Inside Rayon Labs: One Team, a Quarter of Bittensor

Bittensor’s token is having a very good month. TAO trades near $329, up almost 30% on the week, and it has climbed back toward the top 30 crypto assets by market value. The rally has a familiar shape: an AI-linked token catching a bid whenever the wider market rotates toward artificial intelligence. What the price does not tell you is whether the network underneath it sells anything at all. That question, the one that actually matters for a project that brands itself as decentralized AI, runs almost entirely through a single team most investors have never heard of: Rayon Labs.

Rayon Labs builds three of Bittensor’s busiest subnets, and between them they collect close to a quarter of every TAO the protocol mints. One of them, an inference platform called Chutes, is the rare Bittensor subnet that genuinely sells a product to paying customers. It is also the clearest case study in the network’s central tension: emissions, not customers, still pay most of the bills. This is a look at who Rayon Labs is, what it actually earns, why the numbers are so contested, and what a single team controlling that much of a supposedly decentralized network says about the whole experiment.

A token flying, a question standing still

Start with the market, because it is loud right now. According to CoinGecko, TAO recently changed hands around $329.01, up roughly 6% on the day and close to 30% over seven days, with a market capitalization near $3.74 billion at rank #33 and a fully diluted valuation around $6.92 billion. Circulating supply sits at about 11.34 million of a hard-capped 21 million, twenty-four-hour volume runs near $367 million, and the token is still about 57% below its March 2024 all-time high of $757.60. The five-day surge, roughly 25% by one tally of the catalysts, tracked the usual mix: AI sector rotation, fresh trading venues for TAO, and lingering hope for a US spot exchange-traded fund.

None of that answers the harder question. Bittensor is a network of roughly 128 competing subnets, each one a market for a specific machine-learning task, all bidding for a shared pool of newly minted TAO. The pitch is Darwinian: independent teams compete, the useful ones win emissions, and out of that competition emerges decentralized AI that no single company owns. The uncomfortable reality, documented all year by on-chain analysts, is that most subnets produce very little external revenue and survive almost entirely on protocol subsidies. If you want to test whether Bittensor is a business or a subsidy machine, you do not start with the price chart. You start with the one operator that both sells the most and takes the most.

What Rayon Labs actually is

Rayon Labs is harder to define than a normal startup, and that is partly the point. In its own description and in third-party research, it reads less like an incorporated company and more like a tight collective of developers, some of whom work in the open and some of whom stay anonymous. A recent OAK Research profile describes it as a decentralized group of builders with a strong track record of shipping functional products with polished user experience in a matter of months, unusual in a space where most subnets never get past a demo.

Crucially, Rayon Labs does not formally own the subnets it builds. Chutes sits under an entity called Chutes Global Corp, Gradients under Grads LLC, and Rayon itself is the development shop building AI services on top of Bittensor’s infrastructure. That structure matters when you try to answer who is accountable for a network that markets itself as ownerless. The team’s public products span the stack: Chutes for inference, Gradients for fine-tuning, Nineteen for low-latency serving, and a no-code agent builder called Squad that runs on Chutes and sells a Pro tier at about $40 a month. For a network sold on the idea that no one is in charge, one loosely defined team quietly became its most important operator.

That looseness cuts both ways. On one hand, it is exactly the kind of permissionless, ship-fast builder culture Bittensor was designed to attract, and Rayon’s products are among the few on the network that a mainstream developer would actually reach for. On the other, when the most important operator on a supposedly ownerless network is a partly anonymous collective working through a set of corporate shells, the ordinary questions, who is liable, who answers to customers, who is accountable if a subsidy-dependent service is wound down, do not have clean answers. For a project whose entire value proposition is trust-minimized coordination, that ambiguity sits uncomfortably close to the center.

The trifecta: almost a quarter of every TAO minted

Emissions are the subsidy that keeps Bittensor running, and they are not spread evenly. After the December 2025 halving, the protocol mints roughly 3,600 TAO a day and splits it across subnets in proportion to each subnet’s alpha-token price. In March 2026, Rayon Labs’ three subnets pulled close to 23.7% of that flow between them, according to figures compiled by Own Your Mind from Pine Analytics and on-chain data. At the TAO prices of the time, a quarter of daily issuance works out to tens of millions of dollars a year in protocol-minted rewards flowing toward one team’s orbit.

To put that in perspective, no other operator on Bittensor comes close. The network’s whole thesis rests on independent teams competing for a shared reward pool, so a single group holding a quarter of that pool is not a footnote; it is the story. It is worth keeping the concentration figure in mind through everything that follows, because the debate over Rayon Labs is really two debates stacked on top of each other: whether its flagship product is a real business, and whether one team earning this much is compatible with the word decentralized at all.

Inside Chutes: renting GPUs by the endpoint

Chutes, subnet 64, is the crown jewel and the one worth understanding in detail. It is a serverless inference platform, and the unit it sells is called a chute: an endpoint where a GPU operator deploys an open-source AI model. In the mechanics described by tao.media, miners register their hardware, stand up models as these endpoints, and developers pay to route inference requests to them. The experience for a developer looks like any modern serverless API; the difference is that the compute behind it comes from a permissionless swarm of miners rather than a hyperscaler’s owned fleet.

Those miners are ranked, not paid flat. Chutes scores them on a rolling seven-day window weighted toward compute capacity (about 55%), with speed and availability at roughly 20% each and a small slice for bounties, so the operators supplying the most reliable throughput capture the most alpha. The result is a genuine cost advantage: tao.media reports Chutes GPU inventory running at roughly 50% to 60% of market pricing for equivalent hardware, because supply is permissionless and miners are chasing TAO emissions rather than the fat margins a venture-funded cloud needs to justify its raise. On scale, the platform’s own figures are large: more than 696,000 users outside of OpenRouter, upward of 34 trillion tokens processed over its lifetime, and 50 or more models served. Chutes also powers Squad, the no-code agent builder, which makes it the inference backend for a small consumer and agent layer. That agent angle is exactly where the hard, unsolved problem of proving an inference was done honestly comes back to bite, a challenge we have written about in the context of verifiable AI and the agent economy.

Being one of the largest providers on OpenRouter matters because OpenRouter is where developers comparison-shop inference by price and speed, routing each request to whichever backend wins on the metrics. A permissionless swarm competing for emissions can win that auction on price in a way a margin-seeking cloud cannot, which is the genuinely novel thing Chutes demonstrates. The catch is the one that haunts every decentralized-compute market: when a model runs on a stranger’s GPU, proving the output was produced honestly and by the model actually requested is an unsolved problem, and it is why trust, not just price, still steers a lot of production traffic back to incumbents.

The revenue nobody can agree on

Here is where Chutes gets contentious. Rayon Labs reports roughly 120 billion tokens processed per day after monetization, peaking around 160 billion across all of its surfaces. But the one number outsiders can independently verify, the throughput Chutes routes through the public OpenRouter marketplace, tells a very different story. Own Your Mind’s review of the economics shows Chutes peaking near 42 billion tokens per day on 7 February 2026, falling to 8 to 12 billion by late March, and then averaging about 8.1 billion per day between 21 May and 17 August 2026, drifting near 6.8 billion at its low. That is roughly a twenty-four-fold gap between the headline figure and the verifiable slice.

Revenue estimates diverge just as wildly, which is the clearest sign that nobody outside the team can actually pin the number down.

A spread that runs from $1.3 million to nearly $10 million is not a rounding error; it is a disagreement about what kind of business this even is. The self-reported throughput and the verifiable throughput differ by more than an order of magnitude, and every revenue estimate sits somewhere in the fog between them. To be fair to Rayon, some of that gap is real: Chutes serves paying customers directly and through its own API, not just OpenRouter, so the public marketplace only ever showed a fraction of total activity. But the honest reading is that the verifiable portion is small, the self-reported portion is unaudited, and the difference between them is the entire investment case.

The subsidy underneath the discount

Why is Chutes so cheap? Part of the answer is genuine efficiency from permissionless supply. The larger part, by the numbers, is subsidy. Pine Analytics calculated the emissions subsidy at somewhere between 22 and 40 times customer revenue, meaning TAO inflation, not customer payments, funds the overwhelming majority of the operation. Strip that away and the pricing story inverts.

The break-even math is stark. To cover its costs from customers alone, Own Your Mind estimates Chutes would need to charge roughly $1.41 per million tokens, against about $0.88 per million at a centralized comparator like Together.ai. In other words, once you remove the emission subsidy, the decentralized option is more expensive than the venture-funded incumbent it is supposed to undercut. Zoom out to the subnet level and the imbalance is even starker: Own Your Mind’s accounting puts one major subnet, Chutes, on the order of $52 million a year in received emissions against roughly $2.4 million in genuine external revenue. Across the whole network, Pine’s bear case pegged identifiable external revenue at just $3 million to $15 million for early 2026. This is the income desert that Bittensor’s critics keep pointing at, and Chutes, precisely because it sells more than anyone, is the most flattering example of it.

When your best customers are the problem

There is a revealing wrinkle in Chutes’ own history. In a February 2026 announcement, the team admitted that its heaviest users were extracting between 56 and 324 times the value of their subscriptions, according to Own Your Mind’s write-up. That is what happens when you sell a flat-rate subscription against a metered, GPU-priced cost: sophisticated users treat the subscription as an arbitrage and mine it for everything it is worth. Chutes responded by eliminating its free tier and downtiering some models to bring unit economics back under control.

That episode reframes the growth metrics. When usage is partly subsidized, both by TAO emissions on the supply side and by underpriced subscriptions on the demand side, the impressive user and token counts measure appetite for a discount as much as willingness to pay. It also complicates the mechanism Rayon points to as proof of real demand: the platform funnels revenue back into buying its own subnets’ alpha tokens, an auto-staking loop that is supposed to create organic, usage-linked buy pressure. That is elegant when the underlying usage is real. It is circular when the usage is subsidized, because the protocol mints TAO to subsidize inference, that inference generates revenue, and the revenue is used to buy the alpha token whose emissions funded the subsidy in the first place. The loop can look like demand while mostly recycling the subsidy.

The Emission Gate finally bit

Bittensor spent 2026 rewriting its emission rules to attack exactly this problem, and the most consequential change is the Emission Gate, shipped in the protocol’s version 440 update. Rather than letting a subnet’s alpha price alone set its share of new TAO, the Emission Gate ties compensation to demonstrated, market-driven demand, throttling subnets that cannot show real usage behind their token price. It was a direct answer to the emission-farming critique.

It bit, and it bit the biggest earner first. Chutes’ emission share fell from 14.39% in March 2026 to 5.82% by September, per Own Your Mind’s tracking. In roughly two quarters, the single largest recipient on the network lost more than half of its subsidy. Read charitably, the gate is working exactly as designed: reward proven demand, starve the farms. Read skeptically, the size of the drop is itself evidence of how much of Chutes’ original lead was subsidy rather than sales, and it revives the complaint that has shadowed dTAO all year, that the protocol keeps changing the scoring rules while the game is being played. Either way, a team that once collected one in every seven TAO minted now collects closer to one in seventeen, and it happened without its customer base changing much at all.

That volatility is the recurring complaint about dTAO from the people who trade it. When the rules that determine a subnet’s income can be rewritten mid-cycle, modeling a subnet token becomes less like valuing a business and more like guessing the next protocol upgrade. Supporters counter that a young network has to be able to fix broken incentives quickly, and that a mechanism which halves the biggest earner’s subsidy is proof the system is not captured. Both things can be true at once: the Emission Gate is probably good protocol design and a genuine headache for anyone who underwrote a subnet token on the old rules.

Gradients and Nineteen: the quieter two thirds

Chutes gets the attention, but Rayon’s other two subnets are where the concentration debate gets sharper, because they are quieter and, by the strictest public accounting, thinner on revenue. Gradients, subnet 56, is a model fine-tuning platform that lets users customize open-source models with techniques like supervised tuning and reinforcement learning from human feedback, priced around $5 per hour of training according to OAK Research. It held about 6.66% of emissions in March 2026. Own Your Mind is blunt about it, classing Gradients among the emission farms and noting its pricing competitiveness is funded by TAO inflation rather than structural efficiency, with no independent revenue verification.

Nineteen, subnet 19, launched in April 2025 as an ultra-low-latency inference layer aimed at real-time assistants and autonomous agents. It carried roughly 2.71% of emissions in March 2026 and, again per Own Your Mind, has no documented external revenue. So of the trifecta’s near-quarter of the reward pool, the two smaller subnets look closer to emission farms than to businesses on the public numbers. That is not unusual on Bittensor; it is the norm across most of the 128 subnets. What is unusual is how much of the norm is concentrated under one team.

The short list of subnets that actually sell

Rayon’s subnets are not the only ones with a revenue story, but the list of subnets with verifiable external income is short, and that brevity is the real headline. In its revenue rankings, Own Your Mind puts the network’s measured revenue floor at roughly 6.4% of the total emission budget, which is another way of saying that more than nine of every ten TAO minted flow to subnets that cannot yet show matching customer income. A handful of subnets do run real, if modest and often unaudited, businesses.

Targon, subnet 4, is the one subnet that claims more revenue than Chutes, roughly $10.4 million in self-reported annual recurring revenue as of April 2026, built on confidential GPU compute using Intel’s trusted-execution hardware; Own Your Mind flags the figure as unaudited, with no public revenue dashboard to check it against. Score points to a $300 million allocation from a sports-betting hedge fund, and Sportsensor takes a 1% cut of winning trades on Polymarket, but ongoing performance for both is hard to see from the outside. The point is not that Bittensor produces nothing. It is that the entire network’s verifiable external revenue is small enough to fit in a single paragraph, and even after the Emission Gate cut its share, Rayon Labs still sits at the center of the part that is real.

The concentration question decentralized AI keeps dodging

Put the pieces together and the awkward truth is plain: on a network whose entire pitch is that no single entity controls it, one loosely defined team runs a quarter of the reward budget across three subnets held by separate legal shells, staffed by a mix of public and anonymous developers. When one operator captures that much, the phrase decentralized AI needs honest qualification, and the honesty is coming from inside the house.

Cofounder Jacob Steeves said it himself. In a June 2026 decentralization roadmap, covered by Crypto Briefing, he opened with the admission that “Bittensor is currently not a decentralized protocol in the way Bitcoin is,” adding, “It can be, and it will be, but it isn’t yet.” The roadmap set a target of full decentralization by December 2027 and argued that the foundation deliberately traded away decentralization for development speed during the network’s hyper-growth phase. It is a candid statement, but it is also an admission that the core economic layer, the one Rayon Labs sits at the top of, remains centralized today.

The dispute turned bitter in April 2026, when Sam Dare, founder of Covenant AI, the team behind the Templar training subnet, quit the network and torched the decentralization narrative on the way out. As reported by The Block, Dare called the promise that no single entity controls Bittensor “a lie” and described the reality as “decentralization theatre.” TAO fell about 15% on the day, sliding from $338 to $285 within two hours before partially recovering. Whatever the merits of that specific feud, it dragged the concentration question into the open. The closest analogy is Bitcoin mining, where a network can be permissionless in principle and concentrated in practice; the question of who actually secures the network has a Bittensor cousin, and its name is emission concentration.

For anyone holding TAO or a subnet’s alpha, the concentration is not an abstract governance concern; it is a risk factor. If one team’s three subnets drive a large slice of the network’s perceived usage and buy pressure, then that team’s product decisions, its disputes, and its treatment under each rule change ripple through prices that holders never chose to bet on. The Covenant exit, which knocked 15% off TAO in an afternoon over a governance fight most retail buyers had never followed, is the template. A network that concentrates this much influence in a few hands inherits the fragility of those hands, which is precisely the opposite of what a decentralized design is supposed to buy you.

How the money flows: dTAO, alpha and the halving clock

To follow the incentives, you need the plumbing. Since dTAO launched in February 2025, every subnet has its own alpha token and an automated market maker pool that prices it. That alpha price is what sets the subnet’s share of network emissions, which is why a subnet’s token and its subsidy are joined at the hip, and why the alpha markets are watched so closely. Rewards inside a subnet are then split according to the protocol’s emission rules: about 18% to the subnet owner, 41% to miners, and 41% to validators and their delegators, all paid in that subnet’s alpha rather than in TAO directly.

Supply is capped at 21 million, the same as Bitcoin, and issuance halves on a supply schedule rather than a fixed block count. The first halving landed in December 2025, cutting issuance from one TAO per block to half a TAO, roughly 3,600 a day; the next is not expected until around 2029. Around 70% of TAO is staked, at yields in the neighborhood of 10%. For Rayon Labs, both the auto-staking flywheel and the subsidy depend on that emission stream, which makes the 2029 halving a slow-moving deadline: Own Your Mind estimates Chutes would need on the order of fifteen times its baseline revenue to be self-sustaining by then. The staking dynamics that make TAO attractive to hold, and the trade-offs of locking tokens for yield, will feel familiar to anyone who has weighed the economics of running an Ethereum validator.

One more mechanic is worth knowing, because it flatters the supply figures. Launching and maintaining a subnet requires burning TAO in registration fees, a recycling mechanism that permanently removes tokens from circulation and helps explain why circulating supply, around 11.34 million, sits below cumulative issuance. Recycling offsets some of the inflation the emissions create, but it does not change the core problem: those tokens are burned by teams competing for a subsidy, not by customers paying for a service, so the burn is another expression of the same emissions-first economy rather than a sign of external demand.

The ETF that cannot stake, and where subnet tokens go next

The ETF hope driving part of the rally is real but unresolved. Grayscale’s Bittensor Trust, ticker GTAO, is still a private-placement trust, not an exchange-traded fund. Its most recent filing, a Form 8-K dated 10 September 2026 on SEC EDGAR, states that the sponsor intends to rename the trust to Grayscale Bittensor Trust ETF and list its shares on NYSE Arca once the registration statement becomes effective, which it has not. No US spot TAO ETF has been approved, and TAO’s would-be conversion sits in the same queue as the broader wave charted in our global map of 2026 crypto ETF approvals.

There is a deeper catch specific to TAO: the staking paradox. A US spot fund almost certainly cannot stake its holdings, which means an ETF buyer would get price exposure but none of the roughly 10% staking yield, none of the alpha, and none of the governance that make TAO economically interesting in the first place, while the staked majority of the network keeps diluting them. European staked-TAO products, by contrast, do stake and pass the yield through. It is the same tension a home validator weighs, transplanted into a wrapper, and it means the most-hyped path to mainstream TAO exposure may hand investors the least attractive version of the asset.

Meanwhile, subnet tokens are starting to leave home. Project Rubicon, built by General TAO Ventures, wraps subnet alpha into standard ERC-20 tokens and bridges them to Base using Chainlink’s cross-chain messaging, so alpha can trade inside mainstream DeFi rather than only in Bittensor’s native pools. That is the same subnet-as-an-asset logic Rayon’s auto-staking started, pushed a step further into the world of cross-chain bridges and messaging. It also raises the stakes on the revenue question. Once alpha becomes a liquid asset that DeFi users hold without ever touching a subnet, the gap between subsidized emissions and real customer revenue stops being an insider debate and becomes everyone’s exposure.

What to watch: regulation, the summit, and the sustainability clock

On regulation, the frame in the United States is straightforward even if the outcome is not. The SEC governs the securities questions hanging over the network: the ETF path, and the token-sale allegations that surfaced during the Covenant dispute. A subnet token that stakes, yields, and now trades on a mainstream chain starts to resemble a security in the eyes of a US regulator, which is the quiet risk sitting underneath the Rubicon-style financialization. For a project already fighting to prove it is decentralized, more securities-law attention is the last thing the core team wants, and it is a real possibility as alpha tokens go liquid.

On the calendar, the OpenTensor Foundation’s Exploit Summit runs 28 and 29 September in Montreal, the network’s marquee event and the kind of venue where roadmap and subnet news tends to land. Steeves’s December 2027 decentralization target is the governance clock; the roughly 2029 halving is the economic one. Between now and then, the single most important number to track is not the TAO price but whether any subnet, most plausibly one of Rayon’s, can grow verifiable external revenue fast enough to survive with less subsidy.

The risks stack in a specific order. The first is the subsidy-to-revenue gap: if emissions keep funding most of the network’s activity, then every rule change that ties rewards to real demand, like the Emission Gate, will keep repricing subnet tokens downward, and the roughly 2029 halving will do it mechanically. The second is concentration: the more the network’s story depends on one team, the more a single dispute can move the whole token. The third is regulatory: as alpha tokens become liquid, yield-bearing assets on mainstream chains, they look more like securities, and the SEC has shown little patience for that pattern. None of these is fatal on its own. Together they explain why the gap between TAO’s price and Bittensor’s revenue is the number that matters.

The bottom line is that Rayon Labs is simultaneously the best and the most honest test of Bittensor’s thesis. It ships real products, it has real if disputed users, and Chutes is the closest thing the network has to a functioning business. It is also the clearest evidence that emissions, not customers, still pay most of the bills, and that decentralized remains an aspiration rather than a description. If any team can turn subsidy into a self-funding business before the next halving, it is probably this one. If none can, the trifecta is exactly where you will see it break first.

Frequently Asked Questions

What is Rayon Labs?

Rayon Labs is a development collective behind three of Bittensor’s most prominent subnets: Chutes (SN64), Gradients (SN56) and Nineteen (SN19). It is not a single incorporated company, and the subnets sit under separate entities such as Chutes Global Corp and Grads LLC, staffed by a mix of public and anonymous developers. Together its subnets have commanded close to a quarter of all TAO emissions.

What is Chutes on Bittensor?

Chutes is a serverless AI inference platform running as subnet 64. GPU operators deploy open-source models as endpoints called chutes, and developers pay to send inference requests to them. It is one of the largest inference providers routed through OpenRouter and prices GPU time well below typical market rates, though analysts say that discount depends heavily on TAO emission subsidies.

How much revenue does Chutes actually make?

Estimates vary widely. Pine Analytics put annualized customer revenue at $1.3 million to $2.4 million in March 2026, Own Your Mind measured a $1.1 million to $5.6 million range, tao.media cited about $5.5 million, and Rayon Labs itself claimed it was approaching $10 million in annual recurring revenue. Verifiable OpenRouter throughput sits far below the team’s self-reported figures, which is why the real number is contested.

Is Bittensor actually decentralized?

Not fully, by its own founders’ account. Cofounder Jacob Steeves wrote in June 2026 that Bittensor is not yet a decentralized protocol in the way Bitcoin is, and set a December 2027 target to change that. The concentration of emissions in a few teams, including nearly a quarter under Rayon Labs, is a central reason critics question the decentralized-AI label.

Is there a Bittensor (TAO) ETF?

Not in the United States as of late September 2026. Grayscale’s Bittensor Trust (GTAO) remains a private-placement trust that the sponsor intends to convert into an exchange-traded fund on NYSE Arca once its registration becomes effective, but no US spot TAO ETF has been approved. European staked-TAO products exist and, unlike a likely US spot fund, can stake and pass on yield.

By Marcus Okafor, HOGE Wire markets desk.