Peeling back the layers of Bittensor's valuation, what's left after the subsidies recede?

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 is backed by Grayscale (which filed for an ETF listing on the NYSE in December 2025) and has received public approval from Nvidia CEO Jensen Huang. Its token supply narrative is also highly attractive: a total supply cap of 21 million tokens, employing a Bitcoin-style halving mechanism. After the first halving in December 2025, the daily issuance will decrease from 7,200 to 3,600 tokens. Within a year, the number of subnets has increased from 32 to 128, and Templar's Covenant-72B training has demonstrated that decentralized computing power can produce large language models with benchmark competitiveness.

This report does not deny the above facts. What we want to explore is whether the network's economic model can generate real external revenue to support its current valuation, and how competitive it is in relation to centralized service providers and self-hosted computing power.

Bittensor (TAO) token issuance allocation ratio

How does network value transfer?

Bittensor has four types of participants:

Subnet owners who build professional AI marketplaces will receive 18% of the subnet's TAO issuance rewards;
Miners earn 41% by performing AI tasks (inference, training, data processing), totaling approximately 1,476 TAO per day, with an annualized value of approximately US$148 million.
The validator scored the miners' output and received 41%;
Stakers deposit TAO into the subnet liquidity pool in exchange for subnet-specific tokens.

Under the Taoflow model, the reward share of a subnet is determined by the net inflow of TAO staking; if the net inflow is negative, there is no reward. The top ten subnets control approximately 56% of the total network issuance.

TAO is a universally accepted token across the entire network: miner registration, validator staking, subnet token purchases, and service payments all require the use of TAO. Theoretically, subnet activities will create structural demand for the underlying token.

A comparative analysis of the model inference costs of Bittensor subnet Chutes (SN64) and centralized service provider LLaMA 70B.

Current situation on the demand side

Supply transparency vs. demand opacity

Bittensor's supply side is highly transparent: 3,600 TAO tokens are allocated daily according to a program, the halving rules are hard-coded, and the staking rate (approximately 70%), allocation ratio, and liquidity data are all recorded on the blockchain.

However, the demand side is completely opaque. There's no unified dashboard to track external revenue by subnet, and the actual calls to AI services (inference, computation, training) all occur off-chain and aren't recorded on the blockchain. Investors can only infer demand through indirect indicators such as staking flows, subnet token prices, and project-reported data. This lack of transparency is structural, not temporary. The blockchain only records token transfers, not API calls.

The following is the most complete demand-side profile as of March 2026.

Chutes (SN64): Low prices are all thanks to subsidies

Chutes accounts for 14.4% of the total token issuance, the highest among all subnets. Developed by Rayon Labs, it provides serverless inference services for open-source models, priced 85% lower than AWS and 10%–50% lower than Together AI. Its usage data is unparalleled within the ecosystem: over 400,000 users (over 100,000 API users), over 5 million daily requests, a cumulative processing volume of 9.1 trillion tokens, and a three-day average token generation that surged from 6.6 billion to 101 billion. It is also a leading inference service provider on OpenRouter, with some models outperforming centralized competitors.

However, this low price does not come from operational efficiency, but from subsidies.

Based on a 14.4% share, Chutes receives approximately 518 TAOs daily, with an annualized value of about $52 million. Its external annual revenue, however, is only about $1.3 million–$2.4 million (the higher figure is self-reported by the team and has not been independently audited). The protocol's subsidy ratio for this subnet is approximately 22:1 to 40:1. For every $1 a user pays, the network releases $22–$40 worth of TAOs through inflation to subsidize it.

If subsidies are removed, and considering its daily processing volume of approximately 101 billion tokens, the cost is roughly $1.41 per million tokens. The current centralized market price is:

Together.ai's LLAMA 3.3 70B Turbo is approximately $0.88 per million tokens;
DeepSeek V3 costs approximately $0.40–$0.80.
The smallest model can be as low as $0.18.

This means that without subsidies, Chutes will be 1.6–3.5 times more expensive than centralized solutions. The so-called 85% cost advantage is completely reversed; its low price is essentially paid for by TAO holders through inflation, rather than by the structural efficiency brought about by decentralization.

When the next halving arrives (expected at the end of 2026 or 2027), either the price will double, miners will leave, or the gap between subsidies and revenue will widen further.

Some might draw parallels to early internet subsidies for customer acquisition, but Uber, DoorDash, and AWS built switching costs during their subsidy periods: proprietary platforms, driver networks, and enterprise ecosystems. Bittensor subnets, however, have no such barriers: open-source models and standardized interfaces allow users to switch service providers at zero cost. Once subsidies fade, no lock-up mechanism can retain users.

Rayon Labs also operates SN56 and SN19, controlling approximately 23.7% of the total network releases, for which no external revenue has been disclosed. A single team controls nearly a quarter of the network's incentive distribution.

Targon, Templar and other subnets

Targon (SN4), the highest-grossing subnet operated by Manifold Labs, provides confidential GPU computing services to enterprises. Its estimated annual revenue is approximately $10.4 million, corresponding to a valuation of $48 million and a price-to-sales ratio of approximately 4.6, making it the most solid valuation within the ecosystem. However, the $10.4 million figure is merely a projection cited in multiple reports and is not an audited figure.

Templar (SN3) has completed training on the Covenant-72B platform and has a market capitalization of $98 million, but has zero external revenue. Training API and enterprise sales are still underway, and paid products have not yet been launched.

The remaining 120-plus subnets either have no publicly disclosed revenue or are still in the early stages of product development, relying mainly on token issuance subsidies to survive.

Overall Overview

Total verifiable demand-side annual revenue across the entire network is only about $3 million to $15 million. The annualized subsidies for just the Chutes subnet (about $52 million) exceed the upper limit of the entire network's external revenue.

With a market capitalization of $2.6 billion, its revenue multiple is approximately 175–200 times; with a fully diluted valuation of $5.8 billion, it approaches 400 times. In contrast, centralized AI computing power companies have recently raised valuations of only 15–25 times their forward revenue, and high-growth SaaS companies rarely maintain valuations above 50 times for extended periods. Bittensor's valuation multiple is 4–10 times that of aggressively performing companies in the industry.

The huge gap between valuation and demand fundamentals illustrates that the market's pricing of TAO is almost entirely based on supply-side scarcity (halving, staking lock-up), institutional catalysts (Grayscale ETF, Shanghai Stock Exchange expectations), and AI sector sentiment, rather than real economic output. These are indeed price drivers, but they are completely different from the logic of "Bittensor creating sustainable value as an AI service network."

Comparing AI capital expenditures of hyperscale cloud vendors with the annual subsidy scale of Bittensor (TAO)

Pricing Dilemma: Caught Between Two Opportunities

The subnet is facing pressure from both ends:

Above: Self-managed capping

All models on the platform are open source with publicly available weights. The daily cost of running a 70B model on a single H100 image is only $40-50. Tools such as vLLM and Ollam make local deployment extremely simple. NVIDIA's next-generation chips will further reduce inference costs. For organizations with sufficient usage, self-deployment will be even cheaper.

Below: Cloud giants squeeze

Microsoft, Google, Amazon, and Meta's combined AI capital expenditures by 2025 will exceed $200 billion. They possess priority hardware quotas, dedicated data centers, enterprise customer relationships, and can subsidize AI with cash flow from other businesses. Bittensor's annual incentive budget (approximately $360 million) is less than Microsoft's weekly AI infrastructure investment. Professional service providers are also using VC subsidies to compete on open-source models with low prices.

Subnet pricing is compressed into an extremely narrow range, and it also has to bear the unique costs of decentralization: token friction, validator node overhead, subnet owner revenue sharing, network latency, etc.

moat problem

Even if a subnet provides a valuable service, the underlying model and methods are naturally public: Covenant-72B is licensed under the Apache License, and its technical papers are publicly available. Any competitor can directly replicate it without participating in the TAO ecosystem.

Traditional competitive advantages (proprietary technology, network effects, switching costs, brand) are no longer valid:

Open source technology;
Network effects belong to TAO, not to individual subnets;
With consistent model weights, the cost for users to switch is zero.

The community sees the incentive mechanism as a moat, but this relies on a continuous issuance of large amounts of tokens, and each halving will cause the incentive budget to shrink continuously.

What exactly is TAO trading?

With a market capitalization of $2.6 billion, TAO's price does not reflect fundamental demand; annual revenue of $3-15 million is unsustainable under any traditional framework. The market is trading on: Bitcoin-like scarcity, Grayscale ETF expectations, AI sector rotation, and the long-term option value of decentralized AI. These are all legitimate speculative factors, but they stem entirely from the supply side and market sentiment.

If you hold TAO based on scarcity and narrative, you may profit even with weak demand; however, if you believe Bittensor will become a truly large-scale AI service network, there is currently no evidence and it faces insurmountable structural resistance. Investors should clearly distinguish their investment logic.