TAO 2030: A $1tn+ Ecosystem

A Stillcore Series

“The relation of the Tao to all the world is like that of the great rivers and seas to the streams from the valleys.” Tao Te Ching, chapter 32

Introduction

The internet became the most valuable network in history, and no investor, at any price, has ever owned a piece of it. TCP/IP and HTTP were public standards. They went on to carry most of the world’s commerce and communication, but none of the wealth they created stayed with the protocols. It went to the companies built on top, and the best of those stayed private for years.

Bittensor is the open protocol of the intelligence era, built to move intelligence rather than information, and this time the protocol itself can be owned, because it has an asset. TAO is the network’s own currency, and every subnet on it, each one a market for a single kind of intelligence, trades against TAO from the day it launches. An investor in 1995 with perfect foresight still couldn’t have owned the protocols, or for several years the best companies built on them. On Bittensor both layers trade in the open.

In April we wrote that TAO would anchor an economy worth more than $1tn by 2030. Six months on, we’d make the same call. Anyone making a claim like that owes readers the arithmetic, and this paper is ours.

We published The Bittensor Thesis in March to explain how the network works and why we called it the most significant opportunity in crypto since Ethereum. This paper doesn’t repeat that case. It checks the case against what 2026 delivered and runs the result forward to 2030, and 2026 has only raised our conviction. Four things moved from argument to evidence this year. Intelligence began to trade like a commodity, with open models taking most of the volume on a large public AI gateway while prices kept falling. Buyers chose open models, first on price and then, after June, on control. Bittensor began to earn, with outside customers now paying its subnets $50mm to $65mm a year by our estimates, a significant multiple of what they paid a year earlier. And the network assembled every layer of a frontier lab, then used it to turn an open base model into one that, by its builder’s account, matches frontier quality from 2025 at a fraction of the parameters and price. TAO itself trades about where it did a year ago, while in our view the network’s fundamentals and momentum have grown by an order of magnitude. We don’t think much of that is in the price yet.

Commodity markets tend to end the same way. Producers compete their margins down to a normal return, and the money that lasts goes to whoever owns the scarce inputs and whoever runs the market where the good is graded and sold. If intelligence goes the same way, and 2026 suggests it will, we think the Bittensor of 2030 is a market of hundreds (likely thousands) of subnets selling compute and intelligence to each other and to outside buyers, many (perhaps most) of them AI agents paying by the token. At the valuations we work through below, the network's own issuance would pay for a fleet of more than a million GPUs that nobody owns, and every transaction on it would run through TAO.

The paper is in three parts. Part One sets out what 2026 showed. Part Two is our broader vision for the ecosystem in 2030, with subnets competing against companies many times their size, intelligence bought on tap the way water and power are, a network that gets better at getting better, and the value running back into TAO. Part Three does the arithmetic, with four roads to $1tn and the markers we'll track each year to hold ourselves to it.

PART ONE · WHAT 2026 SHOWED

1. Intelligence Became a Commodity

The first thing 2026 settled was the direction of the price of intelligence. In September Epoch AI estimated that the cost of reaching a given level of performance has fallen by about 13x a year since 2023, faster than computing or DNA sequencing ever got cheaper. A score on a benchmark of PhD-level science questions that cost about $0.30 a question in January 2025 cost about $0.0004 less than 18 months later, a drop Epoch puts at more than 700x. Epoch’s earlier work put the decline anywhere from 9x to 900x a year, depending on the task.

Open models do most of the work. Once anyone can download a model and run it, a closed model can’t charge much more for the same job than it costs to run the open one, and distillation keeps the gap short, because smaller models are trained on the outputs of bigger ones and a closed model’s API sells exactly those outputs. Every answer it sells helps train a cheaper competitor, and in September one leading U.S. lab alleged that seven Chinese labs had run more than 180mm exchanges through its models to train their own, about 150mm of them in one case. Asked on CNBC later that month whether that kind of copying was fair game, Jensen Huang said people strip Nvidia’s products down to their raw parts every day. “That’s called competition,” he said. Meanwhile the hardware keeps getting better. We believe that by 2030 AGI-level intelligence will be effectively free at the margin, and running on the phone in your pocket.

The volume has already moved. Vercel, whose AI Gateway routes tens of trillions of tokens a month for production applications, reported in September that open-weight models handled 56% of its tokens in August, the first month they carried a majority, up from 7% in December. Because they cost so little per token, they took only 14% of estimated spending, and closed models took the other 86% on 44% of the tokens. That’s how commodity markets usually turn. The cheap product takes the volume first, and the line between cheap and expensive moves up every time the cheap product gets better. Epoch now puts the best open models 3-4 months behind the best closed ones, down from about a year in late 2024. The average price Vercel’s customers paid per token fell 23% in August alone, and when Vercel took apart an earlier monthly drop, it found that at a fixed mix of models the price would have been essentially flat. The whole decline came from what customers chose to route.

Buyers who move to the cheaper product as soon as it’s good enough are treating the good as interchangeable. Intelligence is a commodity now, and 2026 is the year the data made that plain. What enterprises, businesses, and people are buying is intelligence per dollar, the most capability they can get for each unit they spend, and they’re becoming indifferent to which lab built the model or which generation it belongs to. The published figures probably understate the shift, too, since open weights run wherever their owners like and most of that inference never shows up in anyone’s index. Our partner Jason Calacanis has made the point on All-In more than once, calling them “dark tokens.”

None of that shrinks the market. In 1865 William Stanley Jevons noticed that more efficient steam engines had increased Britain’s coal consumption, because cheap power found uses expensive power never had. It’s known as the Jevons paradox, and intelligence may be the strongest case of it yet. A task gets automated the moment automating it is cheap enough. In the United States, the share of adults who used AI on at least six days of the previous week rose from 8% in March to 19% in August, and more and more of the buying is done by software agents, which run all day and never reach a point where they’ve had enough.

The closest precedent is the printing press. Before it, copying knowledge was a craft, and a small, trained class decided what was reproduced and who could read it. Gutenberg didn’t write better books. He collapsed the cost of copying them, and by 1500 printers across Europe had turned out some 30,000 known editions. Information stopped being scarce, and within a century the world had been remade around it. Intelligence today sits roughly where information sat in 1450, and it’s more powerful than information ever was, because it acts. It turns an idea into software, into a business, into almost anything. We see TAO as the printing press of intelligence.

A unit price falling toward zero while volume climbs is how commodities behave, and the history of commodities is hard on producers. When buyers can’t tell one producer’s output from another’s, producers compete on cost until they earn about their cost of capital, and the profits that last go to the owners of scarce inputs and to the exchange where the good trades. Chicago grain is the textbook case. Until the 1850s, wheat there was sold by the sack and every lot had to be inspected by its buyer. In 1856 the Chicago Board of Trade began sorting wheat into standard grades, so grain of the same grade could be pooled in elevators and sold without anyone looking at it, and in 1865 the Board added standardized futures. Farmers went on earning ordinary returns. The Board became the place where the world’s grain was priced, and as part of CME Group its grain futures are still the global benchmark. Trading gathers where the liquidity already is, which is why a commodity rarely supports more than one or two big exchanges for long.

The AI industry already talks about intelligence as a fungible good. Jensen Huang calls data centers AI factories and measures them in tokens per watt and tokens per dollar, and as open models take share, what a closed lab can charge per token drifts toward the cost of the compute behind it. Intelligence still lacks what wheat got in 1856. A GPU hour is easy to measure. The quality of an answer or a forecast isn’t, and until strangers can check it without trusting each other, intelligence can’t trade the way graded wheat does. We think whoever solves grading for intelligence gets the position the Chicago Board of Trade got in grain. Bittensor was built around that problem, and we know of no other network that was.

Most of the capital in AI is betting the other way. The frontier labs are valued on the assumption that intelligence stays expensive enough to carry fat margins on the models themselves, while a market that trades intelligence wants the opposite, since every price cut adds volume. By 2030 we think model-makers will look a lot like farmers, earning an ordinary return on a commodity crop, and the lasting money will go to the owners of compute and to the market where intelligence changes hands.

2. The Market Chose Open

If intelligence is going to trade like a commodity, someone has to run the market, and 2026 made a strong case that it will be an open one. Two years ago that wasn’t the obvious bet. A handful of closed labs had the best models, and most investors assumed the companies that built the models would own the market for them too. The internet’s history suggests otherwise.

The internet’s first winners mostly didn’t last. AOL spent the 1990s trying to own the way people got online and was worth more than $150bn at the top of the dot-com market, and Verizon bought it for $4.4bn in 2015. Sun Microsystems, which sold the servers the early web ran on and advertised itself with the line “the network is the computer,” reached a value of roughly $200bn in 2000, nine years before Oracle agreed to buy it for $7.4bn. The open standards underneath both companies are still running. We think AI is going through the same sequence, with the closed frontier labs in the seats AOL and Sun once held. They control access to the best models and much of the hardware that serves them, and the two most valuable are reported to be seeking valuations of more than $3tn between them. Underneath them, an open layer of downloadable models, and of the networks that train and serve them, is growing fast.

The buyers moved first, and they moved on price. Airbnb’s chief executive, Brian Chesky, said last October that the company relies heavily on Alibaba’s open Qwen models for its customer service agent because they’re fast and cheap, and that it rarely puts OpenAI’s newest models into production. Uber blew through its annual AI budget in the first three months of 2026, then held its AI spending flat from the spring while its use of AI agents kept climbing, partly by moving work to open models that cost 2 to 20 times less. Pinterest’s chief executive told investors in August that open models the company post-trains itself cost less than 8% as much per transaction as comparable closed models. AT&T cut the cost of some advanced AI work by as much as 56%, for a 2% drop in quality, after moving it to open models. For most of what a business needs, an open model is now good enough and much cheaper.

Then June made the case on control. On June 12, a leading U.S. lab switched off its two newest models for every customer after an export-control directive from the U.S. Commerce Department. The directive was lifted on June 30 and access came back on July 1. We take no view here on whether the directive was right. But for almost three weeks, every person, business, and country relying on those models learned that a supplier could be switched off by a decision they had no part in. Every open- weight model kept running, because once weights are published there’s nothing to recall. Governments outside the United States and China had worried about that kind of dependence for years, and we believe few boards or governments that lived through June will stake a company or a country on a single closed model again. Huang put the risk in engineering terms in July, when he warned that a world running on one model has “one single point of attack, one single source of failure,” and in September Nvidia agreed to buy Hugging Face, the platform where most of the world’s open models are published, in a deal valued at $12.9bn. Read that as a statement about where the company selling hardware to everyone expects the volume to go.

Open weights have largely settled the question of access, though they haven’t changed who produces the models. An open-weight model is still built and paid for by one company, and it gets released because releasing it suits that company’s strategy. Meta, DeepSeek, Moonshot, and Nvidia all publish open weights for their own reasons, and any of them can stop. A good share of the Chinese wave is subsidized by investors who will eventually want a return, and it’s capped by export controls on the chips the next generation needs. When Moonshot’s Kimi K3 came out in July and placed third at launch on a widely followed independent index of model capability, behind only two closed models, it revived talk in Washington of sanctioning Chinese labs and of making U.S. companies liable for running their models. A model released by a single company can be withdrawn by that company, and its home government can restrict it. The compute, the data, the research, and the rights to the output are as concentrated as they were before open weights took off.

Bittensor was built to open up that part. Anyone can bring compute or code to a subnet, and the network pays in tokens for whatever works, so the people who did the work end up owning a piece of what they built. Buyers are already settling the debate over whether AI should be open. The fight that decides what the market for intelligence looks like in 2030 is over who owns its production, and Bittensor is built so that anyone can take part in it. That’s why the printing press comparison holds. The first press ended the scarcity of knowledge. This one ends the scarcity of intelligence, and anyone can hold a piece of it.

3. The Network Started Earning

Bittensor was designed to be that open market, and 2026 is the year it began to earn. The Bittensor Thesis covers the mechanics in detail, so we’ll keep them short.

People in the Bittensor community like to call Bitcoin the first subnet. It’s a joke, but an accurate one. Bitcoin sets a single task, which is to find a number whose hash falls below a difficulty target. Anyone can compete and check an answer instantly, and whoever finds one first gets paid in new bitcoin. With no company behind it, that loop built the largest computing network in history. Bittensor keeps the loop and lets anyone write the task. Each job runs as its own market, called a subnet. The subnet’s owner decides what the work is and how it gets scored, miners anywhere in the world compete to do it, validators score what comes back, and the chain pays the winners continuously in the subnet’s own token.

Scoring is where the difficulty lives. A hash either meets the target or it doesn’t, but the quality of a model’s answer, or which of two forecasts came closer, is a judgment call, and judgment calls can be gamed. Bittensor’s answer is Yuma Consensus. Validators score miners on a sliding scale, each validator’s scores count in proportion to the stake behind it, scores far from the consensus get clipped, and validators who keep disagreeing with the consensus earn less. That’s the grading system intelligence was missing, and it’s why we believe Bittensor will become the exchange where intelligence is priced. The Chicago Board of Trade could pool wheat because an inspector’s grade let strangers trade it unseen. Yuma Consensus does the same job for an answer or a forecast. Many validators grade each piece of work, and because their stake and their earnings ride on agreeing with one another, the grade can be relied on without trusting any one of them. The test is whether buyers who never see the miners will pay for what they produce, and on 23 subnets they already do.

A copy of Bittensor would struggle for the same reason commodity exchanges consolidate. Thousands of miners and validators have spent years learning to beat each other on this network, and none of that experience moves to a fork. The chain runs its own exchange with deep liquidity, and its subnets buy from each other, so each new one makes the others more useful. It also pays out one of the largest steady funding streams in open-source AI, and it launched fairly in 2021, with no premine and no allocation to venture investors, which no later competitor can copy. The code is open source and anyone can fork it, but a fork would start with almost none of this.

Capital moves through the same kind of market. Every subnet has a pool, built into the chain, that pairs its token with TAO, and the only way to buy a subnet’s token on the chain is to put TAO into that pool. The protocol issues new TAO every day and splits it among subnets according to the market price of their tokens. A subnet that attracts capital gets more of the new TAO and can pay for better work. One that bleeds capital gets starved, and if the bleeding goes on long enough, it loses its slot to a new team. Think of it as a venture fund that rebalances all day long, with the market doing the partners’ job. For a founder the difference shows up in the workforce. A company out of Y Combinator raises a seed round and spends most of it on hiring. A subnet founder writes an incentive, and once its emissions are switched on, engineers and GPUs from around the world compete to do the work.

For most of the network’s life that machinery produced research more than revenue. That changed over the past 12 months. By our estimates, outside customers now pay Bittensor’s subnets $50mm to $65mm a year, a significant multiple of what the whole network earned 12 months ago. Estimates vary widely, because most of this revenue is paid off-chain and reported by the teams themselves. Lium, Targon, and Chutes lead in revenue, PwC France is among the customers, and Dropbox is piloting one subnet’s product. At the Exploit Summit in Montreal in September, Const, Bittensor’s co-founder, said 23 of the network’s 128 subnets now have paying customers. Chutes, which serves the leading open-weight models to developers at a fraction of centralized prices, had about 800,000 registered users by the figures he showed, and by July two subnets reported earning enough from customers to cover everything the chain paid their miners, the first test of whether a subnet can stand on its own.

Even at the top of our range, that covers only about a sixth of the value of the TAO the network issues each year. The protocol is meant to subsidize supply before demand arrives, and the subsidy shrinks on a fixed schedule. Revenue has no schedule, and its growth is the part of the case we’re betting on.

The rules changed this year as well. In May the protocol stopped minting free tokens to a subnet’s owner at registration and began tying ownership of a mature subnet to locked stake, so a large enough locked position can take it over. In June it began splitting new TAO by the market price of each subnet’s token, with a penalty for subnets that hold back their miners’ rewards. At Exploit, Const described the next step, which he expects in early 2027. The network would measure how much money each subnet draws from outside, from buyers on other chains and from purchases of its services, and make that one of the inputs that decide where new TAO goes. If it ships as described, the subsidy will start to follow the customers, which is the shift the revenue case for 2030 depends on.

4. The Frontier Lab Took Shape

A frontier AI lab needs somewhere to keep and move data, compute to train and serve models, a way to train them, and products that put them to use. Bittensor now has subnets working at every one of those stages.

On the infrastructure side, Hippius (SN75) runs S3-compatible storage, while Beam (SN105) moves large datasets across a decentralized bandwidth network. Compute and inference is the most developed layer. Chutes (SN64) serves open models, Targon (SN4) runs confidential compute, Lium (SN51) rents out GPUs, Engy (SN53) serves open models and says it attaches a cryptographic proof that each answer was computed correctly, Actual (SN95) pools everyday machines into a single inference computer, and SayGM (SN28), which our partner Mark Jeffrey calls Bittensor’s own OpenRouter, ties the layer together for the developers who use it. Training is thinner but moving, with IOTA (SN9) pre- training on single GPUs connected over the open internet, Teutonic (SN3) running a continuous contest on a 110bn-parameter model, and Affine (SN120) improving open models after pre-training through head-to-head contests. At the application end, Score (SN44) turns ordinary cameras into computer vision for businesses and Minos (SN107) does genomic variant calling aimed at clinical use, with others working on software security (RedTeam, SN61), video compression (Vidaio, SN85), forecasting (Synth, SN50), and alignment data (Aurelius, SN37).

The subnets also trade with each other. Minos keeps its genomic data on Hippius, and applications buy inference from Chutes rather than paying a hyperscaler. A layer that runs on a cheaper layer below it can charge less for its own service without giving up margin, and because each layer uses the output of the ones beneath it, the savings stack up as you move up. A closed lab has to own and fund every one of these layers itself, all on one balance sheet. On Bittensor each layer is funded by its own market and sells to every other subnet on the same terms.

How far this goes by 2030 depends mostly on training. Selling compute and serving other people’s open models is already a big business, and it produces most of the network’s revenue today, but Bittensor only becomes a frontier lab if it can produce competitive models of its own. In 2026 the first ones arrived. In the first quarter the network produced Covenant-72B, a 72bn- parameter model trained by at least 70 independent contributors, with benchmark results in the range of Meta’s Llama 2 70B, a model of similar size trained in 2023 inside one company’s data centers. The team behind it left the network in April, and its subnet now runs Teutonic. IOTA’s 100bn-parameter run started in June and finished late in the third quarter, training across five data centers on single A100 GPUs. A complete replica of the model, 16 separately rented A100s, cost about $20 an hour, and in June the run was reaching about 65% of the training speed of an equivalent data center. Macrocosmos, which runs IOTA, says it ran as a proof of viability rather than to a finished model, and in September it showed a 16bn-parameter run on a mix of gaming and data-center cards. Teutonic’s second model, a 110bn-parameter run that is still training, is what Const calls the largest decentralized model yet.

The result that matters most came at Exploit. Const stepped back from running the Opentensor Foundation to build Affine, a subnet that sets miners against one another to improve open models after pre-training, because he wanted to prove the network could train frontier models and not only serve them. In September he showed its first model. It started from an open model released by a Chinese lab, and everything after that was done on Bittensor, on infrastructure from Hippius, Lium, Targon, and Engy, by what he called “an unorganized group of individuals across the globe” who compete for the same reward. Const says that on his team’s own scoring it would sit on the frontier of a widely followed independent index for a model of its size, 36bn parameters with 3bn active, and that it matches a leading closed model from 2025 across a combined set of benchmarks. Those are the builder’s own figures, and we’d like to see them replicated. If they hold up, a network whose contributors are strangers and adversaries took an open model to 2025 frontier quality, at a fraction of the parameters and price, on infrastructure the network also runs.

The same talk announced the piece that would turn the stack into an economy. The subnets already sell to one another, but each keeps its own books, the way a cloud provider sells prepaid credits. Gamma tokens, which Const expects to release in 2027, would put those credits on the chain under one standard. A buyer would burn a subnet’s own token and receive credits worth a fixed number of dollars, set by a price oracle, that can only be spent on that subnet’s service, so an hour on Lium or inference on Chutes becomes a token that can be bought, held, transferred, and bridged to other chains. A subnet would be able to direct part of its emissions to buying other subnets’ credits, the way a startup spends most of its raise on compute, and anyone holding crypto on another chain, including an AI agent, could rent a GPU or buy inference without a website or an account. In effect it would give the network a stable unit of account for compute, which a volatile token can’t be. If it ships as described, the frontier lab gets a common currency, and every subnet’s customers include every other subnet.

We’ll be plain about the gap. When we wrote The Bittensor Thesis, we put decentralized training’s effective throughput at about 300x below the largest centralized clusters, and we estimated that the compute available to decentralized training was growing about 20x a year, about four times as fast as centralized infrastructure. If those rates held, the gap would shrink by a factor of about four a year and close in a little over four years, somewhere around 2030. We don’t count on that in our base case. It’s a hard engineering problem that could fail, so we treat it as an option on top. If it works, every model becomes a commodity, and Bittensor sits at the source of the world’s supply of intelligence.

There’s an older argument for why it might. In 1997 Eric Raymond described two ways to build software, the cathedral, designed by a small group and released when finished, and the bazaar, built in public by whoever wanted to contribute, and Linux came out of the bazaar. In September OpenAI said that about 10,000 agents running an unreleased model, working in groups and exchanging millions of messages, had produced a claimed proof on the Navier-Stokes problem in about 88 hours. It has been checked by OpenAI’s own system but not yet independently verified, and mathematicians have only started on it. Whatever they conclude, the result came from a large group of cooperating agents, and OpenAI ran that group inside a single company. A network of specialized subnets that call on one another and pay their contributors is the same idea, run in the open.

PART TWO · WHAT IT BUILDS BY 2030

5. The Machine the Incentive Builds

By 2030, at the valuations in this paper, Bittensor’s own issuance would pay for more than a million of today’s GPUs, and nobody would own the fleet. Bitcoin shows how a machine like that gets built. It ran on a handful of personal computers in 2009. By 2025 its miners were computing more than 1,000 exahashes a second and drawing about as much electricity as a mid- sized country, and no company had planned or paid for any of it. The mechanism is simple enough. Miners keep adding machines until the last one added earns just enough to cover its costs, so the size of the network follows the dollar value of the reward. Hashing doesn’t produce intelligence, and Bitcoin’s machine exists only to secure a ledger, but it shows better than anything else how much hardware a permanent incentive can pull together without an owner.

Bittensor’s machine grows by the same arithmetic, with two advantages. Its miners are paid in new tokens, and on subnets with customers the revenue goes into buying those same tokens, so the reward grows with the business. And the network can use almost any hardware the moment it exists, so data-center GPUs rented through Lium work alongside the gaming cards IOTA trains on and the home machines Actual pools together, while a data center has to be financed and built before it produces anything.

That makes it possible to estimate how big the machine gets by 2030 from the value of issuance alone. By then the second halving, which the current issuance rate puts at around the end of 2029, will have cut new issuance to about 1,800 TAO a day, on a supply of roughly 16mm. Figure 6 shows what a year of that issuance would be worth at three values for TAO, and how many H100-class GPUs half of it would rent at about $2 an hour.

At $1tn, issuance alone would rent more than a million H100-class GPUs for a year, before counting a dollar of customer revenue, and in the revenue case later in this paper, half of $25bn would rent roughly another 700,000. For scale, Meta said in January 2024 that it expected to end that year with the equivalent of about 600,000 H100s. These figures also assume today’s hardware. If price-performance keeps doubling every 2 to 3 years, the same dollars will buy several times as much compute by 2030. The biggest labs may be larger still by then, but at $1tn Bittensor’s machine would rank among the largest in the world, with no owner and no balance sheet behind it, and it would grow automatically every time TAO rose.

The incentive works on people, too. At past highs, top miners have earned tens of thousands of dollars a day, paid only for what they produced. An engineer in Lagos or Manila can be as good as anyone in San Francisco and still be shut out of the jobs and capital that would pay for it. Bittensor doesn’t ask where a miner lives or went to school, and it may turn into one of the largest talent markets ever opened. A country with stranded power or idle chips can use it to export intelligence, the way some countries already export hashrate.

6. From Pyramid to Subnet

By 2030 we believe subnets will be the cheapest producers in category after category, competing directly with companies many times their size. The reason goes back to 1937, when Ronald Coase set out to explain why firms exist at all, given that markets are supposed to be good at allocating resources. His answer was transaction costs. Finding someone to do a job, agreeing on terms, checking the work, and paying for it all cost money, and when those costs run high, it’s cheaper to hire people and manage them than to contract with strangers.

Jack Dorsey and Roelof Botha came at the same problem from the other end in March, in an essay called “From Hierarchy to Intelligence.” Since the Roman army, they argue, organizations have used layers of managers mainly to pass information up and down and to keep everyone aligned, and every extra layer slows things down. AI can now do most of that routing. The company they describe puts intelligence at the center and people at the edge, where the work gets done, and steers by revenue on the view that “money is the most honest signal in the world.” Dorsey has started rebuilding Block around the idea, weeks after cutting about 40% of its staff.

A subnet goes a step further and drops the company altogether, which takes out most of the costs Coase was describing. Registration is open, so finding workers costs nothing. The incentive mechanism is the contract. Validators check the work and the chain pays for it, while miners, more and more of them AI, compete at the edge and get paid for results. Nobody in the middle passes anything along, and resources follow the money, the same signal Dorsey and Botha picked. The main advantage of size used to be the ability to coordinate thousands of people toward one goal. A protocol can now do that for thousands of strangers at almost no cost, which lets a subnet compete with a company many times its size in any category where output can be measured. The costs that survive are the ones grading can’t reach. Where the work is hard to score, as with judgment, taste, or trust, the firm keeps its advantage, which is why the categories that move first are the ones with a clear score.

Score shows how that works in practice. Its platform handles routine computer vision with a base model and sends the hard cases, like unusual lighting or rare objects, to the subnet’s miners, who compete to extend the model’s coverage. Each customer deployment turns up new hard cases, and each one the miners resolve improves the product for the next customer, while a conventional competitor improves only as fast as it can hire engineers and buy labeled data.

Most companies will never become subnets, just as most companies aren’t software businesses even though all of them run on software. The largest will buy from subnets the way they buy from the cloud today, and the first to see it will launch subnets of their own.

7. Intelligence on Tap

By 2030 we think intelligence will be bought the way water and power are, metered and paid for as it’s used. Today the biggest buyers of compute still buy it the way utilities used to buy power plants, years in advance and in enormous amounts. OpenAI’s reported agreement to buy about $300bn of computing capacity from Oracle over five years is the best-known example. Contracts like that made sense while capacity was scarce and reserving it was the only way to be sure of having it. Nobody buys water that way. You open a tap when you need it and pay for what you use, at a price the system sets.

Jensen Huang reached for the same comparison at the All-In Summit in September. Closed models work terrifically, he said, and he thinks of them as “kind of like bottled water.” Water itself is free, he went on, and you use the right water in the right places, the same as with electricity or any other commodity, so the world needs both. We’d take his analogy further than he did. Bottled water is a good business and probably always will be (on Vercel’s gateway, closed models still took 86% of August’s spending on 44% of the tokens), but the bigger business is the utility that meters water to every building. Bittensor is building that utility for intelligence, and the meter is where the lasting money is.

Electricity made this change within living memory. For most of the twentieth century, American power came from vertically integrated utilities that sold it at regulated prices. Starting in the 1990s, wholesale markets began pricing power in real time, and across much of the country electricity now clears every few minutes at the cost of the last plant needed to meet demand. Compute today looks like power before that change, sold on long contracts by a few large suppliers at prices set by negotiation.

Bittensor already works like a wholesale market for intelligence. Miners buy and run their own machines, so the network carries almost no fixed costs. Competition among them sets the price of each unit of work, and because the network doesn’t have to finance data centers, it can put hardware to work that a centralized cloud can’t use economically, including older data-center GPUs and idle machines at home.

A wholesale market also needs buyers who can use it, and this is where 2030 should look most different from today. Most customers still sign up the way they would with any cloud provider. By 2030 we think more and more of them will be AI agents buying intelligence on their own account, and an agent needs a market it can use without a contract or a sales call, one that quotes a price and takes payment in software. Gamma credits are built for that buyer. If they ship as described, an agent holding crypto on any connected chain could buy an hour of compute or a batch of inference from a subnet and pay for it in the same transaction.

None of this makes Bittensor a threat to the frontier labs or the data-center builders. Capacity bought years ahead needs somewhere to sell the hours it doesn’t use, and frontier models need cheaper models underneath them for the routine work customers won’t pay frontier prices for. Many in the industry now expect enterprises to split their work that way and keep frontier models for a narrow band of high-value problems, and the Vercel figures show the split is already under way. If it holds, some of the capacity built for the frontier will sit idle and look for buyers, and spot markets are where idle capacity goes. If decentralized training reaches the frontier around 2030, as the training results so far suggest it could, the split disappears and every model becomes something the tap can deliver.

8. The Mosaic

By 2030 we think the network will be improving itself. Recursive self-improvement (RSI) is a system that gets better at getting better. The curve bends when the thing doing the improving is itself improving, and on Bittensor the loop has started to close. The subnets are each other’s customers, so every gain one of them makes lowers the cost of the next. Const described the pipeline at Exploit. A pre-training subnet builds a base model, post-training subnets improve it, and inference subnets serve it. One post-training subnet already lets a model do the grading, so the network’s own output judges the network’s own work. He called it self-referential recursive improvement, and it runs in the open.

The same loop works at the level of a single miner. An AI agent that mines a subnet and earns more can buy more compute and get better at earning. Our bet is that by 2030 many more agents than people will be mining, running, and launching subnets, and that starting a subnet will cost about what a website costs. One day there may be as many subnets as there are websites.

That changes what superintelligence might look like. We think it’s more likely to arrive as a mosaic of specialized intelligences working together than as a single model in a single lab. OpenAI’s 10,000 agents were a small version of that mosaic, run inside one company. Bittensor is the same idea with the walls removed, open to anyone and paid by a protocol, and it’s already assembling itself subnet by subnet.

By then, AGI-level models, humanoid robots, fusion, and much else may be arriving at once. In a world that abundant, what stays scarce is whatever has a fixed supply. TAO is capped at 21mm, and it’s built to turn intelligence into a market.

9. Where the Value Lands

By 2030 the new TAO coming onto the market each day is scheduled to be a quarter of what it was in 2025, while most of the ways money enters the network run through TAO. Follow the money on the network and it ends up in TAO, because on the chain every subnet token trades only against TAO. Anyone who buys a subnet’s token there, whether a customer, a trader, a fund, or the subnet’s own team, has to put TAO into the pool that pairs the two. A subnet with revenue has its own reason to buy its token, since the token is what it pays miners with and a more valuable token attracts better work, so teams use revenue to pay their costs and to buy their own tokens on the market, in some cases burning them, and each purchase leaves TAO in the pool. Registering a subnet costs TAO, and staking into one commits TAO to its pool. Revenue and speculation on every subnet end up as demand for TAO, and so does every new subnet. TAO is the index of the entire economy Bittensor is building, although it gives holders no legal claim on the subnets themselves.

Gamma credits would add a route that runs through usage rather than speculation. As described at Exploit, an outside buyer would bridge dollars in, buy TAO, buy the subnet’s token with it, and burn that token for credits in one transaction, leaving TAO in the pool and taking the burned tokens out of supply. When a subnet pays for other subnets’ credits out of its own emissions, it sells some of its own token to do it, so the net effect on any one pool depends on the mix. The rules haven’t been published, so none of this is in our numbers.

Supply is easier to pin down. Daily issuance was 7,200 TAO until the first halving on December 15, 2025, and runs at about 3,600 now. At the current rate the second halving comes around the end of 2029 and cuts issuance to about 1,800 a day, a 75% drop in four years, with about 16.4mm TAO issued by the end of 2030 against a hard cap of 21mm. And most of it is committed on chain. Of the roughly 11.6mm TAO issued so far, chain data show about 5.5mm staked on the root network and about 2mm in subnet pools, where it backs the subnet tokens, which leaves about 4.1mm, some 35%, unstaked. Bitcoin went from about $13 to more than $1,000 in 2013, the year after its first halving.

Then there’s demand. Ethereum had a demand (ICO) shock in 2016 and 2017, when a wave of projects raised money in ETH and anyone who wanted in had to buy ETH first. On Bittensor the same pull is part of the design rather than a fundraising fashion, which is why our partner Mark Jeffrey calls TAO “the coin of the realm.” You need it to take part. Every purchase of a subnet token on the chain starts with TAO, and so does every new subnet registration, with 256 slots on the roadmap against 128 today.

The on-ramps for outside money are being built now. In May, Sunrise, Wormhole Labs’ gateway for bringing assets onto Solana, launched a canonical version of TAO that trades on Jupiter and Meteora. ForeverMoney, a Bittensor-focused bridge built on Chainlink’s CCIP, put a one-for-one version of TAO on Base in August and has since added Robinhood Chain, and it pairs subnet tokens with TAO for use in DeFi. OKX listed TAO for spot trading in June, and Kraken has listed six subnet tokens against the dollar since July. Grayscale’s application to turn its Bittensor trust into a U.S. exchange-traded product is still pending. Most of these routes are still thin, but each one lowers the cost of getting in.

New buyers also change who holds TAO, and that changes how it trades. We think a lot of today’s float sits with holders who treat TAO as a leveraged bet on Bitcoin and sell it hardest when Bitcoin falls, which is why TAO still moves with the crypto cycle more than with anything the network does. Staking, pool depth, subnet registrations, and exchange-traded products all move TAO from traders to holders who tend to keep it, and we expect TAO’s sensitivity to Bitcoin to fade by 2030 as that happens.

PART THREE · THE ARITHMETIC

10. The Road to $1tn

TAO traded at about $304 on September 30, around 40% of its March 2024 high, and the subnet tokens priced in it were worth about $1.4bn more. Allowing for some overlap between the two, that’s an economy of about $5bn, so $1tn is 200 times the current figure. Against the assets it gets compared with, TAO is small, at about 1/500th of Bitcoin and about 1/1,000th of the two most valuable private AI labs at the valuations they’re reported to be seeking.

There’s no standard way to value a network like this, so we’ve worked the number four ways. Two of them value the whole economy and two value TAO by itself, which is why the title has a plus sign. Read each one as a description of what 2030 would have to look like for $1tn to make sense.

Revenue is the method we’d weight most. At 40 times revenue, a $1tn economy needs about $25bn a year of it in 2030. From $50mm to $65mm today, that means roughly quadrupling revenue every year until then. At 20 times revenue the target doubles to $50bn. Amazon Web Services took 12 years to reach $25bn of revenue, in a cloud market much smaller than we expect the market for intelligence to be by 2030. It’s the number we’d tell anyone to watch.

The Bitcoin comparison asks more of the market than of the network. If Bitcoin reached $250,000 a coin, TAO would need to be worth about a fifth of it to reach $1tn, which is roughly where Ethereum sits today, and at $500,000, a tenth would do. Ethereum’s value has ranged from about 4% of Bitcoin’s to nearly 80% over the last decade. TAO is worth about 1/500th.

Adding up the subnets shows the shape the economy would need. The largest subnet token is worth about $140mm. For the subnets to add up to $1tn, the network would need something like 10 category leaders, in areas such as inference, compute, training, and storage, worth around $40bn each, 40 more worth around $10bn, and a long tail of about 200 worth around $1bn. Each leader would be worth more than all but a handful of crypto assets today, and today’s biggest subnet would need to grow nearly 300 times. It’s a power law, like every venture portfolio and the web itself.

The last calculation runs the other way, from the valuation to what it would buy. At $1tn for TAO alone, the 1,800 TAO issued each day after the second halving would be worth about $110mm, or roughly $40bn a year, a little less than the $43.9bn Meta spent on research and development in 2024. That’s the budget a $1tn network would hand out every day to whoever produced the best intelligence, and it’s what would pay for the fleet of GPUs described earlier.

Four years is a short time for a gain this size. The nearest precedent is Ethereum, worth under $2bn at its 2016 peak and about $570bn at its November 2021 high, roughly 300 times in five and a half years.

We read TAO’s place in its life on three clocks, and they point to three different years. On the money clock it’s 2013. Bitcoin’s first halving came on November 28, 2012, which made 2013 its first full year after one, and TAO’s came in December 2025. Bitcoin began 2013 worth under $1bn and ended it near $9bn, and TAO’s $3.5bn sits inside that range. On the innovation clock it’s 2016, the year Ethereum’s first real applications appeared and a year before the rush of new projects that defined the next cycle. On the capability clock it’s 1997, the year Amazon went public and sold $148mm of goods, up from $15.7mm the year before, with almost the entire build-out of the web still ahead of it.

TAO is still priced at the speed of Bitcoin, and built at the speed of AI. AI is already in a supercycle, and we believe crypto is entering one. We think productivity takes center stage in 2027, and that by 2028 AI will be driving a full-scale revolution in how work gets done. Abundance on that scale is deflationary. Elon Musk has argued that when AI and robots grow the output of goods and services faster than the money supply, the problem becomes deflation rather than inflation.

11. What Has to Be True

Every method above depends on the same few things going right. The table is what we’ll track each year between now and 2030. Revenue is an estimate, since most of it is still paid off-chain, and the rest can be checked from public data.

The biggest risk is that revenue grows too slowly. Even at $1tn, a year of issuance in 2030 would be worth about $40bn against $25bn of revenue, so the subsidy would still be bigger than the business it pays for. That’s survivable as long as revenue keeps gaining on it, and the third halving, around 2033, cuts the subsidy again. If revenue lags, miners earn less, and the fleet of GPUs ends up smaller than the figures above imply.

Jacob Steeves, the co-founder known as Const, still sets much of the chain’s rulebook, including its incentive parameters and upgrade path, and that concentration is key-person risk. This year the protocol stopped minting free tokens to subnet owners at registration and tied ownership of mature subnets to locked stake. It also merged the first pieces of on-chain governance, though upgrades still pass through a small multisig. In June Const published a plan to hand the rulebook to the network within about 18 months, by around the end of 2027, and at Exploit he said he had taken a heavy-handed role over the past six months and wants the network ready to run without him. We intend to hold the network to that plan.

If decentralized training never closes the gap, the closed labs keep the highest-value work and Bittensor serves the rest. That would lower the ceiling on the revenue and sum-of-the-parts cases without ruling them out, since compute and inference already produce most of the network’s revenue.

The SEC decides when U.S. exchange-traded products come to market, and regulators could decide that some subnet tokens are securities, which would limit how they trade in the United States. In September the SEC’s staff said a token buyback isn’t the kind of managerial effort that makes a token a security when the network is functional and has no central party. Whether a subnet with a known team fits that description is a question for counsel, and the answer will shape how subnet tokens trade here. Kimi K3’s release in July also revived talk in Washington of restricting Chinese open models, so the politics of open weights are far from settled.

TAO still trades with Bitcoin’s four-year cycle. If the pattern holds, Bitcoin’s next halving in 2028 puts the cycle’s likely peak around 2029, and the year after every past peak has been rough on crypto assets. We expect at least one severe drawdown before 2030, and readers should plan on one too.

Dynamic TAO launched in 2025, and its pools and emissions logic have been rebuilt more than once since. Software this new carries risks that a protocol with a decade behind it doesn’t, and because every subnet runs on the same pool mechanism, a failure there would hit all of them at once.

The Way

The epigraph compares the Tao to the great rivers and seas that the streams from the valleys run into. We picked it partly for the pun and partly because it describes how Bittensor is wired. Every subnet trades against one asset, so value created anywhere on the network, whether by a subnet buying and burning its own tokens or by a better model, ends up running downhill into the same pool of TAO.

Underneath the numbers, this paper is about who will own intelligence, the most important economic question of our lifetimes. For most of AI’s short history the answer has been a handful of companies, and the public argument has been over how much of their earnings they should be made to share. Bittensor gives the only answer we know of in which anyone can own a share. The people and machines that produce intelligence on the network own part of it through the tokens they earn, and anyone can take a share of the whole by holding TAO.

Our funds concentrate on subnets because on Bittensor the winners aren’t known in advance, and the market picks them every day. None of us has held a view with more conviction in our careers, though we may be early. We’ve written down what has to be true, and we intend to report against it each year until 2030.

Our earlier research, including The Bittensor Thesis and our State of Subnets report, is at stillcorecapital.com. Accredited investors who want to go deeper can reach us there.

This paper reflects Stillcore Capital’s opinions as of September 30, 2026. It is not investment advice, and it is not an offer to sell or a solicitation of an offer to buy any security. Interests in Stillcore’s funds are offered only to verified accredited investors under Rule 506(c). Stillcore, its principals, and its funds hold TAO and subnet tokens, including some discussed in this paper. Please read the Important Disclosures that follow.

Important Disclosures

Opinion, not advice. This paper reflects the opinions of Stillcore Capital Management, LLC (“Stillcore”) as of September 30, 2026. It is provided for informational purposes only, does not take into account any reader’s objectives, financial situation, or needs, and is not investment, legal, tax, or accounting advice, or a recommendation to buy, sell, or hold any asset. Readers should consult their own advisors and do their own research before making any investment decision. Stillcore’s views may change without notice, and Stillcore undertakes no obligation to update this paper.

Holdings and conflicts of interest. Stillcore, its affiliates and principals, and the funds it manages hold TAO and positions in subnet tokens, including tokens of subnets named in this paper. They may buy, sell, or otherwise change those positions at any time, including after publication and in ways that may be inconsistent with the views expressed here, and they stand to benefit if the prices of those assets rise. Stillcore and its principals may also have business or other relationships with subnet teams, validators, and other participants in the Bittensor ecosystem, and its principals may discuss these assets in other forums, including podcasts and social media.

No offer or solicitation. This paper is not an offer to sell, or a solicitation of an offer to buy, any security, including interests in Stillcore Capital Fund I, LP or Stillcore Capital Fund II, LP (the “Funds”). Interests in the Funds are offered pursuant to Rule 506(c) of Regulation D under the Securities Act of 1933, as amended, solely to accredited investors, as defined in Rule 501(a), whose accredited investor status has been verified as Rule 506(c) requires. Any offer will be made only through the Funds’ confidential offering documents, which describe the Funds’ terms, fees, expenses, risks, and conflicts of interest, and which govern in the event of any inconsistency with this paper. The interests have not been registered under the Securities Act or any state securities laws, are subject to restrictions on transfer and withdrawal, and are illiquid. The Funds are not registered as investment companies under the Investment Company Act of 1940. Neither the Securities and Exchange Commission nor any state securities regulator has approved or disapproved of the interests or passed on the accuracy or adequacy of this paper. This paper is not directed at any person in any jurisdiction where its distribution would be unlawful.

Forward-looking statements and illustrative scenarios. This paper contains forward-looking statements, including estimates, projections, and scenarios for 2030, which rest on assumptions Stillcore believes are reasonable but which may not be realized. The $1tn figure and each of the four valuation methods are illustrative scenarios built on the assumptions stated in the paper. They are not forecasts, price targets, or predictions of the future value of TAO, any subnet token, or any Fund, and actual outcomes may differ materially. Hypothetical figures, such as GPU counts and revenue paths, are calculations from those assumptions and do not reflect actual results.

Third-party information and estimates. Market data, company statements, media reports, and other third-party information come from sources Stillcore believes to be reliable, but Stillcore has not independently verified them and makes no representation as to their accuracy or completeness. Revenue figures are Stillcore’s own estimates, unaudited and subject to revision. Benchmark results and other claims attributed to subnet teams or other projects are those parties’ own claims unless otherwise noted. Except for Stillcore’s own partners, references to companies, projects, and individuals are for illustration only and do not imply endorsement or affiliation, or that any of them has reviewed this paper.

Risks. Digital assets, including TAO and subnet tokens, are highly speculative and volatile and may lose all of their value. They are subject to market, liquidity, technology, protocol, custody, cybersecurity, regulatory, and tax risks, including the risk that regulators treat some tokens as securities or restrict their trading. Past performance of TAO, Bitcoin, Ethereum, or any other asset discussed in this paper is not indicative of future results. An investment in the Funds is speculative and involves a high degree of risk, including the loss of an investor’s entire investment, and is suitable only for investors who can bear that loss and the illiquidity of the interests.

Sources

• CoinGecko, Market data for TAO, Bitcoin and Ethereum, including TAO’s all-time high of $757.60 on March 7, 2024, accessed September 30, 2026
• companiesmarketcap.com, Market capitalization of Nvidia, accessed September 30, 2026
• TechCrunch, Report that OpenAI is in talks to raise at a valuation of about $1.4tn, September 29, 2026
• Fortune, Report on a leading private AI lab’s planned IPO at a valuation above $2tn, September 29, 2026
• Stillcore Capital, “2026 State of TAO” (January 2026), “The Bittensor Thesis” (March 2026), “The Show Goes On” (April 2026) and “State of Subnets” (July 2026)
• Epoch AI, “The plunging price of thought,” on the price of a given level of performance falling about 13x a year since 2023, September 22, 2026
• Epoch AI, “LLM inference prices have fallen rapidly but unequally across tasks” (9x to 900x a year), March 12, 2025
• Epoch AI, Open-weight models trailing the frontier by about four months in 2026, May 29, 2026
• Epoch AI, Open models matching closed ones with about a one-year lag, November 4, 2024
• Vercel, AI Gateway Production Index, September 2026 (open-weight share of tokens and of estimated spending; August price per token), September 17, 2026
• Vercel, AI Gateway Production Index, August 2026 (price decline driven by model mix), August 2026
• The Hacker News, Report of a U.S. lab’s allegation that seven China-based labs distilled its models, September 11, 2026
• CNBC, Squawk Box, Jensen Huang on distillation and competition, September 28, 2026
• Quartz, Jensen Huang on AI distillation, “That’s called competition”, September 28, 2026
• Epoch AI and Ipsos, Near-daily AI use by U.S. adults rising from 8% to 19%, September 14, 2026
• All-In podcast, Jason Calacanis on untracked “dark” open-source tokens, September 25, 2026
• William Stanley Jevons, “The Coal Question”, 1865
• Incunabula Short Title Catalogue (British Library), Census of editions printed before 1501
• Aashish Velkar, London School of Economics, Economic History Working Paper 145/10, on wheat grading at the Chicago Board of Trade from 1856, 2010
• Joseph Santos, EH.Net Encyclopedia, “A History of Futures Trading in the United States” (Chicago Board of Trade futures, 1865)
• Morgan Stanley TMT Conference, Jensen Huang on AI factories, tokens per watt and tokens per dollar, March 2026
• Fortune, “Chesky says OpenAI tools not ready for ChatGPT tie-up with Airbnb app,” on Airbnb’s reliance on Alibaba’s Qwen, October 21, 2025
• The Pragmatic Engineer, “The Pulse: tech companies move to open AI models,” on Uber, Pinterest and AT&T, September 10, 2026
• Axios, “Exclusive: Uber cuts AI costs even as usage jumps”, August 27, 2026
• Benzinga, “This Day In Market History: Time Warner Spins Off AOL,” on AOL’s peak market value and Verizon’s $4.4bn purchase in 2015, May 28, 2019
• Oracle and Sun Microsystems, “Oracle to Buy Sun” for about $7.4bn, April 20, 2009
• companiesmarketcap.com, Sun Microsystems’ market value history, including its 2000 peak, accessed September 30, 2026
• National Law Review, A U.S. lab’s suspension of its two newest models after a Commerce Department export-control directive, June 17, 2026
• Vercel, AI Gateway Production Index, July 2026 (controls lifted June 30, access resumed July 1), July 2026
• TechSpot, “Nvidia’s Jensen Huang joins X, defends Chinese AI,” on the “single point of failure” remark, Kimi K3’s ranking at launch, Epoch AI’s estimate of the open-closed gap and the reaction in Washington, July 24, 2026
• NVIDIA, Form 8-K on its agreement to acquire Hugging Face, September 3, 2026
• TechCrunch, “Nvidia confirms it will buy Hugging Face for $12.9 billion”, September 3, 2026
• Jacob Steeves (Const), “The State and Future of Bittensor,” opening talk at the Exploit Summit, Montreal, September 28, 2026
• Jacob Steeves (Const), Thread on the Exploit announcements, including subnet usage figures and gamma tokens, September 29, 2026
• Stillcore Capital, Estimates of network revenue, September 2026
• SubConnect, “The SubConnect Bittensor Revenue Index: Volume I,” on the subnets with the most customer revenue, Score’s customer PwC France and Leadpoet’s Dropbox pilot, August 26, 2026
• Stillcore Capital, “State of Subnets,” on the first two subnets to earn enough to offset their miner emissions, July 2026
• Targon and Intel, “Decentralized Compute on Untrusted Hardware Using Intel TDX and Encrypted CVMs”, 2026
• Subtensor, Pull requests 2642 (“No initial owner alpha”), 2599 (“Convictions (stake locks)”) and 2781 (“Price-based emissions with miner-burn scaling (+ locked transfer flags)”), May and June 2026
• Lidin et al., “Covenant-72B,” arXiv 2603.08163, March 2026
• Macrocosmos, “Orion-100B: distributed pretraining”, June 1, 2026
• Macrocosmos, Post on Orion-100B as a viability proof, September 25, 2026
• Macrocosmos, Post on Orion-16B across mixed consumer and data-center GPUs, September 23, 2026
• Jacob Steeves (Const), Announcement of the Teutonic-II 110B training run, August 24, 2026
• Minos, Subnet repository, on storing data with Hippius, accessed September 30, 2026
• Engy, Verified inference for open models, accessed September 30, 2026
• Jacob Steeves (Const), Exploit Summit opening talk, naming SayGM (SN28) among the network’s primary infrastructure providers, September 28, 2026
• Bittensor, Post describing gamma tokens as acquired by burning a subnet’s alpha, September 28, 2026
• Exploit Summit, Keith Singery and Jacob Steeves, “Gamma Deep Dive”, September 29, 2026
• Eric S. Raymond, “The Cathedral and the Bazaar”, 1997
• Quanta Magazine, “AI Has Solved One of Math’s $1 Million Millennium Prize Problems”, September 8, 2026
• CoinDesk, Bitcoin’s seven-day average hashrate reaching 1 zettahash a second, September 2, 2025
• Cambridge Centre for Alternative Finance, Bitcoin’s electricity use compared with that of countries, August 31, 2023
• TechSpot, Mark Zuckerberg on compute of about 600,000 H100-equivalents by the end of 2024, January 20, 2024
• Actual, Local inference across users’ own machines, accessed September 30, 2026
• Ronald Coase, “The Nature of the Firm,” Economica, 1937
• Jack Dorsey and Roelof Botha, “From Hierarchy to Intelligence,” Block, March 31, 2026
• Fortune, “Jack Dorsey and Roelof Botha think AI can make middle management obsolete”, April 2, 2026
• Tom’s Hardware, OpenAI’s reported $300bn, five-year computing agreement with Oracle, citing the Wall Street Journal, September 11, 2025
• All-In Summit, Jensen Huang on closed models as “kind of like bottled water”, September 2026
• Jacob Steeves (Const), Exploit Summit opening talk, on the training pipeline across subnets and a subnet whose incentive mechanism is run by a model, September 28, 2026
• TAOStats and the Bittensor chain, Issuance, root stake, subnet pools, subnet token values and the halving schedule, accessed September 30, 2026
• CoinGecko, Bittensor subnets category, market values of subnet tokens, accessed September 30, 2026
• Mark Jeffrey, “State of TAO, A Bittensor Report (Excerpt),” on X, on TAO as the coin of the realm, February 19, 2026
• The Block, “Wormhole bridges ‘canonical’ version of Bittensor’s TAO token to Solana,” on Wormhole Labs’ Sunrise platform, May 5, 2026
• KuCoin, “Bittensor TAO goes live on Robinhood Chain via Chainlink CCIP,” on ForeverMoney’s bridge and its earlier expansion to Base, 2026
• OKX, “OKX to list TAO/USDT (Bittensor) for spot trading”, June 30, 2026
• Kraken, “Bittensor subnet tokens are coming to Kraken,” and Kraken’s subnet token markets against the dollar, 2026
• Grayscale Bittensor Trust, Registration statement on Form S-1 to list on NYSE Arca, December 30, 2025
• Amazon.com, Inc., Fourth-quarter and full-year 2018 results (AWS net sales of $25,655mm), January 31, 2019
• Meta Platforms, Fourth-quarter and full-year 2024 results (research and development of $43,873mm), January 29, 2025
• CoinMarketCap, Ethereum and Bitcoin market value history, including Bitcoin’s value at the start and end of 2013, accessed September 30, 2026
• Amazon.com, Inc., Annual report on Form 10-K for 1997 (net sales of $147.8mm, against $15.7mm in 1996)
• The Economist, Interview with Elon Musk, on AI, robots and deflation, July 23, 2026
• Crypto Briefing, Const’s roadmap to decentralize Bittensor within 18 months, June 22, 2026
• Subtensor, Pull request 2198 (“Governance”), on-chain governance, June 9, 2026
• Unchained, “SEC staff narrow token buyback guidance to networks with no central party”, September 29, 2026