The Allocation Machine: Why Bittensor Outallocates Committees, Corporations and Governments.

Part two of a three-part series on incentives, capital allocation, and the machine being built to fix both. Part one showed how the West pays for exactly the wrong things. This part follows the money: who allocates capital, how badly, and what a market that never adjourns does differently.

In 2012, the British government set out to build a railway. The full HS2 network, London to Birmingham to Manchester and Leeds, was costed at around £37.5 billion. Fourteen years later, everything north of Birmingham has been cancelled, £40 billion has already been spent, the surviving stub, one line, less than half the original network, is estimated at £88 to £103 billion, the trains will run slower than promised, and the first passengers are now expected somewhere between 2036 and 2039, with Euston station possibly unserved until 2043. Britain’s own Transport Secretary said it plainly in May 2026: “If this seems like an obscene increase in time and costs, it’s because it is.” Three times the money, for less than half the railway, a decade late.

In 2011, NASA quietly published a study asking its own cost model, NAFCOM, what it would have cost the agency to develop SpaceX’s Falcon 9 rocket the traditional way, through cost-plus contracts with the usual aerospace primes. The model’s answer: about $4 billion. Under a leaner “commercial” NASA approach, $1.7 billion. SpaceX’s actual spend, verified by NASA’s own accountants, was roughly $390 million for the Falcon 1 and Falcon 9 programmes combined. The government’s own maths, published by the government, showing the government’s procurement machinery inflating costs roughly tenfold against a founder spending largely his own money.

Same civilisation. Same decade. Same species of engineer. What differed was the allocation mechanism: who decides where the capital goes, how fast the feedback arrives, and who feels the pain when it’s wasted.

The first essay in this series toured the bounties for illness, idleness, arrival and theft, and ended on a question: what would an institution look like if the incentive could never drift from the outcome? This essay answers it. But first you have to see how capital allocation actually works in the institutions that control most of it, and why the most talented allocator alive, armed with a presidential mandate and a chainsaw, recently bounced off the problem.

Strip any government of its flags and its anthems and what remains is an allocation engine. It takes roughly 40 per cent of everything the country produces and posts bounties with it: bounties for medical treatment, for housing, for education, and, as part one showed, occasionally for not working, for shoplifting, and for rat farming. Three structural features make it the worst allocation engine in the economy, and none of them is fixable by electing better people. It spends other people’s money on other people, so the feedback loop between bad allocation and personal pain is severed. It defines value upstream, once, in prose, by committee, rather than letting the people consuming it discover value continuously; the committee writes “dead rats”, pays on “tails”, and can’t patch the contract until the next legislative session, by which time the rat farms have lobbyists. And it has no mechanism for defunding failure: a programme that misfires doesn’t shrink, it acquires beneficiaries, and beneficiaries vote. HS2 was re-baselined upward for a decade because at every decision point, the people deciding were spending someone else’s money on a project whose cancellation would embarrass them personally.

And notice what taxation actually does in this frame. The money the state gouges from payslips and profits isn’t destroyed; it’s transferred, from the one class of people who have just demonstrated, by earning it, that they can turn capital into value, to the institutions with the allocation record you have just read. Taxation’s deepest cost isn’t the blunted incentive on the way out of the worker’s payslip. It’s the reallocation on the way in: every pound moved from a proven allocator to a proven misallocator, at scale, annually, forever. An economy doing this at 40-plus per cent of GDP is running its capital through its least competent hands on purpose, and calling the result redistribution.

Corporations are better, but only by degree, because a corporation is still a central planner internally. Capital gets allocated by an annual budgeting process; political capture happens between divisions instead of between constituencies; and pay is tied to proxies, hours, headcount, revenue booked, that drift from value creation exactly as Goodhart’s law predicts. The average company reallocates meaningful capital roughly once a year, through a political process, using measures that corroded years ago. That is the incumbent technology. That is “best practice”.

Elon Musk, whatever else you think of him, is the most instructive capital allocator of the age, because he has now run both halves of the definitive experiment: what happens when you fix the allocation mechanism, and what happens when you attack misallocation without fixing it.

The first half is SpaceX, and the crucial detail is that its advantage was never just genius; it was contract structure. Under cost-plus contracting, the incumbent aerospace model, the contractor is reimbursed whatever it spends, plus a fee. Read that back slowly: revenue rises with cost. It is a bounty on expenditure, as purely perverse as anything in part one, and it produced the SLS rocket, which NASA’s own Inspector General priced at $4.1 billion per launch and called “unsustainable”, on a programme the GAO later described as “unaffordable”. SpaceX took the other structure: fixed price, milestones, its own capital at risk. Same industry, same physics, tenfold cost difference. And when NASA ran the two structures head to head on Commercial Crew in 2014, awarding Boeing $4.2 billion and SpaceX $2.6 billion for the same job, the fixed-price terms meant Boeing’s decade of Starliner failures cost Boeing over $2 billion, absorbed by its own shareholders, while SpaceX has been flying astronauts since 2020. The overruns didn’t vanish. They landed on the people empowered to prevent them. That is what a correctly wired incentive does.

The second half is DOGE, and it is the half your author finds genuinely poignant. In late 2024 Musk stood on stage and promised to cut “at least $2 trillion” from the US federal budget. He was handed a department, a presidential mandate, root access to the federal payment systems, a team of ruthless engineers, and the loudest megaphone on earth. Eighteen months later the scoreboard read as follows: claimed savings of $215 billion, about a tenth of the target; independent analyses of the “wall of receipts” finding triple-counting, contract ceilings booked as savings, and an $8 million contract recorded as $8 billion; outside estimates putting verified savings in the low tens of billions; the Partnership for Public Service estimating the exercise itself cost taxpayers around $135 billion in paid leave, rehiring and lost productivity; agencies scrambling within months to rehire for the jobs that turned out to matter; total federal spending up during the entire period; Musk gone by May 2025 after a spectacular falling-out with the President; and the department quietly dissolved into the bureaucracy it came to cut. “We have no plans to do kind of a closing DOGE report,” said the OMB director, an epitaph nobody will need to fact-check.

The point is not partisan. The most effective capital allocator alive attacked government misallocation with the only tools that exist inside the current paradigm: personnel, willpower and authority. And the machine absorbed him, the way it absorbs every reform-minded minister and every efficiency tsar, because he was pulling levers connected to people rather than to payouts. At SpaceX, Musk didn’t exhort aerospace to be efficient; the contract structure made efficiency the profit-maximising strategy, and efficiency duly appeared. At DOGE there was no contract structure, only a man, and men leave the room. Willpower doesn’t scale and doesn’t persist. Mechanisms do. If you want a one-sentence summary of this entire series, that is it.

And if eighteen months of DOGE feels like too small a sample, history already ran the same experiment for eight years, with the most celebrated business mind of an earlier generation. In November 1960, Robert McNamara became president of the Ford Motor Company, the first man outside the Ford family ever to hold the job. He was the star of the “Whiz Kids”, the statistical-control officers who had optimised America’s bomber fleets in the war and then rebuilt Ford with the same discipline: measure everything, model everything, let the numbers decide. Five weeks into the job, Kennedy asked him to run the Pentagon, and Washington cheered for exactly the reason people cheered DOGE: finally, a real businessman in the swamp. The early procurement reforms genuinely saved money. Then came Vietnam, and McNamara did what the best businessman of his age was trained to do: he found a metric. Enemy body count became the war’s revenue line, kill ratios its margins, bombing sorties its output figures, and every incentive in the chain of command promptly went full Hanoi. Units inflated counts, civilians became combatants in the paperwork, careers rose on numbers that measured nothing, and the dashboard in Washington glowed green while the war was being lost. The sociologist Daniel Yankelovich later gave the pathology McNamara’s name: the McNamara fallacy, making the measurable important instead of the important measurable. McNamara himself wrote the epitaph in his memoirs, decades on: “we were wrong, terribly wrong.”

What makes the pairing devastating rather than merely ironic is that McNamara at Ford and McNamara at the Pentagon were the same man, with the same brain, using the same methods. What changed was the feedback architecture around him. At Ford, his metrics were audited by reality every quarter: build a bad car and customers, dealers and the share price told him so, fast, in numbers he couldn’t fudge because they arrived from outside. At the Pentagon, his metrics were produced by the very hierarchy being measured by them, with no customer, no competitor and no price signal anywhere in the loop, and so they curdled into exactly the rat tails that every unpriced bounty in this series produces. Sixty years apart, the two most capable business operators of their eras walked into government carrying management, and government handed them back Goodhart’s law. The lesson was never that business people are better than politicians, though they often are. The lesson is that business people are only as good as the mechanism auditing them, and government is what a business becomes when you remove the audit. Which is why the interesting project was never sending better men into the ministry. It’s replacing the ministry’s operating system.

Which brings us, via a seventeen-year-old precedent, to the machine.

To a mechanism designer, Bitcoin is one thing: the first incentive scheme in history that has never been gamed at the payout layer. For seventeen years it has posted a bounty, currently around $10 billion a year, for exactly one commodity: valid proof-of-work on top of the honest chain. Not tails. Not proxies. The thing itself, verified by mathematics rather than by inspectors who can be fooled by a tailless rat. The result is the most secure computer network ever built, constructed by anonymous strangers, many of them actively hostile to each other, with no CEO, no budget committee, and no HR department. Bitcoin took the exact human energy that farms rats, greed, and converted it into security the way a turbine converts wind into electricity.

But proof-of-work is the only commodity Bitcoin knows how to buy, because hashes are the only commodity a blockchain can verify natively. The question that hung in the air for a decade was whether the trick generalises. Could you build a Bitcoin-style incentive machine that buys anything measurable? Machine intelligence, weather forecasts, protein folding, trading signals?

That is precisely, and explicitly, what Bittensor is.

Bittensor, launched by Jacob Steeves (“Const”) and collaborators with a deliberately Bitcoin-shaped monetary design (21 million hard cap on its token TAO, no premine, no VC allocation, four-yearly halvings, the first of which occurred in December 2025), is best understood not as a crypto project but as a factory for building Hanoi-proof bounties. The architecture has three layers, and each one attacks one of the three failure ingredients we began with.

Layer one: subnets, or programmable bounties. The network hosts around 128 subnets, each an independent competitive market for one digital commodity. One subnet buys AI inference. Another buys model training. Others buy GPU compute, weather forecasts, deepfake detection, protein folding, sports prediction, trading signals. Each subnet has a written incentive mechanism, code that defines what “good” means for that commodity and how it’s measured. Crucially, the mechanism is code, not legislation: when miners find a way to game it (and they always do; humans are relentless), the subnet owner patches the reward function in days, not parliamentary sessions. The metric and the target are forced back together every time they drift. Goodhart’s law, with a patch cycle.

Layer two: Yuma Consensus, or paying for rats instead of tails. Inside each subnet, miners produce the work and validators score it, submitting weight vectors that rank every miner’s output. The chain computes a stake-weighted consensus: each miner’s reward is set at the level at least half the stake agrees on, and any validator’s score sitting above consensus gets clipped, earning that validator nothing for the excess. Validators themselves earn by bonding to miners, and the bonds build only when their scores consistently track consensus. Score honestly and accurately, earn more; collude or freeload, earn less. Honest evaluation becomes the profit-maximising strategy, the exact inversion of the Hanoi rat bounty, where dishonest harvesting was. Emissions in each subnet split 41 per cent to miners, 41 per cent to validators and their stakers, 18 per cent to the subnet owner. Everyone in the loop eats what they measure.

Layer three: dTAO, or the market that fired the committee. This is the part capital allocators should study, because in February 2025 Bittensor did something almost no institution in history has done voluntarily: it identified its own central-planning bottleneck and shot it. Until then, the split of TAO emissions across subnets was decided by 64 root validators. A committee. And the committee behaved exactly like every committee in this series: the top five validators held more than half the voting power, favoured their own subnets, and cut side deals. Bittensor had, embarrassingly, reproduced a politburo at its own centre. The dTAO upgrade replaced it with a price system. Every subnet now has its own token, alpha, sitting in an automated market-maker pool against TAO. Anyone who thinks a subnet is producing value stakes TAO into its pool; anyone who thinks it’s junk pulls out. Each subnet’s share of network emissions is set continuously by the market price of its alpha. That’s the whole mechanism: capital allocation across 128 competing product lines, decided block by block by anyone on earth with skin in the game, with no allocation meeting, ever. Subnets the market believes in get bigger bounties to attract better miners; subnets that stop producing watch their emissions starve in real time. Failure defunds itself, and no beneficiary lobby can save a dying subnet, because there is no one to lobby.

Put the three layers together and you have the institutional inversion of everything in the first essay. The West’s systems define value once, in prose, by committee, and let the payout drift from the outcome for decades. Bittensor defines value in code, re-measures it every block, pays only on measured output, makes honest measurement itself the most profitable activity in the room, and reallocates the entire capital budget continuously by market price. Barry Silbert, who built Digital Currency Group and has seen every crypto cycle from the inside, calls Bittensor “the thing I’ve gotten most excited about since Bitcoin” and describes it as “the world wide web of intelligence”. DCG built a whole subsidiary, Yuma, to incubate subnets, and has committed over $105 million to the ecosystem. Grayscale filed in December 2025 to convert its Bittensor Trust into a spot ETF, and its analysts treat alpha tokens as de facto subnet equity: a liquid, real-time price on the output of every team in the network. Publicly listed treasury companies (xTAO on the TSX-V, TAO Synergies on Nasdaq) now exist purely to hold and stake TAO. The institutions are not confused about what this is.

Fair question. The honest answer: early, noisy, and yes, visibly.

The flagship is Chutes, subnet 64, a serverless AI inference market where permissionless miners with GPUs compete to serve open-source models. Chutes became the number one provider on OpenRouter, the main routing layer for AI inference, serving on the order of a hundred billion tokens a day at prices 10 to 50 per cent below centralised rivals. It claims $5.5 million in annualised revenue; independent analysts at Pine Analytics verified $1.3 to 2.4 million of it, and the gap itself tells you this ecosystem still marks its own homework, of which more below. Targon, subnet 4, sells confidential compute, claims $10.4 million ARR (self-reported, unaudited) and raised a $10.5 million Series A. Lium, a GPU rental subnet, did over a million dollars in its first five months. Zeus sells machine-learning weather forecasts to energy traders. Macrocosmos has folded over 160,000 proteins. Score’s computer-vision products are used by an actual football club.

The most interesting datapoint for the capital-efficiency argument is Ridges, subnet 62, a market for autonomous coding agents. Within about 45 days of launch, with less than a million dollars of cumulative incentives paid to miners, its top agent hit 80.3 per cent on SWE-Bench, a score Grayscale noted outperforms leading frontier-lab models on that benchmark. Attach every caveat you like to benchmarks, and you should attach several; the shape of the result survives them all. A bounty of under $1 million, posted to anonymous global talent through a well-designed incentive mechanism, bought performance that centralised labs spend billions approaching. That is the Hanoi rat bounty running in reverse: the same human ingenuity that bred rats for the tail money, pointed at the actual target, because the mechanism paid for dead rats and not tails.

And in March 2026, Templar, a subnet doing permissionless distributed training, produced Covenant-72B, the largest model ever trained over a decentralised network, notable enough that Jensen Huang acknowledged it on the All-In podcast. The machine is producing real commodities.

None of it is a finished cathedral, and it doesn’t claim to be. Verified external revenue across the network, somewhere between $3 and $15 million a year, is still a fraction of what the emission subsidy pays out; that gap is the revenue desert, and the halvings are the clock that will force every subnet to stand on real customer demand, exactly as Bitcoin’s block subsidy gave way to fees. The scoring mechanisms are still being patched under adversarial pressure, most recently with commit-reveal after researchers showed stake counting for more than output on some subnets. This is scaffolding around the most interesting economic construction site on earth, and its flaws are visible, priced, and patched in public, which is more than can be said for a single programme in part one of this essay. The NHS does not publish an adversarial analysis of its own incentive drift. When Bittensor’s incentives misfire, the miners find the exploit in days and the mechanism gets patched. When a government’s incentives misfire, the exploit gets a lobby group, and the patch takes a generation.

Now the investment thesis, which is really an organisational thesis. The competition, remember, is the incumbent technology from the top of this essay: capital reallocated once a year, by committee, through a political process, on corroded measures. Against that, a company built on Bittensor rails, running its product as a subnet or building atop one, gets a different machine entirely. Its cost of R&D is a bounty posted to the entire planet’s talent pool, paid only on measured output. Its capital allocation across product lines is repriced every block by a market rather than re-argued every year by a committee. Its measurement layer is adversarially stress-tested by profit-seeking miners who get paid to find the gaps, then patched, continuously. And its access to capital has no cap table gatekeeper: anyone, anywhere can stake into its alpha token the moment its output improves, and out the moment it doesn’t. As Steeves puts it, good supply is amplified and poor supply is naturally eliminated. Compare Ridges buying frontier-adjacent coding performance for under a million dollars against the billions a frontier lab spends, and “a million times better at exchanging capital for value” stops being a metaphor and starts being a ratio.

Set the two incentive stacks side by side and the contrast is almost comically stark. Britain, taken as a system, pays you not to work, taxes you progressively harder for each additional unit of work you insist on doing anyway, charges you again for employing someone else, and until this spring declined to charge anything at all for helping yourself to the contents of a shop. Bittensor pays you nothing for existing, nothing for enrolling, nothing for your credentials, your postcode or your sob story, and pays you instantly, every block, for exactly one thing: measured value delivered. There is no taper, because the second unit of value earns the same as the first. There is no cliff, because there is no threshold to fall off. There is no gouge on the marginal hour, because there is no taxman between output and reward. There is no wage floor either, and this cuts the other way to how it sounds: because no law forbids small contributions, there is no such thing as an unemployable person on Bittensor. A student in Lagos whose model serves five dollars of inference gets five dollars, no employer NI attached, no permission required, no first rung missing; and if their output improves tomorrow, their pay improves tomorrow. And there is no theft margin, because unearned extraction is the one strategy the consensus mechanism is explicitly built to starve. Every pathology in the first essay is a broken marginal incentive; Bittensor is nothing but marginal incentive, running clean. People respond to both systems with identical rationality. Only one of them compounds.

This does not mean every Bittensor company wins. Most subnets will die, exactly as most companies die, and the revenue desert has to bloom before the market cap makes sense on fundamentals. The claim is narrower and stronger: the discipline is the moat. The teams that learn to write incentive mechanisms that pay for rats and not tails, that survive adversarial miners probing every seam of their reward function, will have acquired the single scarcest skill in economic history. Every institution in the first essay, every health system, welfare state, tax code, subsidy scheme and Fortune 500 budget process, is a customer for that skill whether it knows it yet or not.

And follow the logic one step further, because it doesn’t stop at companies. Every programme catalogued in the first essay is a government-authored incentive mechanism, written in prose instead of code, unpatched for decades, with no measurement layer, no adversarial testing, and no market pricing its outputs. Governments aren’t bad at allocation because the people in them are stupid, and, as DOGE demonstrated, they can’t be fixed by sending in the smartest person alive with a chainsaw. They’re bad at it because they are running the worst incentive-mechanism technology in existence, and the technology is the thing that has to change. Which means the comparison Bittensor invites is not really with Nvidia or OpenAI. It is with the ministry. A tuned incentive mechanism squeezes more value from whatever you point it at, a business, a charity, a city, a country, because the underlying trick, pay precisely for the outcome and let the whole world compete to deliver it, is substrate-independent. Ridges bought frontier-grade coding for under a million dollars. Ask what the same architecture does pointed at, say, Britain’s asylum accommodation contracts, re-costed from £4.5 billion to £15.3 billion, a bounty currently structured to reward cost overruns, and you begin to see the actual size of the market this technology addresses. It is not the AI market. It is the allocation market, and the allocation market is every budget on earth.

For a hundred years we’ve had exactly two technologies for coordinating human effort at scale: the price system, which is brilliant but blunt, and the org chart, which is precise but corruptible and slow. Bitcoin hinted at a third: programmable incentives with ungameable payouts. Bittensor is the general-purpose version of that hint, an incentive machine you can point at any measurable commodity, wired to a capital market that never adjourns.

The French administration in Hanoi had rat catchers, rat farmers, and a bounty that paid for the wrong thing. The modern West has the most talented population in history, trillions in capital, and bounty structures that pay people to be ill, to arrive, to inflate the invoice, to hold the empty house, to refuse the pay rise. The talent was never the problem. The rat catchers are magnificent. They always were.

Change what the bounty pays for, and you change what the world produces. That is the entire trade.