Bittensor (TAO) Bearish Thesis: An Income Desert Beneath the Computing Power Myth

Author: Pine Analytics

Translated by: Saoirse, Foresight News

TAO is currently trading at approximately $275, with a market capitalization of $2.6 billion and a fully diluted valuation of $5.8 billion. The project has received institutional backing from Grayscale (which filed an application with the NYSE for a TAO ETF in December 2025) and public endorsement from NVIDIA CEO Jensen Huang. Its token supply narrative is highly compelling: a hard cap of 21 million tokens and a Bitcoin-style halving mechanism. Following its first halving in December 2025, daily issuance dropped from 7,200 to 3,600 tokens. Within one year, the number of subnets increased from 32 to 128. Templar’s Covenant-72B training further demonstrated that decentralized compute can power large language models competitive with industry benchmarks.

This report does not dispute the above facts. Instead, we examine whether the network’s economic model can generate genuine external revenue sufficient to justify its current valuation—and how competitively it stacks up against centralized service providers and self-hosted compute solutions.

Bittensor (TAO) Token Distribution Breakdown

How Network Value Flows

Bittensor has four participant categories:

• Subnet owners build specialized AI markets and receive 18% of the subnet’s TAO issuance rewards;
• Miners execute AI tasks (inference, training, data processing) and earn 41%—approximately 1,476 TAO per day, worth roughly $148 million annually;
• Validators score miner outputs and earn 41%;
• Stakers deposit TAO into subnet liquidity pools to receive subnet-specific tokens.

Under the Taoflow model, a subnet’s reward share is determined by net TAO staking inflows; negative net inflows yield zero rewards. The top ten subnets control approximately 56% of total network issuance.

TAO serves as the universal utility token across the network: miners register, validators stake, subnet tokens are purchased, and services are paid for—all using TAO. In theory, subnet activity generates structural demand for the underlying token.

Cost Comparison: Bittensor Subnet Chutes (SN64) vs. Centralized LLaMA 70B Inference Services

Demand-Side Reality

Transparent Supply vs. Opaque Demand

Bittensor’s supply side is highly transparent: 3,600 TAO are distributed daily according to programmable rules; halving schedules are hardcoded; staking rates (~70%), distribution ratios, and on-chain liquidity data are all publicly verifiable.

But the demand side remains entirely opaque. There is no unified dashboard tracking external revenue by subnet. Actual AI service usage—inference, computation, training—occurs off-chain and leaves no trace on the blockchain. Investors must rely on indirect proxies such as staking flows, subnet token prices, or project-reported metrics to infer demand. This opacity is structural—not temporary. Blockchains record token transfers, not API calls.

Below is the most comprehensive demand-side snapshot available as of March 2026.

Chutes (SN64): Low Prices Entirely Subsidized

Chutes accounts for 14.4% of total network issuance—the highest among all subnets. Developed by Rayon Labs, it offers serverless inference for open-source models at prices 85% lower than AWS and 10–50% lower than Together AI. Its usage dominates the ecosystem: over 400,000 users (including >100,000 API users), more than 5 million daily requests, cumulative processing of 9.1 trillion tokens, and a three-day average token generation volume surging from 6.6 billion to 101 billion tokens. It ranks among OpenRouter’s top inference providers, with certain models outperforming centralized competitors.

Yet this low pricing stems not from operational efficiency—but from subsidies.

Based on its 14.4% share, Chutes receives ~518 TAO daily, valued at ~$52 million annually. Its reported external annual revenue stands at only $1.3–2.4 million (the higher figure comes from team disclosures and lacks independent audit). The protocol’s subsidy ratio for this subnet ranges from 22:1 to 40:1. For every $1 users pay, the network inflates $22–40 in newly minted TAO to subsidize operations.

Removing subsidies and back-calculating cost based on its ~101 billion daily token throughput yields an unsubsidized cost of ~$1.41 per million tokens. By comparison, current centralized market rates are:

• Together.ai’s LLaMA 3.3 70B Turbo: ~$0.88 per million tokens;
• DeepSeek V3: ~$0.40–0.80 per million tokens;
• Smaller models: as low as $0.18 per million tokens.

Thus, without subsidies, Chutes’ pricing would be 1.6–3.5× higher than centralized alternatives. Its claimed 85% cost advantage reverses entirely—the “low price” is effectively borne by TAO holders via inflation, not enabled by structural efficiencies of decentralization.

At the next halving (expected late 2026 or 2027), Chutes faces three unpalatable options: double its prices, lose miners, or widen the gap between subsidies and revenue even further.

Some may draw parallels to early internet-era customer acquisition subsidies (e.g., Uber, DoorDash, AWS). But those platforms built real switching costs during their subsidy phases: proprietary infrastructure, driver networks, enterprise ecosystems. Bittensor subnets have none: models are open source, interfaces standardized, and users can switch providers at zero cost. Once subsidies end, no lock-in mechanisms exist to retain users.

Rayon Labs also operates SN56 and SN19, collectively commanding ~23.7% of total network issuance—neither of which discloses external revenue. A single team effectively controls nearly one-quarter of the network’s incentive allocation.

Targon, Templar, and Other Subnets

Targon (SN4), operated by Manifold Labs, is the highest-revenue subnet, offering confidential GPU compute services to enterprises. Its estimated annual revenue is ~$10.4 million, implying a valuation of ~$48 million and a price-to-sales ratio of ~4.6x—the most grounded valuation within the ecosystem. However, the $10.4 million figure is drawn from multiple third-party reports and remains unaudited.

Templar (SN3) successfully trained Covenant-72B and holds a $98 million market cap—but reports zero external revenue. Its training API and enterprise sales efforts remain ongoing, with no paid products launched yet.

The remaining 120+ subnets either disclose no revenue or remain in early product stages, surviving primarily on token issuance subsidies.

Overall Picture

Total confirmed annual demand-side revenue across the entire network amounts to only ~$3–15 million. Chutes alone receives ~$52 million in annual subsidies—exceeding the upper bound of the network’s total external revenue.

With a $2.6 billion market cap, Bittensor trades at a revenue multiple of ~175–200x; at its $5.8 billion fully diluted valuation, the multiple approaches 400x. By contrast, centralized AI compute firms raised capital recently at forward-revenue multiples of just 15–25x, while high-growth SaaS companies rarely sustain multiples above 50x long term. Bittensor’s valuation multiples sit 4–10× higher than even the most aggressive peers in the sector.

This massive divergence between valuation and fundamentals reveals that TAO’s pricing is driven almost entirely by supply-side scarcity (halvings, staking lockups), institutional catalysts (Grayscale ETF, exchange listing expectations), and AI-sector sentiment—not by real economic output. These are valid price drivers, but they bear no logical connection to the proposition that “Bittensor functions as a sustainable AI service network.”

Comparison of Hyperscale Cloud Providers’ AI CapEx vs. Bittensor (TAO) Annual Subsidy Budget

Pricing Dilemma: Squeezed From Both Ends

Subnets face dual pressure:

• From above: Self-hosting caps pricing

All models on the platform are open source and weights publicly available. Running a 70B model on a single H100 costs only $40–50 per day. Tools like vLLM and Ollama make local deployment trivial. Next-generation NVIDIA chips will further reduce inference costs. Institutions with sufficient scale will find self-hosting cheaper.

• From below: Cloud giants intensify competition

Microsoft, Google, Amazon, and Meta collectively spent over $200 billion on AI capex in 2025. They enjoy hardware priority allocations, dedicated data centers, enterprise relationships, and cross-subsidization from other business lines. Bittensor’s annual incentive budget (~$360 million) is less than Microsoft’s weekly AI infrastructure spend. Even specialized service providers compete aggressively on open-source models, backed by VC funding.

Subnet pricing is thus compressed into an extremely narrow band—while bearing inherent decentralized costs: token friction, validator overhead, subnet owner fees, and network latency.

Moat Question

Even if a subnet delivers valuable services, its underlying models and methods are inherently public: Covenant-72B uses the Apache License, and technical papers are openly published. Competitors can replicate it directly—without participating in the TAO ecosystem at all.

Traditional moats—proprietary technology, network effects, switching costs, brand—simply don’t apply:

• Technology is open source;
• Network effects accrue to TAO—not individual subnets;
• Model weights are identical across providers, making user switching costless.

The community often cites incentive design as the moat—but this depends on sustained, large-scale token issuance, which shrinks with every halving.

What Is TAO Really Trading?

At a $2.6 billion market cap, TAO’s price reflects neither demand fundamentals nor the $3–15 million in annual revenue—which is insufficient to support such valuation under any traditional framework. What the market is actually trading is: Bitcoin-style scarcity, Grayscale ETF expectations, AI-sector rotation, and the long-dated option value of decentralized AI. These are legitimate speculative factors—but they derive entirely from the supply side and market sentiment.

If you hold TAO based on scarcity and narrative, you may profit even amid weak demand. But if you believe Bittensor will evolve into a large-scale, economically viable AI service network, there is currently no evidence supporting that view—and formidable structural headwinds stand in the way. Investors must clearly distinguish their own investment logic.