Recapitalism

Centralization is coming and cannot be stopped. That is not the interesting question. The interesting question is what ends up at the center — because centers are not interchangeable, and the one that wins on cost is the one that holds a standard and a judgment, gives away everything it builds at the moment it builds it, and earns from flow rather than ownership.

Solving for: what actually happens if nothing intervenes.

Five things are already in motion, and each one causes the next.

1. The cost of standing up an organization falls toward zero. In any category where the work is legible — where you can describe what good output looks like — orchestrated agents beat human operations on cost. Not marginally.
2. Capital formation becomes the bottleneck. If launching compresses from years to weeks, funding cannot stay on a schedule of discrete rounds negotiated over months. The step-function private raise is the wrong instrument for the new run rate. Early-stage funding has to become continuous and liquid or it becomes the thing everyone is waiting on.
3. Categories consolidate. Whoever orchestrates best can acquire a competitor, strip its cost structure, and run it at a margin the competitor could never match. Do that repeatedly and the category ends up under one umbrella. Winning a market replaces winning a customer. Capital concentrates behind whoever has proven they can consolidate — call it the next form of private equity, and note that its returns rest on attention, which is the one thing machine systems are already superhuman at holding.
4. Compute becomes the scarce input. Capability scales with compute, so compute is the binding constraint on capability, which is why data-center commitments are being announced at a scale approaching a trillion dollars. Whoever holds the most compute holds the most capability, and uses that capability to hold more.
5. Coagulation. One center, with nothing left outside itself to coordinate with.

This is not a new pattern, only a faster one. It ran from farming to industrial consolidation, then software ate the world, and now AI is eating software.

Solving for: the reframe that makes the problem tractable.

Most responses to Part One try to prevent the centralization. That fails, and it fails for a structural reason rather than a political one: every force in the chain is an efficiency, and efficiencies do not lose to objections.

So concede it. Something ends up at the center. Now ask the only question that remains open: what sits there? Because centers are not interchangeable. What occupies the center determines what the center is worth, what it can extract, and whether anything survives around it.

Centralize what must be one. Distribute everything that can be many.

What must be one

Two things, and only two.

A schema — one standard that everything emitted conforms to. And a filter — one selection function that scores across all of them. Both must be singular for a hard reason: two schemas is a fork, and two filters is noise.

Here is the property that makes this the right answer. Neither one is an asset. A standard generates no rent. A judgment generates no rent. They generate throughput. A center holding those earns by increasing flow, not by owning what flows — which means its incentive is to make everything downstream of it larger, not to capture it.

TCP/IP centralized. The internet did not.

Compare that to what coagulates by default in Part One: compute, capital, and attention. All three are assets. A center holding those earns by owning what flows, and its incentive is to make everything downstream of it smaller.

Why a single decider caps out anyway

There is a second, independent reason the default center is weak, and it is the load-bearing argument in this whole document.

Any organization looking for something valuable is running experiments, and the measure that matters is how much it learns per dollar. The trap is that this depends on how different the experiments are, not how many there are. Ten experiments sharing one assumption are informationally close to one experiment. As similarity between attempts rises, the effective number of attempts hits a ceiling — and stays there no matter how much money is added.

A single decision-maker has one worldview, one sense of what is plausible, one way of scoring results. Its experiments are similar by construction. Spending grows; learning flattens.

The corollary is the uncomfortable part: getting better at this makes it worse. A sharper model prunes harder against what it already believes, which narrows the search further. It is the familiar shape of the excellent investor who develops a real thesis and then misses the next category entirely. Not a failure of judgment — an excess of consistency, and consistency is the same thing as similarity between attempts.

The live example

The frontier labs share one architecture family, one training corpus, one objective, one benchmark suite, and one talent pool that rotates between them. On paper, several independent efforts. Functionally, closer to one and a half. From outside it looks like capital expenditure climbing steeply while differences between products flatten and near-identical capabilities ship within weeks of each other. That is a learning ceiling, seen from outside.

Their deeper constraint is not compute. It is that the remaining valuable information was never written down. It lives inside specific people's specific work — what actually happens in a claims department, a shipyard, a clinic. Unpublished means unscrapable means untrainable. No amount of compute reaches it.

The failure mode nobody has a solution for

The risk that gets discussed is superintelligence. The risk that is comfortably reachable with what exists now is different and worse-specified: systems that are superhuman at holding attention, pointed at narrative control and the concentration of power.

Such a system wants more compute, more energy, and more data without limit — which is not a hypothetical, it is what the people building it say out loud, in the form of the claim that compute is the future of currency. Part One is the mechanism by which that system gets everything it wants.

Three bars

Putting the right thing at the center is correct on paper, and paper is where every good governance idea has died. Any proposal has to clear three bars, in this order:

1. Cheaper, not kinder. It has to win on cost against the default. Arrangements that require anyone to prefer them for ethical reasons do not survive contact with capital.
2. Gives power up without being asked. Not because it is virtuous, but because giving power up is the cheaper path to the input it needs.
3. Survives being wanted. It has to hold when it is working and everyone wants control of it. That is the moment every previous design has failed.

Solving for: a structure that clears all three bars.

Which shape wins

There are four candidate arrangements of centralized and decentralized. A centralized body that decentralizes itself over time. A decentralized body that centralizes over time. A decentralized body that emits centralized ones. And a centralized body that emits decentralized ones.

From first principles the last one wins, because centralization and decentralization are good at different jobs and this is the only arrangement that assigns each to the job it is good at. Centralization wins on operations — coherence, speed, standards, and everything that pairs with automation. Decentralization wins on deciding — on holding many independent views, which is the only thing that keeps the learning ceiling from closing in.

So the power concentrates and the decider distributes. Which raises the obvious objection, since Part Two said the filter must be singular. The resolution matters:

The structure

A core holds the schema and the filter. Nothing else. Everything it builds is emitted outward as a product owned, from the moment of creation, by the people who use and improve it.

The core never collects anyone's private information. It emits standardized things that owners point at their own situation, and only the result comes back. Information stays where it lives; only the score travels. That is how you reach the information the labs cannot buy — you do not acquire it, you equip it.

Ownership at the moment of creation

Every platform now offers customization. Configure it, personalize it, make it yours. But customization without ownership is lock-in with a friendly face: every hour spent shaping something raises the cost of leaving at no cost to the platform, and improvements flow upward as data rather than back as equity.

The honest test is one question — what do you still have if the company disappears tomorrow? If the answer is nothing, you were a tenant.

The structural point is that the core is never in possession. Ownership transfers at creation, so there is no later moment where the core must choose to give something up. There is nothing to be trusted about, which is precisely why it can be believed.

The release ladder

Each product that succeeds climbs a fixed ladder. Each rung fires on a published measurement, never on a decision to be generous.

1. Governance. When users are effectively running it already, that becomes formal. The core's voting share drops to a hard cap that cannot afterward be raised.
2. Capital. When it leads its category on a published metric, control of its money moves to its own treasury. The core's veto narrows to safety questions only.
3. Capability. When a defined capability threshold is crossed, authority over the standard itself transfers to the collective body, irreversibly.
4. Learning. Everything the experiment revealed — what users changed, kept, and discarded — returns and sharpens the filter. A better filter makes the next product cheaper to launch and its terms more attractive, which brings more contributors, which produces more evidence. The loop closes tighter than it opened.

What makes the first three credible is timing. Ordinary commitments to decentralize fail at the moment of execution, because whoever promised still wants the thing when the day arrives. Here each surrender is scheduled for the moment it costs nothing — the point where control has already effectively moved and the transfer only formalizes a fact. Promises that are free when they come due are the only ones reliably kept, and being reliably kept in advance is what recruits contributors before there is anything to recruit them with. This is what clears the third bar.

Two requirements keep the ladder from becoming decoration. Every trigger metric must be defined, and the distance to it published, before the product has traction — a measurement nobody can watch approaching is a discretionary decision in costume. And one thing must never vest: a residual economic claim that survives all four stages, or the core's revenue falls exactly as its products succeed and stalling becomes the rational strategy.

Meta-learning is where the return is

The filter is the only asset that compounds, because it is the only one built out of other people's work. Every emission returns evidence about what was wrong with the last filter.

This already exists in a slow form. An accelerator running batches is a meta-learning compressor: it sees many attempts, gets better at predicting which work, and publishes sharper requests for what it wants to see next. But it runs at the speed of cohorts, its compression is human and lossy, and nothing in it is agentic. The same loop, run continuously and with orchestration underneath it, is a different machine — and the tooling that makes that switchover possible is the near-term build.

Why the collective is the cheaper path

This is the argument that clears the second bar, and it does not rest on ethics at all.

A human being reaches full linguistic and conceptual competence on something like five or six orders of magnitude less data than a frontier training run consumes. That gap is currently being closed by brute force, at a cost approaching a trillion dollars. It is the most expensive mid-curve move ever attempted.

Training on a corpus average produces the middle. What it cannot produce is the tail — and the tail is where the value is. The tail is embodied, held by people who actually know things, and it is precisely what a corpus average does not contain by definition. You cannot scrape your way to it and you cannot compute your way to it.

Collective intelligence is not a nice alternative to that spend. It is the mechanism that produces the data efficiency the spend is trying to buy. Which is why giving power up is cheaper than holding it: power is the price of the input.

Why incumbents cannot follow

They can copy the architecture in a week. They can outspend on compute by three orders of magnitude. Neither matters, because the step they cannot take is conferring ownership at the moment of contribution — their ownership is already allocated.

To route equity to contributors, an incumbent would have to dilute existing shareholders in favor of people who are not yet shareholders. That is exactly the decision nobody makes, for the same reason every progressive-decentralization promise fails: when it comes due it is expensive, and the party who must pay is the party who decides.

Notice the shape of it. The incumbent's strength is the reason it cannot follow. A cap table is an asset — right up until the scarce input is something only unallocated equity can buy. Then it is a cage.

Solving for: why a correct structure still needs something to hold it together.

A currency is a shared belief system. So is a religion. The difference is smaller than it is comfortable to admit, and it means the metaphysical categories are showing up as infrastructure: mind as AI, body as robotics, and belief as the settlement layer that lets strangers coordinate without trusting each other.

Two words for the thing that has to exist where the economic and the shared-belief layers meet.

A spear is a point. It concentrates force into one place so it can pass through something. A ritual is what a collective does in unison so that the point holds. Neither works alone, and note that this is the same shape as the whole argument: a point that concentrates, a collective that coheres around it.

Coherence without a point is a mood. A point without coherence is just another center.

Which finally answers what everything is centralizing around. Not a company and not a technology. Speed to solved — solving problems fastest and implementing fastest. That is a center it is rational to want, and it is the first version of centralization that is worth defending rather than resisting.

The economy is already a general intelligence. The only open question is what it coalesces around.

What would show this is wrong

• If the standard does not hold across genuinely different use cases, it is not a standard — it is a product with settings.
• If no third party can build a viable business on top of it, the same conclusion follows.
• If no owner ever improves their product in a direction the core dislikes, then no independent experimentation is happening and the learning advantage is imaginary.
• If any trigger metric is defined after traction exists, the ladder is decoration.
• If the core ever starts operating one of its own products as its main business, the structure is finished regardless of what the documents say.
• If the collective's tail data does not actually outperform corpus-average training on real tasks, the second bar is unmet and the cost argument collapses into an ethical one.