Bittensor’s Training Bet: AI Models Without a Data Center

Bittensor’s token is having a good week for reasons that have almost nothing to do with artificial intelligence. TAO changed hands near $254 on 8 September 2026, up close to 10% over seven days and still about 66% below its March 2024 record of $757.60, according to CoinGecko. The bid is a financial-engineering story: spot-ETF filings in the United States, a European product that stakes the token, and a macro tape coiled ahead of a heavy month of policy decisions. None of it measures whether Bittensor is any good at the thing it was built to do.

Underneath the price, the most genuinely novel work on the network is quieter and much harder than the market is pricing. Over the past year, groups of strangers using Bittensor have trained large language models with no shared data center, no corporate cluster, and no single company in charge. Two of those runs, a roughly 72-billion-parameter model called Covenant-72B and a 100-billion-parameter model called Orion-100B, reached a scale that looked out of reach for permissionless networks as recently as 2024. This piece is about that capability: what it is, how it is even possible, what it actually costs, who else is chasing it, and why the token keeps looking the other way.

Inference is the easy half. Training is the hard problem.

Most crypto-AI coverage, including our own, has centered on inference: running an already-trained model to answer a prompt. Inference is comparatively easy to decentralize because each request is small and self-contained. A single GPU in someone’s basement can serve a query without ever talking to another GPU. The whole decentralized-inference sector, from Chutes on Bittensor to Akash and io.net, is built on that property, and its real constraint turns out to be demand rather than physics, a point we made in our look at who actually buys decentralized inference.

Training is a different animal. To build a model’s weights from scratch, thousands of processors have to march through the same calculation in lockstep, and after every step they must exchange gradients, the tensors that encode how each of billions of parameters should nudge. In a hyperscale data center those exchanges ride specialized interconnects like NVLink and InfiniBand that move terabytes per second between GPUs sitting centimeters apart. Try to run the same job across homes and small server rooms on different continents, wired together by ordinary broadband, and the math falls apart. The compute is willing; the network is not. That gap is why serious people spent years insisting frontier-scale training simply could not be decentralized. Bittensor’s training subnets are the loudest counterargument to that claim, and understanding them starts with the wall they had to climb.

The bandwidth wall, and the trick that beat it

Put a number on the problem. A typical consumer upload link runs around 60 megabits per second. Exchanging full gradients for a modern model over links like that, naively, would take on the order of 5,000 years for a single DeepSeek-v3-scale run, by the reckoning of researcher Jaime Sevilla at Epoch AI. No incentive design fixes a five-millennia training run. The only way through is to move dramatically less data between machines without wrecking the model’s ability to learn.

That is exactly what a wave of 2023-to-2026 research delivered. Google DeepMind’s DiLoCo showed that workers could train mostly on their own and synchronize far less often, cutting communication by roughly 500 times. Nous Research went further with DeMo, or Decoupled Momentum, which separates each worker’s momentum, applies a fast frequency-domain transform, and then ships only the largest components, reducing per-step communication by around two orders of magnitude; the method is public in an arXiv paper and became the backbone of Nous’s DisTrO optimizer family. Layer in gradient quantization (sending 4-bit instead of 32-bit numbers) and streaming variants that overlap computation with communication, and the effective bandwidth requirement drops by a factor that turns 5,000 years into something a determined network can finish in months.

Bittensor’s builders did not just borrow these ideas; they extended them. Templar’s team introduced SparseLoCo, which combines top-k sparsification with 2-bit quantization to reach, by their accounting, more than 146 times compression versus dense gradients. Macrocosmos took a different route for its pipeline-parallel design, squeezing the activations passed between pipeline stages from roughly 140 megabytes down to 2.2 megabytes. These are the unglamorous engineering wins that make everything downstream possible. Without them, there is no story here at all.

Covenant-72B: the run that proved it could work

The first Bittensor training run to command outside attention came from Subnet 3, known as Templar, operated by a team called Covenant AI that included researchers affiliated with the Mila institute in Montreal. Between September 2025 and March 2026, that subnet trained Covenant-72B, a model of roughly 72 billion parameters, on about 1.09 trillion tokens of web text plus a further 14 billion tokens of annealing data. It finished on 10 March 2026, and the network coordinated the whole run through Bittensor’s incentives rather than a private cluster, using DeMo for communication and SparseLoCo on top, as documented by industry trackers including 800.works.

What made it notable was not size alone but that it was any good. On MMLU, a standard knowledge benchmark, Covenant-72B scored 67.1, edging past Meta’s LLaMA-2-70B at 65.6, per an analysis by Own Your Mind. That is not a frontier model, and LLaMA-2 is a 2023 reference point, but it is a real, competitive, openly trained model that came out of a permissionless swarm rather than a single lab. The run averaged fewer than 17 contributing peers per round, peaking around 24. A handful of independent operators, scattered geographically and never meeting, jointly produced a 72-billion-parameter model. Two years earlier the consensus view was that this was impossible. It is the clearest evidence yet that Bittensor’s training thesis is more than a pitch deck.

Orion-100B: Bittensor’s biggest model to date

If Templar was the proof of concept, Macrocosmos is the team that kept scaling. Macrocosmos runs several Bittensor subnets and is among the network’s most prolific contributors by number of active networks, including Subnet 9, which hosts its training architecture IOTA (Incentivised Orchestrated Training Architecture), described in a technical primer on arXiv and its developer docs. IOTA breaks a model across pipeline stages so that a contributor can supply as little as a single GPU and still hold one slice of a much larger network.

In mid-2026 the team unveiled Orion-100B, a 100-billion-parameter distributed run split across 16 pipeline stages with three replicas, reported by TAO Media. The headline claim is efficiency: more than 30% model FLOP utilization and, by Macrocosmos’s own measure, roughly 65% of the training efficiency of a conventional data center, achieved on globally distributed GPUs. The project began in June 2025 aiming at a 15-billion-parameter target, was pulled back to a 1.5-billion-parameter test bed, and then, after more than 700 experiments, scaled from 1.5 billion to 100 billion parameters inside roughly a month in April 2026. Getting to 65% of data-center efficiency over the open internet, if it holds up under scrutiny, is the kind of result that quietly moves the frontier of what decentralized training can attempt next.

Two caveats keep the result honest. The efficiency and utilization figures are Macrocosmos’s own, published ahead of independent replication, so they should be read as a credible claim rather than a settled fact until outside researchers reproduce them. And a 100-billion-parameter model, impressive as it is for a distributed run, is still an order of magnitude below the largest closed frontier systems and trained on far fewer tokens than a flagship release. What Orion-100B demonstrates is not parity with the labs; it is that the efficiency penalty for going distributed has fallen from prohibitive to merely large, and is still falling. In a field where that penalty was recently assumed to be permanent, shrinking it is the whole game.

The decentralized-training scoreboard

Bittensor’s two flagship runs sit inside a small but fast-moving field. The table below lines up the runs that have actually shipped, rather than the ones that live only in whitepapers, so the on-network work can be judged against the off-network competition.

Two things stand out. First, on raw parameter count the Bittensor runs are currently the largest openly documented decentralized pretraining efforts, ahead of Prime Intellect’s 32-billion-parameter INTELLECT-2. Second, almost every serious project leans on the same small toolbox of communication-compression tricks, several of them traceable to Nous Research. This is a research community as much as a set of competitors, which is worth remembering when the marketing turns tribal.

How Bittensor pays strangers to train a model

The reason Bittensor can attract that compute at all is money, paid continuously and in the open. The network is a collection of subnets, capped today at 128 active slots and set to expand toward 256 later in 2026, up from roughly 32 at the start of 2025. Each subnet defines a task, from serverless inference to, in these cases, pretraining a model. Miners do the work, validators score it, and the protocol pays out block emissions accordingly. Since the December 2025 halving the network issues about 3,600 TAO a day, and within each subnet the split is fixed, as summarized on CoinGecko’s subnet primer.

Since the dTAO upgrade of February 2025, every subnet also has its own alpha token and a small automated market pool, and the market price of that alpha decides what share of network-wide emissions the subnet earns. Rewards are paid in alpha, not TAO directly. That design is elegant and it has a sharp edge for training in particular. Inference subnets like Chutes can at least point to outside customers paying for tokens served; training subnets sell nothing to anyone. Their entire economic engine is emissions, denominated in a token whose price is set by speculation on the subnet’s future. When Templar’s alpha ran up several hundred percent in March 2026 on the back of the Covenant-72B milestone, that was the market rewarding a research result, not revenue. It is a powerful way to bootstrap expensive work, and a fragile one, because the moment sentiment turns, the training budget evaporates with the token.

The other hard problem: proving a stranger trained honestly

Paying for work you cannot see raises an obvious question: how does the network know a peer actually did the training it claims, rather than submitting noise, recycling someone else’s updates, or quietly training a smaller model and pocketing the rewards? For inference this is tractable, because you can re-run a prompt and compare outputs, which is roughly what Prime Intellect’s TOPLOC verifier does for rollouts. For training it is genuinely hard, because reproducing a step means owning the same expensive hardware and burning the same compute, which defeats the point of checking.

Bittensor’s answer is its validators and the Yuma Consensus scoring that sits under every subnet. Validators repeatedly evaluate whether a miner’s submitted updates genuinely lower the model’s loss on held-out data, score them, and let the stake-weighted median decide who gets paid, which makes faking a contribution costly if you expect to be caught and down-weighted. It is not bulletproof. Weight-copying, where a validator lazily mirrors another’s scores instead of doing its own evaluation, has been a recurring exploit the protocol has fought with commit-reveal schemes. And an operator submitting plausible-but-degraded work is harder to catch than one submitting garbage. Verifying honest remote compute remains one of the genuinely unsolved problems in this field, on Bittensor and off it, and it is the quiet reason training subnets still lean on a small set of reputable operators rather than throwing the doors open to anyone.

About that commodity-hardware claim

The romantic version of decentralized training has your idle gaming PC contributing to a global brain overnight. The reality is more selective. The marketing around Covenant-72B leaned on phrases like commodity GPUs and ordinary residential and commercial internet, and the internet part is largely true. The hardware part is doing a lot of work in that sentence. By Own Your Mind’s accounting, a Templar peer needed on the order of eight high-end data-center GPUs, B200-class cards, to participate, a capital outlay in the range of $240,000 to $320,000 per participant. With fewer than 24 peers at peak, that is a federation of well-capitalized operators, not a crowd of hobbyists.

This matters because decentralization is being sold two ways at once. There is decentralization in the sense that no single data center or company owns the run, which Bittensor genuinely delivers. And there is decentralization in the populist sense that anyone can join with whatever they have, which, at frontier-adjacent scale, remains aspirational. Macrocosmos designs IOTA so a contributor can bring a single GPU, and consumer-tier hardware is the stated long-term goal, but the models that make headlines still run on serious kit. Investors evaluating a training subnet should read past the word permissionless and ask how many entities actually mattered to a given run. The honest answer, for now, is often a dozen or two.

The bitter irony: the team that proved it works walked out

Here is the twist that makes Bittensor’s training chapter a genuine drama rather than a spec sheet. One month after Covenant AI produced the network’s most impressive training result, it quit. On 9 April 2026 the team shut down its three subnets, Templar (SN3), Basilica (SN39) and Grail (SN81), sold a large block of TAO, and published a blistering exit statement. TAO fell about 15% within two hours, from roughly $338 to $285, as The Block reported.

Founder Sam Dare did not mince words. The founding promise that no single entity controls the network, he wrote, “is a lie.” What Bittensor calls decentralization, he added, “is not. It is decentralization theatre.” The specific charges were aimed at cofounder Jacob Steeves: that he retained effective control over the network’s governing group, suspended subnet emissions unilaterally, stripped operators of moderation rights, and used large, visibly timed token sales as pressure during disputes. Whatever the merits, the symbolism was brutal. The people who had just demonstrated that permissionless AI training could work were leaving because, in their telling, the platform underneath it was not permissionless at all.

Steeves has been unusually candid on the broader point, if not the specific allegations. In June 2026 he acknowledged to The Crypto Times that Bittensor is not yet decentralized in the way Bitcoin is, framing the current centralization as a deliberate choice to move fast; Bitcoin, he argued, had to decentralize from day one because it launched into a tightly policed financial system, whereas AI is early enough that Bittensor can grow first and hand over control later. Both founders stepped back from executive roles in February 2026, with a stated roadmap to full decentralization by the end of 2027. The tension is structural, and it rhymes with a lesson from a very different corner of crypto: distributing a job across many independent operators, as distributed validator technology does for Ethereum staking in the institutional DVT bid, removes a technical single point of failure without automatically removing the human one at the top. Bittensor solved the hard engineering problem of training without a center. It has not yet solved the softer, arguably harder problem of governing without one.

Bittensor is not the only one racing

The most useful context for Bittensor’s training subnets is that they are one entry in a crowded race, and the other entrants have made a very different bet on how to fund the work. Prime Intellect, founded in 2024, trained INTELLECT-1 at 10 billion parameters and then INTELLECT-2, a 32-billion-parameter model it calls the first globally distributed reinforcement-learning run, releasing the code and weights along with purpose-built tooling like PRIME-RL and the TOPLOC verifier, per its release notes. Crucially, Prime Intellect has no liquid token. It raised a $130 million Series A at a roughly $1 billion valuation in July 2026, led by Radical Ventures with Nvidia Ventures, Intel Capital and others, as TechCrunch reported, and it books an annualized revenue run rate near $100 million from enterprise customers. Nous Research, which authored much of the compression toolkit everyone else uses, runs its Psyche network with coordination on Solana and has trained models on decentralized infrastructure, as VentureBeat covered. Pluralis is pursuing Protocol Models, where contributors literally co-own the resulting weights.

The contrast is the whole point. Prime Intellect captures the value of its work in equity that its investors and customers can price against revenue. Bittensor captures it in a token that anyone can buy and that trades on narrative. When a training subnet succeeds, Prime Intellect’s cap table gets more valuable in a way a spreadsheet can defend; a Bittensor subnet’s alpha gets more valuable in a way that can reverse in a day. Neither model is obviously right. But if you are buying TAO because you believe in decentralized training, you should know you are taking the token version of the bet, not the equity version.

It also means the projects are not really enemies at the research level. Nous Research’s optimizers underpin runs on networks that compete with it commercially; DiLoCo came out of a corporate lab and now turns up almost everywhere; INTELLECT-2’s code is open for anyone, including Bittensor subnets, to study. The competition is over who can turn shared science into a sustainable operation, and the two live answers are Bittensor’s token-incentive machine and Prime Intellect’s equity-and-revenue company. Watching which one funds more real training over the next few years will say more about crypto-AI than any single benchmark.

How far can this really go?

It is tempting to draw a straight line from Orion-100B to a decentralized model that dethrones the labs. The arithmetic says otherwise, at least for this decade. Epoch AI’s analysis puts the largest decentralized runs to date at roughly 1,000 times less compute than genuine frontier models. Decentralized training has grown at an astonishing pace, on the order of 20 times a year since 2020, which does outrun the roughly 5-times-a-year growth of frontier training. But it is starting from so far back that, in Epoch’s blunt phrasing, we will not see decentralized runs catch the frontier in scale this decade.

The binding constraint remains communication. Compression buys orders of magnitude, but each trick has limits, and stacking them introduces its own overhead and instability. There is plenty of latent supply; Epoch estimates something like 15.7 million H100-equivalent GPUs exist worldwide, far more than any single lab commands. The hard part is orchestrating a meaningful fraction of them across bad links without the whole run grinding to a halt. So the realistic near-term prize is not a decentralized GPT-5. It is capable open models, in the 30-to-100-billion-parameter range, that no company owns and no board can quietly switch off, trained fast enough to stay useful. That is a smaller claim than the hype, and still a meaningful one. A model that cannot be revoked has properties a closed API never will, whatever its benchmark scores.

That reframing is why serious money keeps flowing into the category even with the frontier out of reach. A government, a bank, or a hospital that cannot legally or politically pipe its data into a US closed model still needs capable AI it can run and audit itself. So does anyone building in a jurisdiction where the leading labs will not operate. Open models that a swarm can keep improving, and that no single vendor can deprecate, price-hike, or geofence, are a different product from a frontier API, aimed at a different buyer. Bittensor is one of a handful of networks racing to be the place that product gets built, and on current evidence it is among the leaders on scale. Whether it can convert that lead into durable demand, rather than durable emissions, is the open question.

What the token market is actually pricing

Set the engineering beside the tape and the disconnect is stark. TAO carries a market capitalization around $2.44 billion at rank 40, on a fully diluted valuation near $5.34 billion, with a hard cap of 21 million coins and about 9.6 million circulating, per CoinGecko. Capability has climbed over the past year while the token sits roughly two-thirds below its all-time high. The September rally is not a training story. It is a financialization story.

The divergence shows up inside the ecosystem too, not just in the headline token. By early 2026 the combined market value of all subnet alpha tokens had reached somewhere around $1.12 billion, roughly a quarter of TAO’s own value, with the leaders clustered near or above $130 million each; Templar’s alpha and the Chutes inference subnet’s alpha were among the largest, per CoinGecko’s subnet data. Yet those valuations swing on narrative and emissions rather than cash flows, and Templar’s stayed richly priced right up until its operators walked out. The lesson repeats at every layer of the stack: capability and price are only loosely tethered here, and the tether is sentiment.

The pieces of that story are concrete. Grayscale’s Bittensor Trust exists, but as its own SEC filings make plain, it remains a private-placement vehicle for accredited investors rather than an approved spot exchange-traded fund; conversion filings from Grayscale and Bitwise are pending, not granted. In Europe a Safello staked-TAO product already lets holders earn the network’s yield inside a listed wrapper, a structure a plain US spot ETF would struggle to replicate. And subnet alpha tokens are escaping into broader decentralized finance: General TAO Ventures’ Project Rubicon wraps subnet alpha onto Base using Chainlink’s cross-chain messaging, which imports a fresh set of dependencies worth weighing against what cross-chain trust actually costs. All of this is happening against a jittery macro backdrop, with crypto broadly coiled ahead of a dense run of US data and policy events we mapped in the setup into September’s decision week. The market, in short, is buying wrappers and rate cuts. It is not, on this evidence, pricing Orion-100B.

Who provides a model nobody trained alone?

Distributed training also lands in a regulatory blind spot that neither crypto nor AI rules were written for. On the token side, the joint SEC and CFTC interpretation of March 2026 named a specific set of digital commodities, the majors like Bitcoin, Ethereum, Solana and XRP, and pointedly said nothing about DePIN or AI-compute tokens. That silence leaves TAO in a classification gray zone for US investors: not blessed as a commodity, not charged as a security, simply undefined. Bittensor’s own services layer, where custodians and exchanges touch the token, is where the SEC’s attention would fall first if it comes, not the protocol itself.

The AI side is stranger still. Europe’s AI Act, whose enforcement powers for general-purpose models took effect in August 2026, assigns obligations to the provider of a model. But who is the provider of Orion-100B? No single company trained it. A subnet coordinated a shifting set of anonymous peers, several of whom may sit outside the European Union entirely. The Act was drafted with corporate labs in mind, and a model that emerges from a permissionless swarm does not map cleanly onto its categories. That ambiguity is not a loophole to celebrate; it is an unresolved risk that could later attach to token holders, subnet operators, or the foundation, depending on how regulators decide to read a training run with no obvious author.

The bottom line

Bittensor’s training subnets are the strongest evidence the network is doing something a company structurally cannot: producing capable AI models that no single entity owns, trained across an open network rather than inside a walled cluster. Covenant-72B and Orion-100B are real, documented, and ahead of the off-network competition on scale. That is the bull case, and it is more grounded than the token’s chart suggests.

The same chapter is also the network’s harshest mirror. The team that best demonstrated the thesis walked out accusing the foundation of the exact centralization the project is supposed to abolish. The commodity-hardware framing oversells who can really take part. The economics rest on token emissions rather than customers. And Epoch AI’s arithmetic says the frontier stays out of reach for years. For readers deciding what to make of it, the questions to track are specific: does Macrocosmos keep scaling past 100 billion parameters, does Subnet 3 survive its founders’ departure, does the frontier gap start closing faster than 20 times a year, and do the US ETF decisions that currently drive the price ever give way to something the training work can support. Until then, the most interesting thing about Bittensor and the thing moving its token are, remarkably, not the same thing.

Frequently Asked Questions

What is Bittensor Subnet 3 (Templar) and Covenant-72B?

Subnet 3, called Templar, was a Bittensor subnet focused on decentralized model pretraining, operated by a team named Covenant AI. Between September 2025 and March 2026 it trained Covenant-72B, a model of roughly 72 billion parameters, on about 1.1 trillion tokens across a small set of independent peers rather than one data center. It scored 67.1 on the MMLU benchmark, slightly ahead of Meta’s LLaMA-2-70B, making it one of the most capable openly documented decentralized training runs to date.

Can you really train a large AI model without a data center?

Yes, up to a point. Techniques like DiLoCo, Nous Research’s DeMo, and Templar’s SparseLoCo cut the data exchanged between machines by two orders of magnitude or more, which makes training over ordinary internet links feasible where it once would have taken millennia. Bittensor subnets have trained models up to 100 billion parameters this way. The catch is that the largest runs still require high-end GPUs at each node, and Epoch AI estimates decentralized training remains about 1,000 times short of true frontier scale.

Is Bittensor (TAO) decentralized?

Partially, and it is contested. The training and inference work is genuinely distributed across independent operators. Governance is another matter: in April 2026 the Covenant AI team quit and called the network’s decentralization theatre, alleging cofounder Jacob Steeves retained unilateral control. Steeves has acknowledged the network is not yet decentralized like Bitcoin and describes current central control as a temporary choice to move fast, with a roadmap to full decentralization by the end of 2027.

How is Bittensor different from Prime Intellect and Nous Research?

They pursue the same goal, decentralized AI training, with different funding models. Bittensor pays contributors in its TAO and per-subnet alpha tokens, so the work is financed by token emissions and priced by speculation. Prime Intellect raised $130 million in venture equity at a roughly $1 billion valuation and has no liquid token, capturing value through its cap table and enterprise revenue instead. Nous Research, which developed much of the shared compression tooling, runs its Psyche network with coordination on Solana.

Is there a Bittensor (TAO) ETF in 2026?

Not a US spot ETF, as of September 2026. Grayscale’s Bittensor Trust operates as a private-placement product for accredited investors, and conversion filings from Grayscale and Bitwise remain pending with the SEC rather than approved. In Europe, a Safello staked-TAO exchange-traded product already lets holders earn the network’s staking yield, a feature a plain US spot ETF would find difficult to offer.

By Marcus Okafor, HOGE Wire markets desk.