subnets control about 56% of the total network issuance.
TAO is a universal token across the network: miners register, validators stake, subnet tokens are purchased, and service payments all require TAO. Theoretically, subnet activities will bring structural demand for the underlying token.
Comparison analysis of the inference cost of Bittensor subnet Chutes (SN64) and centralized service provider LLaMA 70B model
Demand side status
Supply transparency vs demand opacity
The supply side of Bittensor is highly transparent: 3600 TAO are allocated daily according to the program, halving rules are hard-coded, staking rates (about 70%), distribution ratios, and liquidity data are all on-chain.
But the demand side is completely opaque. There is no unified dashboard to track external revenue by subnet, and the actual calls for AI services (inference, computation, training) occur off-chain and are not recorded on the blockchain. Investors can only infer demand through indirect indicators such as staking flows, subnet token prices, and self-reported data from project parties. This opacity is structural, not a temporary phenomenon. The blockchain only records token transfers, not API calls.
The following is the most complete demand-side portrait as of March 2026.
Chutes (SN64): Low price relies entirely on subsidies
Chutes accounts for 14.4% of the total network issuance, the highest among all subnets. Developed by Rayon Labs, it provides open-source model serverless inference services, priced 85% lower than AWS and 10%-50% lower than Together AI. Its usage data is unparalleled in the ecosystem: over 400,000 users (over 100,000 API users), daily request volume exceeding 5 million times, processing a cumulative 9.1 trillion tokens, with average token generation over three days soaring from 6.6 billion to 101 billion. It is also a leading inference service provider on OpenRouter, with some models outperforming centralized competitors.
But this low price does not come from operational efficiency, but from subsidies.
Based on a 14.4% share, Chutes receives about 518 TAO daily, with an annualized value of about 52 million dollars. Its external annual revenue is only about 1.3 million to 2.4 million dollars (the higher value is self-reported by the team and has not been independently audited). The subsidy ratio for this subnet is about 22:1 to 40:1. For every 1 dollar paid by users, the network needs to release 22-40 dollars worth of TAO through inflation to subsidize.
If subsidies are removed, based on its daily processing capacity of about 101 billion tokens, the cost price is about 1.41 dollars per million tokens. The current centralized market price is:
Together.ai's LLaMA 3.3 70B Turbo costs about 0.88 dollars / million tokens;
DeepSeek V3 costs about 0.40–0.80 dollars;
Small models can be as low as 0.18 dollars.
This means that without subsidies, Chutes' price would be 1.6 to 3.5 times more expensive than centralized solutions. The so-called 85% cost advantage is completely reversed; its low price essence is that TAO holders pay through inflation, rather than structural efficiency brought by decentralization.
When the next halving arrives (expected at the end of 2026 or 2027), either the price will double, miners will exit, or the gap between subsidies and revenue will further widen.
Some people will compare customer acquisition subsidies in the early internet, but Uber, DoorDash, and AWS established switching costs during the subsidy period: proprietary platforms, driver networks, corporate ecosystems. However, Bittensor subnets have no barriers: models are open-source, interfaces are standardized, and users can switch service providers at zero cost. Once subsidies retreat, no lock-up mechanism can retain users.
Rayon Labs also operates SN56 and SN19, collectively controlling about 23.7% of the total network issuance, without disclosing external revenue. A single team almost controls a quarter of the network's incentive distribution.
Targon, Templar, and other subnets
Targon (SN4) is the highest revenue subnet, operated by Manifold Labs, providing confidential GPU computing services to enterprises, with estimated annual revenue of about 10.4 million dollars, corresponding to a valuation of 48 million, with a price-to-sales ratio of about 4.6 times, making it the most solid valuation in the ecosystem. However, 10.4 million is merely predicted data cited by multiple reports and not an audited figure.
Templar (SN3) completed the Covenant-72B training, with a market value of 98 million dollars, but external revenue is zero. The training API and enterprise sales are still progressing, and no paid products have been launched yet.
The remaining 120+ subnets either have no public revenue or are still in the early product stage, mainly relying on token issuance subsidies to survive.
Overall overview
The total annually confirmable demand-side income across the network is only about 3 million to 15 million dollars. Just the annual subsidy for one subnet, Chutes (about 52 million dollars), exceeds the upper limit of the entire network's external revenue.
At a market value of 2.6 billion dollars, its revenue multiple is about 175-200 times; at a fully diluted valuation of 5.8 billion, it approaches 400 times. In contrast, centralized AI computing firms have recently raised valuations of only 15-25 times projected income, and high-growth SaaS rarely maintains above 50 times long-term. Bittensor's valuation multiple is 4-10 times that of aggressive industry targets.
The huge gap between valuation and demand fundamentals indicates that the market prices TAO almost entirely based on supply-side scarcity (halving, staking lock-up), institutional catalysts (Grayscale ETF, expected listing), and sentiment in the AI sector, rather than real economic output. These are indeed price drivers, but they are entirely separate from the logic of 'Bittensor creating sustainable value as an AI service network.'
Comparison of AI capital expenditures of super-large cloud vendors with the annual subsidy scale of Bittensor (TAO)
Pricing dilemma: squeezed from both sides
Subnets face pressure from both ends:
Above: Self-hosted cap
All models on the platform are open-source, weights are public, and the comprehensive cost of running a 70B model on a single H100 is only 40-50 dollars per day, with tools like vLLM and Ollama making local deployment extremely simple. NVIDIA's next-generation chips will also significantly reduce inference costs. Institutions with sufficient usage will find self-built deployments cheaper.
Below: Cloud giants squeeze
Microsoft, Google, Amazon, and Meta's total AI capital expenditure in 2025 exceeds 200 billion dollars, with hardware priority quotas, dedicated data centers, and corporate customer relationships, and can subsidize AI with cash flow from other businesses. Bittensor's annual incentive budget (about 360 million dollars) is less than Microsoft's weekly AI infrastructure investment. Professional service providers are also competing on low prices using VC subsidies on open-source models.
Subnet pricing is compressed within a very narrow range and also bears the unique costs of decentralization: token friction, validation node expenses, subnet owner shares, network latency, etc.
Moat issue
Even if a certain subnet provides valuable services, the underlying models and methods are inherently open: Covenant-72B adopts the Apache protocol, and the technical paper is publicly published. Any competitor can directly replicate without participating in the TAO ecosystem.
Traditional moats (proprietary technology, network effects, switching costs, brand) do not hold:
Technology open-source;
Network effects belong to TAO, not to individual subnets;
Model weights are consistent, and user switching costs are zero.
The community believes that the incentive mechanism is the moat, but this relies on continuous large-scale token issuance, and each halving will continuously shrink the incentive budget.
What exactly is TAO trading?
At a market value of 2.6 billion dollars, the price of TAO does not reflect demand fundamentals; annual income of 3 to 15 million cannot support it under any traditional framework. What the market trades is: Bitcoin-like scarcity, Grayscale ETF expectations, rotation in the AI sector, and the long-term option value of decentralized AI. These are all reasonable speculative factors but come entirely from the supply side and market sentiment.
If you hold TAO based on scarcity and narrative, you might profit even if demand is weak; but if you believe Bittensor will become a truly scalable AI service network, there is currently no evidence and it faces structural resistance that is difficult to overcome. Investors should clearly differentiate their investment logic.