Part II · the thesis

The Last Allocator

When demand learns to allocate itself, the allocator is the last of its kind.

Part I ended on a question: if companies become temporary and nearly free, what is left worth owning? Here is one answer. It is about the people who decide where money goes, and why AI is about to make them the last of their kind.

SERIES 2026 DENOMINATION ONE ALLOCATION THIS NOTE IS ISSUED BY THE HAND THAT HOLDS IT
§

The allocator is a guess about demand

Every venture fund, every allocator of capital, is doing one thing underneath the theatre: guessing what people will want before they say it, and pricing that guess early. That is the whole job. Partners are paid to be a proxy for a signal they cannot see directly. The signal is demand. The proxy is taste, network, pattern memory, nerve.

A proxy is a lossy thing. It samples a few founders, reads a few decks, trusts a few instincts, and stands in for the wants of millions it never meets. When the real signal is expensive to read, the proxy earns its fee. The allocator exists because demand, until now, could not speak for itself at scale, in time, with money attached.

Take that constraint away and the reason for the middle layer goes with it. This part is about what happens the moment demand can read itself, price itself, and pay for what it wants. The allocator does not get disrupted by a better allocator. It gets dissolved by the thing it was always only imitating.

THREE ROLES COLLAPSING INTO ONE LOOP: ALLOCATE, PRODUCE, CONSUME

I·IV

How the middle disappears

Four moves, in order. Each one is small. Together they retire a profession. Read them as a sequence, because the order is the argument: demand becomes legible, then it allocates, then it issues, then it governs the thing that builds for it.

I

Demand becomes legibleAggregated behaviour, read continuously by machines, becomes a cleaner picture of what people want than any partner's intuition of it.

What it replacesThe pitch meeting, the survey, the focus group. All of them ask a small sample to describe a want. Behaviour at scale shows the want instead, revealed rather than reported, and it never stops updating.
Why it winsA proxy loses to the source the instant the source is readable. Once a model can watch a million real choices a second, the partner reading twelve decks a week is not competing on the same axis. Legibility is the whole game, and machines are better at it.
Made concreteThe clearest proof already runs: ad markets. Google's open exchange runs on OpenRTB real-time bidding, Meta prices its feed through its own internal auction, and both pair those with recommender models that match intent at millisecond granularity, no venture partner in the loop. TikTok's ranker predicts content demand faster than any studio exec's taste. Extend that toward agent commerce (x402 settling agent purchases on-chain, the proposed ERC-8004 giving agents attributable identity) and an LLM run continuously over the transaction stream could read revealed preference cheaply, where a Sequoia partner samples a handful of pitches a week.
The strongest counterBehavioural data only captures demand for options that already exist. The judgment a VC is paid for is funding markets with no trace yet: there was no click data for the iPhone, Uber, or Airbnb before they shipped. Aggregated behaviour is a rearview mirror, blind to latent demand, and it degrades under Goodhart once it is optimised against.Genuine concession at the zero-to-one frontier. But most capital flows into incremental reallocation inside existing categories, where legible aggregate demand clearly beats intuition, so the claim holds for the bulk of demand even if it fails on the genuinely new.
II

Users become the allocatorsWhen demand can price itself, the user is the capital router. Not user feedback feeding a fund. The user, in aggregate, doing the fund's job.

The shiftAllocation stops being a decision a few people make about the many, and becomes something the many do by existing legibly. The signal and the cheque merge. Wanting a thing, at sufficient scale and legibility, becomes the act of funding it.
Why it holdsA demand-native allocator has no vintage year, no fund cycle, no thesis to defend past its sell-by date. It reprices the moment behaviour moves. It is not smarter than the VC on any single call. It is faster, cheaper, and closer to the source on all of them at once.
Made concreteWith x402 (shipped May 2025), any API can answer a request with HTTP 402 and get paid in USDC on the spot, and the proposed ERC-8004 sketches a portable identity and reputation registry so a service could bill a counterparty it has never met. Still an early standard, not a proven system, but put those together and every agent that pays per call is casting a priced vote: the revenue stream becomes the allocation signal, so a service a million agents actually pay for gets funded by its own metered demand rather than by an a16z partner deciding which of ten pitches deserves a Series A. Cloudflare's pay-per-crawl points the same way, letting demand price itself request by request.
The strongest counterPriced demand only allocates revenue to things that already exist and already have users. It says nothing about who fronts the large, irreversible, pre-demand capital (chip fabs, foundation-model training runs, years of R&D) for things with zero paying agents on day one, and that upfront risk capital is exactly what venture funds supply.Honest concession: legible demand disperses the allocation of ongoing revenue, but it does not solve who bears the upfront risk before any demand is legible. The allocation of scarce risk capital can stay just as concentrated even as billing decentralises.
III

Demand mints its own instrumentsOnce you can allocate, you can also issue. Users create the financial instruments that reward the exact product they want built, and hold them themselves.

The moveA crowd that wants a thing stops petitioning someone to fund it and instead issues the claim that pays for it: a token, a bond, a bounty on an outcome, redeemable when the thing exists. Demand becomes its own treasury. The reward for building flows to the people who wanted it built.
Why it mattersThis is the step that closes the door behind the allocator. When the people who want the product also own the upside of its creation, there is no seat left for a party whose only contribution was believing in the want early. The want and the capital were never two things. They just used to be held by two people.
Made concreteThe clearest working version is the Advance Market Commitment: in 2009 a group of donors put up $1.5B through Gavi, guaranteeing a capped price for pneumococcal vaccine doses, which got GSK and Pfizer to sign 10-year supply commitments and build capacity. Operation Warp Speed ran the same play in 2020, paying for COVID doses before they existed. The crypto-native forms are Optimism's RetroPGF, which pays builders retroactively for public goods already delivered, and Gitcoin quadratic funding, which matches a pool by how many distinct people fund a thing. Extended to agents this is still speculative: an escrowed outcome bounty settled over x402 could release payment automatically once an oracle confirms the work was done.
The strongest counterMinting an instrument does not repeal the free-rider problem. A diffuse crowd rationally under-contributes, hoping others pay, which is why Kickstarter chronically underfunds ambitious goods and token presales fill with speculators who want to flip the instrument rather than use the thing it funds.Honest concession: every version that actually scaled (the AMC, Warp Speed, RetroPGF's matching fund) is anchored by a few large, credibly committed payers plus a subsidy. So "a crowd mints its own instrument" is in practice "a few whales and a matching pool do," and the crowd framing oversells who carries it.
IV

The autonomous DAO is the containerThe vehicle where legible demand both funds and directs, and a machine builds. Behaviour guides development. No capital class in the loop.

What lives hereA standing pool of demand, read by AI, that issues its own instruments and points them at outcomes. The DAO is not a company with a token bolted on. It is the loop itself given a wallet: wanting, paying, and building, held in one autonomous body instead of passed between three.
The end stateProduct development steered autonomously by demand, with AI as the builder that reads the pool and ships against it. Nobody green-lights. The roadmap is just what the aggregate is already paying to exist. The allocator, the product manager, and the founder's instinct all fold into one reading function.
Made concreteA DAO holds an on-chain treasury and posts legible demand as bounties, and AI agents build against it, paid on delivery over rails like x402 with identity carried in a registry like the proposed ERC-8004. The pieces exist in isolation: Nouns DAO funds builders directly from its ETH treasury by vote, and Optimism's RetroPGF has paid out tens of millions in OP tokens for demonstrated impact rather than an upfront investor bet. The catch, and why it is still speculative: RetroPGF only rewards work after the fact, so a forward-looking version that commissions agents in advance does not exist yet, and the allocation decision does not vanish, it moves from a VC to token-holder votes.
The strongest counter"Legible demand" is doing all the work and it does not actually exist. DAO governance is token-weighted voting with low turnout, whale capture, and delegate politics, which is a proxy for whoever accumulated the most tokens, not a clean signal of what users want or will pay for.Honest concession: the treasury still has to be capitalised and its stablecoins issued by someone (Circle, a whale, an early backer), so the model relocates the capital class to whoever holds governance weight rather than removing it from the loop. This is exactly the New Rentier future in Part III.
×

Watch the middle dissolve

One control: how legibly demand can read itself. Slide it up and watch the allocator fade, the direct loop brighten, and the instrument mint itself. This is the whole thesis in one moving part.

The collapse
Drag the signal. At the left, demand is illegible and the allocator earns its fee. At the right, demand reads and pays for itself, and the middle has nothing left to do.
ILLEGIBLE LEGIBLE
Who routes capitalVC partners
Allocation cadenceVintage years
What gets fundedSomeone's guess

The picture has a direction built into it, and the direction is the point. Move left to right and you are not upgrading the allocator, you are deleting the reason it existed. The fee the middle used to earn was rent on illegibility. Pay off the illegibility and the rent goes to zero. What is left is a crowd that wants, an instrument that pays, and a machine that builds, closed into a loop that never needed a partner's signature.

∞

The limit case: perfect information

The slider above stops where behaviour ends. The ladder does not. A click is already downstream of the thing a market actually runs on: the wanting itself, upstream of language, upstream of the finger. The last rung of legibility is the interface that reads the want at its source, and it is being built right now, as a medical device that does not yet know what else it is.

Run it to the end. A brain-computer interface over a shared neural stream does not sample demand, it measures it. A product's price stops being discovered by watching who pays and starts being read directly: how hard a million nervous systems lean toward the thing, in the second they lean. Pricing becomes physiology. The survey, the ad auction, the recommender, every tool for guessing a want from its residue, collapses into a sensor that touches the want itself.

And the stream is labour, so the stream is paid. This is move III wearing skin. If your signal prices the market, the market owes you a wage for the reading: a data dividend, income for being legible. The instrument demand mints attaches at the source, and every connected mind holds a claim on the value its own wanting creates. You are not the product here. You are the oracle, and oracles bill.

Now take the limit and ask the only interesting question. Everyone connected, every want shared, every valuation common knowledge: how do markets react? Badly, and then by becoming something else. Most of a market's organs eat information asymmetry. The bid-ask spread is rent on ignorance, and it starves when nobody is ignorant. Arbitrage is a bet that you know something the other side does not, and there is nothing left to know. Advertising is the art of injecting noise into a demand signal, and a channel that reads the want directly no longer accepts the injection. Insurance thins, because risk-pooling needs both sides blind to whose house burns. Speculation, research, the analyst, the roadshow: all of them are ways of paying to know first, and there is no first. What survives is the part that was never about knowing: matching, building, delivering. The market stops being a discovery machine, because nothing is left to discover, and becomes pure allocation at the speed of the read.

The spread
Drag the reading deeper. The shallow end is behaviour: clicks, purchases, residue. The deep end is the want itself, read at the source and paid a wage. Watch what the market loses the more it knows.
BEHAVIOUR THE WANT ITSELF
The spreadRent on ignorance
Who pricesThe auction guesses
The wageNothing. You are the product

Made concreteThe reading layer is already in humans. Neuralink's N1 implant has had paralysed patients driving cursors by intention since early 2024. Synchron reaches the motor cortex through a blood vessel, no open skull. Precision Neuroscience's thin-film array cleared the FDA in 2025 for short-term use. All of it is assistive medicine today, and every bit of it is a demand sensor the moment its stream is shared. The payment half has precedents too: California floated a data dividend in 2019, and Posner and Weyl's Radical Markets put the argument plainly: data is labour, and the signal you already emit deserves a wage.
The strongest counterPerfect information is self-cancelling, and economics has known it for decades. Grossman and Stiglitz showed that a price system revealing everything pays no one to learn anything, so the discovery a market runs on stops being produced. Hirshleifer showed that some information destroys value the moment it is shared: insurance exists because neither side knows whose house burns. And the deepest objection is neural: wants are not coins in the skull waiting to be counted. Much of wanting is constructed in the act of choosing, so a signal read before the choice may be noise dressed as demand.Concession: the limit is never reached, and that is the design. The equilibrium is a standing gradient, a market that pays exactly enough data wage to keep enough minds legible, with a spread that never touches zero, only the cost of reading one more mind. Perfect information is the asymptote. Markets do not die on the way to it, they migrate: every fee that used to reward knowing first moves to reading deeper.
!

What breaks it

A thesis you cannot break is not a thesis, it is a wish. Four honest places this fails, and the tell to watch for each. Hold on to them, because Part III turns each one into a future.

Whoever shapes behaviour allocatesIf the allocator is now the reading of behaviour, then the party who shapes behaviour becomes the new allocator, quietly, without a fund or a name. The manipulator inherits the seat the VC left. The tell: does the loop reward what people want, or what people were nudged to do just before the reader looked.
Revealed preference is a trapAggregated behaviour is not the same as what people would choose if they were rested, informed, and not being farmed for engagement. A perfectly legible demand signal can point straight at things nobody is glad they wanted. Legibility measures the pull, not the worth of being pulled.
Nobody asks for the discontinuousSome products have no demand until they exist. Pure demand-allocation funds the faster horse forever and never the car, because the want for the car cannot be read off behaviour that has never seen one. The loop is brilliant at optimisation and blind to invention. Taste, the thing the allocator sold, may be exactly what has no demand signal.
Autonomous hides a new small classAn autonomous DAO centralises on whoever writes the reading function and the issuance rules. Autonomy in the loop can be a costume worn by a handful of people who wrote the code that reads the crowd. The last allocator may not disappear. It may just move into the repository and stop filing as a fund.

The allocator does not fully die. The one who priced yesterday's demand dissolves into everyone. The one who funds what nobody can see yet is the last of its kind, and that is who Part III follows.