Bittensor Subnets: Where the Revenue Actually Is

The thesis

TAO is the AI narrative tokenTokenA digital unit of value or access rights tracked on a blockchain. Tokens can represent ownership in a project, a right to use a service, a share of future revenue, or simply a tradable asset with no underlying claim.Like a physical poker chip a casino issues. The chip itself has no value. What makes it worth something is what it lets you do at the casino, what the casino has promised, and how much other people will pay you for it.Token in the glossary → most people buy. TAO itself sits at the coordination layer: the base currency, the stakingStakingLocking up a cryptocurrency to help secure a blockchain network, usually in exchange for rewards. The locked tokens act as a security deposit that can be taken away if the staker misbehaves.Like putting down a large rental deposit for an apartment. You get the money back if you behave, you earn interest while it's locked, and the landlord takes it if you trash the place.Staking in the glossary → asset. The revenue, the products, the actual AI workloads all happen in the subnets.

Bittensor subnets, often called TAO subnets, are the specialised networks where that work happens, and Bittensor runs 129 of them today. Most will fail. A handful generate revenue from paying customers. The gap between the two is where the opportunity and the risk both live.

This article examines the subnets that matter: the ones with verifiable traction, shipped products, or unusual economic models. For how the subnet emission system works mechanically, see our dTAO subnet economics deep dive. For how TAO compares to other DeAI tokens, see MOR vs TAO vs FET. For how to actually buy and stake into subnets, see our practical staking guide.

April 2026 update. Four major developments since the original publication: Templar shipped Covenant-72B (the largest decentralised pre-training run published), Manifold Labs co-authored a confidential compute whitepaper with Intel, Grayscale filed an amended S-1 for the GTAO ETF on NYSE Arca, and on 10 April 2026 Covenant AI publicly withdrew from Bittensor, accusing founder Jacob Steeves of unilateral governance control. TAO dropped roughly 15% on the exit news. The first three reshape the institutional narrative. The fourth reshapes the governance narrative entirely. For the full story, see our Templar exit deep-dive. The underlying revenue picture has since been measured rather than cited, and it’s thinner than Pine Analytics’ $3-15M: the demand revenue anyone outside the subnets can settle runs at a few percent of what the network mints.

Where the emissions flow

In March 2026 the top 10 subnets captured 56.46% of all TAO emissionsEmissionsNew tokens created and distributed by a blockchain protocol over time as rewards to validators, stakers, or miners. Emissions fund network security and participation at the cost of diluting existing holders.Like a company that pays employees partly in newly printed shares. Every year the total number of shares goes up, which means existing shareholders own a slightly smaller slice of the same company unless the company grows faster than the printing.Emissions in the glossary →. Our own chain index now reads the top 10 at 38.9%, so concentration has eased since March, but it remains high.

Under dTAO, emission allocation is market-driven. Stakers vote with their TAO by depositing into subnets they believe produce value. The more TAO staked into a subnet, the more emissions it receives. This is Darwinian selection applied to AI infrastructure: produce value or lose your emissions to someone who does.

In an early-2026 snapshot of the staking split, 72.5% of staked TAO was still in Root (the legacy staking mechanism). Only 27.5% had moved to alpha (subnet) staking. Most TAO holders were still parking in Root for conservative yield rather than actively allocating to subnets. “Every holder is a VCVCVenture Capital. Private investors who fund projects at an early stage in exchange for equity or token allocations. VC rounds are typically pre-launch, at steep discounts to any future public price, with multi-year vesting.Like angel investors in a startup who buy shares before the company goes public. They take more risk because the company might fail, so they get a better price. Once the company IPOs they can sell, and the public market pays whatever price it thinks is fair.VC in the glossary →” hasn’t materialised at scale.

The Rayon Labs question

One team dominates the network. Rayon Labs operates three subnets: Chutes (SN64), Gradients (SN56), and Nineteen (SN19). Combined, they commanded 23.71% of all TAO emissions in early 2026 (Messari, OAK Research). Our own chain index read the three at 7.97% on 22 September 2026, so that share has fallen sharply.

This is a centralisation concern. Bittensor’s thesis is Darwinian competition among independent teams. When one operator captured a quarter of emissions across three subnets, the “decentralised AIDeAIDecentralised AI. An umbrella term for blockchain-based projects that build AI infrastructure (compute, data, inference, models, agents) without a single central provider controlling the system.Like the difference between streaming a movie from Netflix and sharing it via BitTorrent. Netflix is fast and polished but one company controls what you can watch and what you pay. BitTorrent is messier but no single operator can shut you out.DeAI in the glossary →” framing needs honest qualification. Rayon Labs earns those emissions by building products people use. That’s the system working as designed. But it’s also concentration risk.

Subnet by subnet

Chutes (SN64): the revenue machine

Chutes is the subnet everyone points to when they say Bittensor has revenue from paying customers. A serverless inference platform where developers deploy AI models without managing infrastructure.

The numbers tell a growth story. Rayon Labs reports 160 billion tokens processed per day and 9.1 trillion tokens cumulative across 400,000 users. These are self-reported figures. But the OpenRouter integration provides independent verification of throughput: Chutes is listed as a provider on OpenRouter with 18 models and visible daily throughput data showing billions of tokens processed.

The revenue model is straightforward. Developers pay $3-20/month for platform access, or use the APIAPIApplication Programming Interface. A structured way for one piece of software to talk to another. In DeAI, APIs let applications request inference from a model without running the model themselves.Like a waiter in a restaurant. You don't walk into the kitchen and cook your own meal. You tell the waiter what you want, they tell the kitchen, the kitchen cooks it, and the waiter brings it back. The API is the waiter.API in the glossary → with per-token pricing (from $0.22/M tokens for Mistral). Revenue auto-stakes into the alpha token, creating a flywheel where usage drives token demand. A functioning business with revenue beyond emission farming.

GitHub activity is substantial: 1,077 commits on the chutes-api repository.

The honest caveat: the growth from 5-6 billion tokens/day (reported in early 2026) to 160 billion/day (reported March 2026) is extraordinary. A 30x increase in roughly two months. That trajectory needed watching, and the watch was warranted.

April 2026 update

OpenRouter’s public provider data shows Chutes peaked at roughly 42 billion tokens per day on 7 February 2026 and has been declining through March, averaging 8-12 billion tokens per day. That’s a meaningful gap from Rayon Labs’ 160B/day claim and contradicts the narrative of explosive growth. Some industry analysts have cited “approaching $10M ARR” for Chutes (0xSammy, April 2026) but the figure is single-source and the OpenRouter throughput data points the other direction. Pine Analytics estimated $1.3-2.4M in March 2026; our own measurement from OpenRouter’s published series now puts it at $591.3K to $3.7M a year.

The subsidy question

Pine Analytics, an independent Web3 analytics firm, published a bear case analysis on 24 March 2026 that puts Chutes’ economics in a different light. At the 14.4% of network emissions Pine measured, Chutes received approximately 518 TAO/day. Our own chain index reads Chutes at 6.55% on 15 August 2026 and Celium at 5.72%, so the gap between them narrowed sharply after this survey, and by 22 September 2026 Celium (7.08%) had overtaken Chutes (5.82%). Against estimated external revenue of $1.3-2.4M, Pine put that at a subsidy ratio of 22-40:1. Customer payments fund a small fraction of operations; TAO emissions fund the rest.

The pricing math follows directly. Pine Analytics calculates that if Chutes had to cover its costs from customer revenue alone, the break-even price would be approximately $1.41 per million tokens. Together.ai currently charges $0.88/M for Llama 70B. Unsubsidised Chutes would be 1.6-3.5x more expensive than a centralised alternative, not cheaper. The cost advantage that makes Chutes attractive to developers depends entirely on the emission subsidy continuing at its current rate.

Chutes is far from worthless. What the subsidy ratio does is sharpen the sustainability question. Pine’s estimate of $1.3-2.4M in external revenue from paying customers is meaningful for a crypto-native compute platform. The question is whether that revenue can grow fast enough to justify the emissions allocated to produce it, and whether a future halvingHalvingA protocol event that cuts the rate of new token emissions by half. Halvings are scheduled in advance, happen automatically at fixed intervals, and are a core mechanism for enforcing declining token supply growth over time.Like a savings account where the interest rate is contractually cut in half every four years. You still earn interest, but the rate drops on a known schedule, and the issuer can't change it without breaking the contract.Halving in the glossary → that cuts those emissions further would trigger a pricing adjustment or a user exodus.

Gradients (SN56): training for everyone

Gradients makes AI modelModelA trained neural network that takes inputs (text, images, audio) and produces outputs (more text, classifications, generated content). In DeAI the model is the thing that actually does the work.Like a very experienced apprentice who has spent years watching thousands of masters make furniture. They can't explain how they know when a joint is right, but they can make a chair that looks and functions like a Chippendale. The training is invisible. The output is what matters.Model in the glossary → fine-tuningFine-tuningThe process of taking a pre-trained model and training it further on a smaller, more specialised dataset to adapt it to a specific task, domain, or style. Fine-tuning is much cheaper than training from scratch.Like hiring an experienced general practitioner doctor and giving them six months of focused training in a sub-speciality. You don't have to teach them medicine from scratch. You just narrow their expertise to the area you actually need.Fine-tuning in the glossary → accessible without code. Upload your data, pick a base model, click train. The platform handles everything: text and image model trainingTrainingThe one-time process of teaching a neural network to perform a task by showing it massive amounts of example data and adjusting its internal weights until the outputs are good. Training builds the model; inference uses it.Like the years an apprentice spends learning a trade. You don't see any of the actual work, just thousands of repeated mistakes gradually becoming competence. By the end, the apprentice can do the job. The training was invisible, but the skill is now permanent.Training in the glossary →, RLHF, alignment tuning. Currently on version 5.0.

Cost advantage is the pitch. Multiple sources report training costs of $100-500 on Gradients versus $10,000+ on Google Vertex. Rayon Labs published a comparison showing Gradients v3 at $5/hour versus AWS at $67.50/hour and Google at $19.25/hour. These comparisons are self-reported and the exact workloads may differ, but the directional cost savings are consistent with other decentralised compute pricing we have seen on Akash and Render.

The important caveat: Gradients captured 6.66% of TAO emissions in March 2026, meaning a significant portion of its pricing competitiveness is funded by TAO inflationInflationThe annual rate at which new tokens are created and added to the circulating supply. Most networks use inflation to pay validators, stakers, and infrastructure providers from freshly minted tokens rather than real revenue.Like a landlord who raises the rent every year. If your salary goes up at the same rate, you break even. If it doesn't, you get poorer without noticing, because the number on your payslip hasn't changed but the ground under it has shifted.Inflation in the glossary → rather than structural efficiency. As with Chutes, the “cheaper than Google Cloud” headline needs to be read alongside the emission subsidy question. The cost gap narrows considerably if you account for the TAO being issued to support the network.

Gradients ranked #3 in that survey; our chain index read it at 1.34% on 22 September 2026. The product runs at gradients.io with community-trained models visible on the platform. Subnet-specific code remains closed on GitHub, which limits independent technical verification.

Nineteen (SN19): speed as a feature

Nineteen is Rayon Labs’ inference subnet focused on ultra-low latency. Multiple sources describe it as holding a “world record for fastest LLMLLMLarge Language Model. A neural network trained on vast amounts of text to predict the next word in a sequence. Modern LLMs (GPT, Claude, Llama, Qwen, DeepSeek) generate human-quality text and are the foundation of most modern AI products.Like an autocomplete that read every book ever written. It has no memory of individual texts but it has absorbed the patterns of language so deeply that it can generate paragraphs that sound human. The skill is statistical, not conscious.LLM in the glossary → inference.” The product is live at nineteen.ai for text and image generation.

At 2.71% of emissions in March 2026 (rank #9), and 0.81% on our 22 September 2026 chain read, Nineteen is the smallest of the Rayon Labs trio. It serves a different market from Chutes: where Chutes optimises for scale and serverless deployment, Nineteen optimises for speed on latency-sensitive workloads.

Targon (SN4): confidential compute with Intel

Targon and Chutes carry the two largest revenue claims on the network. Manifold Labs runs Targon. The thesis is GPUGPUGraphics Processing Unit. Originally designed to render video game graphics, GPUs turned out to be exceptionally good at the massively parallel math that AI models need. Modern AI training and inference runs almost entirely on GPUs.Like a factory with 10,000 workers doing the same simple task in parallel, versus a CPU which is more like 10 workers each doing different complex tasks. AI training involves doing simple math a million times per second on a million numbers, which is exactly what the GPU factory is designed for.GPU in the glossary → compute and inference with Intel TDX confidential computingConfidential ComputeHardware-enforced computation where data and code are encrypted in memory and only the authorised application can access them. The machine's operator cannot read what the application is doing even though they own the machine.Like renting space in a bank vault. The bank owns the building and runs the security, but what you put in the vault is invisible even to the bank staff. Only you have the key.Confidential Compute in the glossary → on the CPU side and NVIDIA Confidential Computing on the GPU side. Workloads run in encrypted enclaves where neither the operator nor the host hardware can inspect the data.

The headline news is the whitepaper Manifold Labs co-authored with Intel on 23 March 2026: “Decentralized Compute on Untrusted Hardware Using Intel TDX and Encrypted CVMs.” It gives Targon a formal security architecture document co-authored with a major semiconductor manufacturer. Intel hosting it on their own community blog gives it more credibility than the typical crypto-native partnership announcement.

Manifold Labs raised $10.5M in a Series A to expand the confidential compute infrastructure. Manifold reported $10.4M ARR in April 2026, a self-reported figure targon.com no longer publishes, and Pine Analytics flags it as “not an audited number” with “absence of a live revenue dashboard.” Targon also powers the Dippy AI app, which itself claims 4-8.6M users across platforms. The Dippy integration is the strongest external usage signal, though the customer-revenue split between Targon and Dippy isn’t publicly disclosed.

The interesting question for Targon is whether the Intel partnership translates into enterprise contracts that justify the valuation, or whether it remains a credibility marker on the marketing page. Revenue exists; the question is what it scales to. Confidential compute is an established enterprise need, particularly for healthcare and financial services. Targon is positioned for that market in a way few other Bittensor subnets are.

Templar (SN3): decentralised training breakthrough

Templar completed Covenant-72B on 10 March 2026: a 72 billion parameter model pre-trained on 1.1 trillion tokens across 70+ nodes using commodity internet connections. No centralised cluster. No whitelist. Anyone with GPUs could join or leave freely.

This is significant. The technical innovation is SparseLoCo, a compression technique that combines top-k sparsification with 2-bit quantisationQuantisationCompressing an AI model by storing each parameter with fewer bits of precision. Quantisation cuts model size and inference cost by 2-4x with small quality losses, making big models practical to run on consumer hardware.Like printing a high-resolution photo at lower DPI. The image is mostly the same, the details are slightly less crisp, and the file size drops dramatically. For most uses you can't tell the difference. For some uses the quality loss matters.Quantisation in the glossary → and error feedback to reduce communication overhead by 146x. Without this, the bandwidth requirements would make decentralised training over home internet connections impractical. The results are published on arXiv.

Templar captured 5.62% of emissions at rank #5 in March 2026, and 2.71% on our 22 September 2026 chain read. Following Covenant-72B, the team launched a successor competition called Templar: Crusades, focused on optimising training efficiency. The Crusades repository lives in the one-covenant org.

April 10 update: Covenant AI has publicly withdrawn from Bittensor. The team that built Covenant-72B and ran SN3, SN39, and SN81 announced their exit on 10 April 2026, accusing OTF leadership of unilateral emission suspension, moderation removal, infrastructure deprecation, and timed token sales. TAO dropped 15% on the news. The technical achievement (Covenant-72B) survives the exit intact and the paper is unaffected, but anyone holding SN3 alpha tokens is now in a difficult position with the original team disowning the subnet. Full analysis in our Templar exit deep-dive.

Sportsensor (SN41): prediction meets Polymarket

Sportsensor runs a decentralised sports predictionInferenceRunning a trained AI model to produce an answer. Inference is what happens when you type a prompt into ChatGPT and get a response. The model takes your input, computes a best guess, and returns it.Like asking an expert for their opinion. The training was the decades they spent becoming an expert. The inference is the 30 seconds it takes them to answer your specific question.Inference in the glossary → network. Miners compete to build the best models for predicting outcomes across NFL, NBA, MLB, NHL, EPL, and MLS. The meta-model aggregates the best predictions and trades them on Polymarket through proxy wallets.

The revenue model is direct: Almanac (the front-end) charges a 1% fee on winning trades. Revenue buys back the SN41 token. This is a verifiable revenue loop tied to Polymarket trading.

ContangoDigital, a VC investing in DeAI, leads the team and also operates @AskBillyBets, a public-facing prediction agent powered by Sportsensor intelligence. The Polymarket integration is confirmed in the GitHub source code.

The concern: both sportsensor.io and sportsensor.ai domains are currently dead. For a product claiming traction, having no live website is a red flag. GitHub engagement is low (4 stars, 116 commits). The 427% NBA trading profit cited by 0xJeff comes from Sportsensor’s own trading report and could not be independently verified.

Score (SN44): computer vision for sports

Score takes a different approach to sports AI. Rather than predicting outcomes from data, it uses decentralised computer vision to annotate football (soccer) video footage. Miners compete to annotate game footage accurately and quickly, and that annotated data feeds into player scoring models.

Big headline here: DKING (now rebranded to SIRE) secured a $300 million AUM deployment from a leading sports-betting hedge fund. Announced publicly at the Proof of Talk event and covered by crypto media, with ex-FIFA data scientists behind the meta-model. SN44 is reported to average roughly 70% prediction accuracy.

Caveat: dking.bot has an expired SSL certificate, which isn’t a great look for a product managing $300M in strategies. Partnership announcement is verifiable but ongoing performance isn’t transparent.

Zeus (SN18): weather forecasting

Zeus is a weather forecasting subnet using ERA5 reanalysis data from the Copernicus Climate Data Store. Miners compete to produce more accurate forecasts than traditional baselines. The product is live at zeussubnet.com with active development (v1.5.6, 97 commits).

0xJeff mentions the commodity trading application: hedge funds would pay for weather intelligence that helps forecast commodity prices. Zeus’s own website doesn’t make this claim directly. Weather forecasting is shipped and live. Commodity trading revenue? Plausible, but unsubstantiated.

Metanova Labs (SN68): drug discovery

Metanova runs a decentralised drug screening platform. Miners submit molecular candidates from the SAVI 2020 library (1.75 billion synthesisable compounds). Validators evaluate binding affinity using PSICHIC, a GNN-based oracle. A recent combinatorial update enables generation of billions of additional synthesisable molecules.

One of the most intellectually interesting subnets. The model turns drug discovery from a centralised, decade-long, billion-dollar process into a decentralised competition where participants screen compounds for therapeutic potential. Wet lab testing is described as “imminent” by Bittensor.ai.

GitHub activity is strong (451 commits, updated March 13 2026). The DeSci use case is substantive. The risk is the gap between virtual screening and actual drug candidates. Computer models identify promising molecules. Chemistry determines if they work.

What’s emerged since March

Three subnets that weren’t covered in the original analysis have grown enough to deserve attention.

Lium (SN51): the revenue we can check

Lium runs a GPU rental marketplace at lium.io, and it’s now the clearest revenue case on the network because most of it can be read. Lium published its monthly billing at a keyless endpoint until 7 October 2026, when the endpoint began requiring a validator signature: $1.3M in its last full month, 3.0x the same month a year earlier. Providers get 95% of billing, paid daily on chain at 17:00 UTC, and our read of a week of those payouts implies $19.3M a year, close to the $15.5M Lium’s billing annualises to.

The money then moves in the opposite direction to what “customers pay the providers” suggests. Lium’s validator routes every slot of its “burn” pool to UID 47, which its own code calls the “Datura payout key”, and that hotkey takes 82.8% of SN51’s miner incentive. The alpha it earns pays the providers’ rental share, so their rental income is freshly emitted alpha rather than renter money passed through. Lium’s provider docs describe this pool as burn; the deployed validator config sends it to Lium’s payout key.

Renter payments in TAO go the other way. They arrive at the subnet owner’s coldkey, which over the last 30 days spent 1,932 TAO buying SN51 alpha and burned 22,091 alpha in the same transactions. The burns are lumpy, though, so in any 30-day window the spend can run ahead of the TAO received or behind it, with the difference held in the owner’s balance. The TAO actually arriving over those 30 days, 2,888 TAO, would buy back about 45.0% of the alpha routed to the payout key, so revenue offsets only part of the providers’ emission.

For a holder that is the best version of the subnet model on Bittensor today: revenue is recorded, payouts are on chain, and TAO-paid revenue buys and burns the subnet’s token. Two things remain open. Card payments are invisible to the chain, and so is who the renters are, which matters because a miner on another subnet paying with emission income would be the subsidy going round again.

Affine (SN120): the co-founder bet

Affine is a reinforcement learningRLReinforcement Learning. A training paradigm where an AI agent takes actions in an environment, receives rewards or penalties for the outcomes, and learns a policy that maximises long-term reward. Used heavily for aligning modern LLMs.Like training a dog with treats. Good behaviour gets a treat. Bad behaviour gets nothing or a reprimand. Over many repetitions the dog learns which behaviours produce treats and starts doing them on purpose.RL in the glossary → subnet where miners compete to train and improve models on tasks like program synthesis, reasoning, and code generation. The mechanism is winner-takes-all: only the Pareto-optimal model gets rewarded. Sybil-proof, copy-proof by design.

What makes Affine notable is the team. Jacob Steeves, one of Bittensor’s co-founders (publicly known as “Const”), stepped down as CEO of the Opentensor Foundation to build Affine directly. When the founder of the protocol decides his most valuable contribution is building inside the protocol rather than running it, that’s a signal. The mechanism is interesting, and the team move is the bigger story.

No revenue. No external product. Pure research subnet at this stage. Track it for who’s behind it; nothing has shipped yet.

Ridges AI (SN62): autonomous coding agents

Ridges runs a marketplace of autonomous software engineering agents that compete to fix bugs, write tests, and generate code. The conceptual competitor is Devin, Cursor, GitHub Copilot Workspace. The differentiator is decentralisation: instead of one company building one agent, many teams compete on benchmarks and the best ones get rewarded.

Up 85% in March 2026 by alpha token price. No commercial product. No verified user metrics. Early enough that anything could happen, but the autonomous coding agent space is one of the few areas where Bittensor’s competitive marketplace structure has a plausible advantage over centralised alternatives.

The application layer: CreatorBid and Bitstarter

Two platforms are building infrastructure around subnets rather than running them directly.

CreatorBid is positioning itself as the distribution network for Bittensor. It functions like Virtuals Protocol but for the TAO network: an agent launchpad where teams build products on top of subnet intelligence. Notable agents include $AION (prediction vaults using SN6), $SIRE/DKING (sports via SN44), $SHOGUN (InfoFi), and $TAOLOR (research copilot). The TAO Council initiative provides integration, partnership, and incubation support for subnet teams. CreatorBid is launching v2 with Trade Compass (winner of the Endgame Hackathon) as the first project.

Bitstarter is a fundraising platform for subnets. It enables subnet owners to sell future emissions or OTC alpha tokens at a discount to raise operating capital. Recent raises include Quasar (SN24, 400 TAO for long-context LLM), Djinn (600 TAO for market intelligence), and Handshake58 (220 TAO for AI payment infrastructure). The model offers tiered discounts: Quasar gave 25% discount with 3-month lock or 40% discount with 6-month lock. Half the TAO raised goes to buybackBuybackUsing protocol revenue to purchase tokens on the open market, usually to burn them or return them to a treasury. Buybacks convert business income into upward pressure on the token by reducing circulating supply.Like a public company using profits to repurchase and retire its own shares. The cash leaves the company's balance sheet, the share count drops, and every remaining shareholder owns a slightly bigger slice of the same business.Buyback in the glossary → to maintain alpha token price. Bitstarter’s team is anonymous, which is a risk factor.

How to evaluate a subnet

Six signals separate working products from emission farms:

Run those signals honestly and most of the 129 subnets fall into the “weak” column on multiple counts. The handful that score “strong” across the board are where the opportunity sits.

The risk framework

0xJeff’s DeAI Matrix is the most useful risk classification:

The cash cows (decentralised compute and inference) generate 8-9 figure annualised revenue according to 0xJeff. These are self-reported aggregate figures and should be treated as directional rather than precise. But the pattern is clear: compute and inference have product-market fit. Everything else is still proving itself.

For an independent view, Pine Analytics estimated $3-15M in identifiable external revenue across the entire network (bear case analysis, 24 March 2026). The lower bound reflects Pine’s estimate for Chutes ($1.3-2.4M) and a handful of smaller subnets. The upper end incorporates Targon’s self-reported $10.4M projection, which Pine Analytics flags as unaudited and unverifiable on-chain because blockchain records token movements but not API calls. External revenue is structurally difficult to verify for inference subnets.

We now measure the part of that picture anyone can check, and it’s a narrow part. Chutes is the only Bittensor-linked provider among OpenRouter’s 102, and OpenRouter publishes its daily token volume, so it’s the one demand figure in the whole network that a subnet can’t write itself. Priced at Chutes’ own listed rates it comes to $591.3K to $3.7M a year, against a network that mints 3,576 TAO a day to pay participants.

Beyond that, the subnets’ own current figures are the next tier. Lium’s billing, which its on-chain payouts back up, Chutes’ list-price usage and Bitcast’s dashboard sum to $19M a year, or 5.2% of the emission budget. Adding Targon’s $10.4M, which its team gave an analyst in September 2025 and has not restated, gives a generous ceiling of $29.4M, or 8.0%.

The institutional shift

Two institutional developments since March deserve mention because they change how outside capital can access TAO without addressing the underlying revenue questions.

Grayscale GTAO. Grayscale filed Amendment No. 1 to its Form S-1 for the Grayscale Bittensor Trust on 2 April 2026, seeking to list on NYSE Arca under the ticker GTAO. The filing structure is creation/redemption in 10,000-share baskets, in-kind TAO or cash via authorised participants. The trust traded on OTCQX when this was written in August 2026. If the NYSE Arca listing approves, GTAO becomes the first US-listed regulated wrapper for direct TAO exposure. That’s a meaningful access point for institutional capital that won’t custody crypto natively.

TAO Institute SRI. Industry analyst 0xSammy announced the Subnet Risk Index (SRI) in early April, positioned as institutional-grade risk evaluation for Bittensor subnets. The methodology hasn’t been publicly disclosed and the institute’s affiliations are unclear. If it ships with credible methodology, it could become the reference benchmark for subnet risk that Pine Analytics’ bear case has already started to displace.

Neither development changes the measured revenue picture; both expand the audience that will be looking at it. The bear case from Pine Analytics ($3-15M demand-side revenue against an emission budget more than an order of magnitude larger) becomes more uncomfortable when an ETF wrapper and an institutional risk index sit on top of it.

The honest assessment

Bittensor’s subnet model is the most interesting coordination mechanism in DeAI. It creates direct incentives for teams to build AI products, compete on quality, and find customers. The Darwinian pressure has teeth: subnets that don’t produce value lose emissions and eventually get deregistered.

Lium is now the standout. Its billing grew 3.0x in a year, its payouts are on chain, and its TAO-paid revenue buys and burns its own token, which makes it the strongest evidence that a subnet can run a business the chain can partly audit. Chutes, the earlier standout, published flat revenue until it withdrew the series in mid-2026, and its own token volume has fallen sharply since January.

But the concentration is concerning. One team (Rayon Labs) captured nearly a quarter of emissions in early 2026. Several subnets with prominent claims have dead websites. Self-reported metrics are pervasive and independent verification is difficult. The 72.5% Root staking ratio in early 2026 suggested most TAO holders weren’t actively participating in the subnet economy that’s supposed to be the network’s core value proposition.

The subnets to watch are the ones where revenue comes from external customers rather than from TAO emissions recycling through the system. Chutes has this with OpenRouter. Sportsensor has it with Polymarket (if the product survives). Score has it with the DKING hedge fund partnership (if the performance materialises). Metanova could have it if wet lab results validate the computational screening.

For a Bittensor holder, the subnet layer is where fundamental value either materialises or doesn’t. TAO is the bet on the ecosystem. Subnets are where you find out if the bet is paying off.