Directed Automation and the Cost Floor: Constraint Migration, the Water Demonstrator, and the Concentration Default
Directed Automation and the Cost Floor
The constraint-migration thesis, demonstrated on water, and the concentration default it cannot cross
Unified treatment · feasibility and distribution · v1.0
What this document is
This is a single program with one central feasibility claim, a tight boundary around it, and the distributive analysis the boundary fences off, presented as one document rather than split into modules so the whole argument sits in one place. The claim: directed engineering effort drives the human-labour content of providing a need toward zero and relocates the binding cost constraint from labour, to capital, to energy and materials, where a thermodynamic floor stops the descent short of zero. The boundary: this is a claim about feasibility and cost, not about distribution. Whether the abundance reaches everyone is a separate question, governed by ownership rather than physics. The document runs in three parts. Part I states the framework. Part II works one need, water, end to end as the existence proof in the friendliest domain. Part III crosses the boundary and asks what decides whether the feasible abundance is universal, and finds that the same technology that makes it feasible pushes its distribution toward concentration. Every claim carries a label for what kind of move it is and a confidence for how secure it is, on the Thermodynamic Realism scheme. The strongest thing the program claims is narrow; the component results are well-established, their fusion is a conjecture, and the labels keep the two apart.
What is established and what is synthesised
The architecture at a glance
Part I — The framework. Four premises: P1 experience curve · P2 thermodynamic floor (three honesty clamps) · P3 variable automatability · P4 near-free software replication. Dependency D1 energy abundance. Core C1 constraint migration · C2 bounded abundance · C3 self-replication and the closure residue · C4 the migration frontier (open). Open O1 whether deliberate direction rides the curve. Boundary N1 the distribution firewall. Edge J1 Jevons · F1 falsification · standing disciplines · a counterexample watch.
Part II — The demonstrator (water). WP1 water's experience curve · WP2 water's thermodynamic floor. WC1 migration · WC2 the floor is negligibly small · WC3 the staged pathway, realised and projected · WC4 the residue. The adequacy level as a value choice; the friendly-case weight cap; WO1 to WO3 open.
Part III — The distribution (the concentration default). DP1 the concentration attractor · DP2 redistribution rode on threat · DP3 peaceful leveling is rare · DP4 cheap automated repression. DC1 the two-blade severance · DC2 the floor reduces to the owner's reasons · DC3 the broken power symmetry · DC4 the wielder is not exempt. The lever menu; counters; DO1 to DO4 open.
Two layers, judged separately. The feasibility layer (Parts I and II) is empirical and can be wrong in ways evidence would show. The distributive layer (Part III) is value-laden and agent-relative, defeasible by argument, not by experiment. You may accept one and reject the other.
Premises
Wright 1936; Farmer and Lafond 2016; Lafond et al. 2018
For a producible good, unit cost falls on a regular trajectory as cumulative production accumulates: each doubling of cumulative output yields roughly a constant fractional cost decline (Wright's law), \(C(x)=C_0\,x^{-b}\). This is one of the better-tested regularities in technology economics and has been turned into forecasts with calculated error bounds. Scope: a statistical regularity across many technologies that stayed on a curve, not a guarantee for any single technology and silent on whether a given need even has a learnable, curve-following technology. Survivorship matters: we observe curves for technologies that scaled.
Landauer 1961; Gutowski et al. 2009; Gutowski, Sahni et al. 2013
Every physical provisioning process has a nonzero minimum energy and materials requirement, a minimum exergy \(B_{\min}\) to bring raw inputs to the finished state. The second law forbids a zero-energy transformation. This is the lower bound that stops P1's descent short of zero.
(a) The floor exists but need not bind. The thermodynamics camp finds mature material processes approaching their floors; the experience-curve camp (Way et al. 2022) reports no empirical evidence for cost floors over the observed range and declines to impose them. Both hold: the floor is real, but for many technologies it sits far below current cost, so it has not yet constrained the price. "Cost goes to the floor" means toward a bound that may be near (mature material processes) or far below (technologies still descending).
(b) The floor is not uniformly low. Advanced, information-dense processes (semiconductors, nanomaterials) use energy per unit mass orders of magnitude higher than conventional ones (Gutowski et al. 2009). So bare-survival material goods have low floors, and high-tech goods have high and rising ones. Abundance is cheapest exactly for survival needs and most expensive for sophisticated comforts.
(c) The materials floor can drift up. As high-grade ores deplete, extraction energy rises; recycling and closure are what hold the floor down. Gutowski, Sahni et al. (2013) found aggressive best-available technology plus aggressive recycling reached only about half the targeted intensity reduction, with demand growth outpacing the gains. So the floor is not static and depends on recycling, which couples to C3.
Baumol and Bowen 1966; Acemoglu and Restrepo 2018
The degree to which a task's human-labour content can be displaced by capital differs across tasks. For some the irreducible input is energy, materials, and mechanical or computational work, which automate well; for others it is human relation, presence, or judgement, which resist productivity gain (the cost-disease sectors). This variation determines which needs can have their labour content driven to zero and which cannot.
The informational layer of an automated system, its designs, control policies, and learned models, copies at a marginal cost approaching zero. The intelligence that runs a machine is not the scarce part; the machine's matter and the energy to build and run it are. Replication is not execution. Copying the model is near-free; running it is not. Local autonomous control, including the inference that would drive stages 2 to 3 of Part II, consumes real electricity and specialised silicon, a high and partly rising floor (P2b), so the executed intelligence undergoes its own constraint migration rather than escaping it. P4 is a claim about the marginal cost of copying the design, not a claim that running the intelligence is free. In most physical tasks this inference cost is small beside the actuation and manufacturing energy, so it is rarely the binding term, but it is nonzero and floored, and it is not what P4 sends to zero.
The dependency the program reduces to
from P2; Way et al. 2022
Because C1 relocates the binding cost of automatable provisioning to energy and materials (P2), the program's payoff reduces to a single external condition: energy made abundant and cheap relative to the floor. This is a named dependency, not a footnote. It has strong support, since solar, wind, batteries, and electrolysers sit on steep experience curves and a rapid green transition is forecast to be cheaper than continuing with fossil fuels (Way et al. 2022). It is not guaranteed, and it is bounded by the materials for the energy system itself (battery and grid metals): the energy experience curve pulling cost down and ore-grade depletion pushing the energy system's own materials floor up (P2c) are a coupled race, and whether mining learning and recycling outrun depletion is open, with current forecasts (Way et al. 2022) leaning toward continued descent rather than a stall. Energy abundance is necessary but not sufficient: Georgescu-Roegen's warning against the "energetic dogma" holds, since cheap energy cannot substitute for scarce materials and material transformation has its own limits (P2c). If D1 fails, C2 fails with it.
Core results
from P1, P2, P3, P4
Decompose the unit cost of providing a need into labour, amortised capital, energy-and-materials, and a residual (land, coordination, risk). Directed automation drives the labour term of an automatable task toward zero (P1), while the design and control that achieve it replicate near-free (P4). As labour falls out, the binding constraint migrates: to capital, itself a produced good under P1 and so also falling, and ultimately to the energy-and-materials term, which P2 floors. The cost of an automatable need does not go to zero; it goes toward its thermodynamic floor, and the currency of the residual cost changes from human hours to joules and processed matter. The load-bearing claim. It reframes "make it free" as "drive its cost toward the energy floor and make energy abundant." The residual term (land, coordination, risk) is where the clean migration leaks; see C4.
from C1, P2, D1
Post-scarcity in the strict sense (zero cost) is unreachable for any physical good. What is reachable is abundance in the bounded sense: provisioning cost driven toward \(B_{\min}\), with that floor made small relative to an abundant energy supply (D1). "Can we make survival free" resolves to "can we make the energy and materials floor of survival small against abundant energy," an engineering and energy-supply question, not a question of whether labour can be eliminated.
from C1, P4; von Neumann 1966; Freitas and Gilbreath 1982; Freitas and Merkle 2004
Driving the labour term fully to zero, not merely low, requires capital that builds and maintains capital: self-replicating production. Von Neumann established in-principle possibility; the 1980 NASA study designed a concrete case. Their central problem bounds the claim: closure, whether a machine system can produce every kind of part it is made of from local inputs. Full material closure is not established even on paper; the realistic regime is partial closure, where most of the system's mass is self-produced and a residue of high-complexity inputs is imported. That residue is where human or external input persists. The software half of self-replication is near-free (P4); the physical half is bounded by closure and by P2. Self-replication lowers labour and capital dramatically; it does not abolish the energy-materials floor, and it depends on the recycling that holds P2(c) down.
from P3; Baumol and Bowen 1966; and the residual term in C1
C1 holds cleanly only where the irreducible input is energy and materials. Three modes of provisioning resist, for different reasons:
Relational (Baumol). Where the irreducible input is human relation, presence, or judgement (care, teaching, trust), productivity gain is limited and relative cost rises as automatable goods cheapen. Whether advancing AI moves this boundary is contested and live.
Positional and locational. Land, location, and status are inherently rivalrous and scarce, not reducible to energy and materials, and do not follow an experience curve. Housing's binding constraint is often the land, not the structure: the fabric migrates, the location does not. Positional goods are zero-sum by construction and no amount of automation democratises them.
No learnable curve. Some needs may have no technology that stays on an experience curve at all, or one whose floor is high (P2(b)). Novel pharmaceuticals, bound by discovery and trials rather than production, are a candidate.
Current disposition (asserted, not a result): bare-survival provisioning is largely material and largely migrates; flourishing includes relational and positional goods that do not. So "no human work required for survival" takes a qualification rather than standing bare. Closing condition: the frontier closes toward C1-general if a domain now believed resistant is shown to migrate to the energy floor without loss of the thing provided; it closes against C1-general if a clean, persistent non-energy floor is demonstrated under genuine automation pressure. Either is informative.
Open question on the program's own premise
The program is named for directed automation, but P1 is evidence that cost falls with cumulative production however that production is driven, not evidence that a deliberate political programme can summon the decline on demand. Whether directed, planned provision rides the experience curve as well as distributed market production is open. There is supporting evidence (the World War II demand-stimulation natural experiment suggests directed demand can drive costs down, Lafond, Greenwald and Farmer 2022) and contrary evidence (the mixed historical record of planned innovation). The honest position: the program's physics (P1, P2) is stronger than its policy premise (that direction works), and the policy upshot rests on the weaker link. This is flagged, not assumed.
The boundary: feasibility is not universality
Nothing in C1 to C4 entails that the abundance reaches everyone. "The floor should reach everyone" is an agent-relative, value-laden commitment, not a result derivable from the physics or the framework. This is the same firewall as TR's C1 firewall and is forced by the same fact: no global optimisation target, no agent-independent "ought" (TR Module 2, Part 1). Worse for the optimistic reading, the default vector points away from universality: when capital is cheap enough relative to wages the equilibrium automates all tasks and the labour share falls (Acemoglu and Restrepo 2018, 2022), so the same technology that makes the floor cheap removes the leverage that would distribute it. Under existing ownership of self-replicating capital the attractor is concentration. That claim is not left as a slogan here; it is the whole subject of Part III, which crosses this firewall and works out why.
The edge
Jevons 1865; Sorrell 2009; and Gutowski, Sahni et al. 2013
Falling unit cost does not entail falling total resource throughput. Cheaper provision raises consumption, so driving cost per unit to the floor can leave, or raise, the aggregate energy and materials drawn; material-production data show demand growth outpacing intensity gains, so total impact has kept rising. C2's "make the floor small against abundant energy" is a claim about per-unit cost and required supply, not a promise that total demand falls. Energy supply (D1), not unit cost, is the real long-run constraint.
The feasibility layer is wrong, or narrower than claimed, if: (a) a mature, high-volume material provisioning process shows no approach to a thermodynamic floor as it scales (against P2); (b) automation's cost reductions fail to relocate the binding constraint to energy and stall at a persistent non-energy floor across material domains too (against C1); (c) directed provision reliably fails to ride the experience curve that market provision rides (against O1, and against the program's policy premise); or (d) the frontier (C4) is shown to swallow even the relational and positional domains, which would remove the qualification on "no human work" and strengthen the thesis. Each is a measurable outcome, not a matter of definition.
Standing disciplines
Three rules guard against the program's characteristic failure modes, mirroring TR's flag-is-not-a-verdict and open-is-not-refuted disciplines.
No aggregation into a percent-of-GDP headline. Costs here are per-need and dynamic. They must never be summed across needs into a single static share-of-GDP figure. That move double-counts existing provision, treats GDP as a spendable budget, and freezes a dynamic quantity, which is the error this program was built to replace. Cost claims are stated per need, as trajectories toward a floor, with the denominator named.
The demonstrator never establishes the distributive claim. Part II's demonstration that water can be driven to its floor shows feasibility-in-principle only. It is never to be cited as showing that the floor will or should reach everyone. N1 governs, and Part III, not Part II, owns the distributive question.
Open is not refuted, and supported is not proved. C4 and O1 are open: the program does not expect to win them, and lists what would close them in each direction. C1's general form is a conjecture until measured across a basket of needs, not a theorem inherited from its premises.
Counterexample watch
The needs most likely to resist C1, logged with a current disposition, open to a submission that defeats the disposition. The pattern, not any single row, is the content. Water, the first row, is taken from disposition to worked demonstrator in Part II.
| Need | Resistance mode | Current disposition |
|---|---|---|
| Water, basic shelter fabric, baseline energy | None (material) | Plausibly migrates. Low floors, automatable; water is the friendly case and the demonstrator worked in Part II. |
| Food calories at scale | Mixed: land and biology bound | Contested. Arable land is positional and photosynthesis does not ride an experience curve; vertical and cellular agriculture offer a migration route but at a high energy floor (P2b). Not the clean case it first appears. |
| Housing as a place to live | Positional / locational | Splits. Structure migrates; land and location do not. The binding constraint is often the land. |
| Care, eldercare, early teaching | Relational (Baumol) | Open, contested. Irreducible human input by current evidence; AI is actively testing the boundary. |
| Modern high-tech devices | High exergy floor (P2b) | Migrates to a high floor. Cheaper over time, but the floor is high and rising with sophistication. |
| Novel pharmaceuticals | No learnable production curve | Open. Bound by discovery and trials, not production; the experience curve may not apply. |
What this part is
Part I's C1 is a synthesis, untested across a basket of needs. This part does not test it across the basket; it tests it in one place, the most favourable place, and reports exactly how far the result carries. It carries two disciplines on its face. First, it proves nothing about distribution: showing a need can be driven to its floor is feasibility-in-principle only, never evidence the floor reaches everyone (N1; Part III owns that). Second, its weight is capped by being the easy case: a worked instance in the most favourable domain establishes that the mechanism runs in at least one real case, not that it runs in most. It is also where the basket question of N1 is discharged for one need, because one cannot cost "enough water" without first deciding what enough means.
The adequacy level is a choice, not a fact
Water need is usually quoted at four levels, differing by more than an order of magnitude. Survival, replacement of physiological losses by drinking, is about 3 litres per person per day in a temperate climate and 5 in the tropics. The widely cited basic water requirement, covering drinking, hygiene, sanitation, and modest food preparation, is 50 litres per person per day (Gleick 1996). Actual use in rich countries runs far higher, on the order of 300 to 700 litres per capita per day, the upper part of which is plainly wants rather than needs.
This part reports at two reference levels: 50 litres per day as the basic-adequacy floor, and 150 litres per day as a comfortable level that sits well above physiological need and well below profligate use. Stating both shows the result is robust to the choice. Flagging the choice records that it is not a physical fact: where the line falls between need and want draws on human-need theory and the right-to-water tradition (Doyal and Gough 1991; Nussbaum; Gleick's framing of a basic requirement as a guaranteed minimum), not on thermodynamics. The physics does not tell you 50 or 150; values and need theory do. This discharges N1's basket point for the one need, and it sits on the distributive side of the firewall.
The cost decomposition for water
Apply C1's decomposition. The unit cost of delivered water splits into labour, amortised capital, energy-and-materials, and a residual.
Labour. Operations and maintenance staffing, and construction labour. Already low per unit and already substantially automated: modern treatment is run by supervisory control and data acquisition systems with automated dosing and monitoring. This is the term directed automation drives toward zero, and the term with least distance left to fall.
Capital. Plants, membranes, pumps, and the distribution network. Capital is itself a produced good on its own experience curve (P1), and for desalination that curve is measured, not assumed (WP1). So the capital term falls over time rather than setting a fixed floor.
Energy-and-materials. Pumping and separation energy, plus membranes, chemicals, and brine handling. This is where the binding constraint lands once labour and capital fall, and it is floored by physics (WP2).
Residual. Water rights, land and watershed access, and risk (drought, contamination). This is the locational and positional term, and unlike the others it does not ride an experience curve and does not migrate to energy (WC4).
Imported facts about water
Caldera and Breyer 2017; Mayor 2020
Seawater reverse osmosis capital cost has followed a measured Wright's-law trajectory: a 15 percent learning rate across more than four thousand plants commissioned between 1977 and 2015, so capex fell about 15 percent for each doubling of cumulative installed capacity. When economies of scale are separated from learning, learning is found to be the dominant driver of historical desalination cost reductions. Specific energy use tracked the same descent: from roughly 20 kWh per cubic metre for early-1970s seawater RO, to about 8 in the 1980s, to about 5, to roughly 2.5 to 4 for modern plants. This is P1 instantiated in one technology, with a real number attached.
Elimelech and Phillip 2011; Lin and Elimelech 2015; Voutchkov 2018
The minimum energy to separate fresh water from seawater is set by thermodynamics: about 1.06 kWh per cubic metre for 35 g/L seawater at 50 percent recovery, with a practical limit nearer 1.6 once unavoidable process irreversibility is included. Current best single-stage RO runs close to that limit, while total plant energy sits above 3 kWh per cubic metre because of pre-treatment, post-treatment, and inefficiencies. Crucially, the floor depends on the source: conventional treatment of surface water needs only about 0.37 kWh per cubic metre and groundwater about 0.48, against roughly 2.5 or more for seawater. This is P2 with both clamps visible at once: the floor is real (P2a), current cost sits a small factor above it and is approaching (P2a), and the floor is not uniform but rises steeply with the difficulty of the source (P2b).
Core results for water
from WP1, WP2, and C1
As labour is automated out, already largely done in operation, and capital falls along its experience curve (WP1), the binding term in delivered-water cost becomes energy-and-materials, floored by WP2.
from WP2
Take the two adequacy levels through three provisioning routes. The arithmetic is elementary and worth showing, because the conclusion is the whole point of choosing water.
| Provisioning route | Energy intensity (kWh/m³) | At 50 L/day (kWh/yr) | At 150 L/day (kWh/yr) |
|---|---|---|---|
| Surface or groundwater, conventional treatment | ~0.4 | ~7 | ~22 |
| Seawater desalination, current best total | ~3.0 | ~55 | ~164 |
| Seawater desalination, thermodynamic minimum | ~1.06 | ~19 | ~58 |
from WC1, P4, and C3
The dynamic claim, that directed automation takes water's labour content to near zero end to end, runs through four stages, and honesty requires marking which are real and which are extrapolation. Stage 0, current: labour in operations, maintenance, and construction; capital in plant and network; energy in pumping and separation. Stage 1, automated operation, realised: treatment plants already run under automated supervisory control with sensor networks and automated dosing; the labour content of operation is already low and falling, the present rather than a forecast. Stage 2, automated maintenance and self-repair, partly realised: robotic and acoustic leak detection and pipe inspection exist; autonomous repair and predictive membrane replacement are emerging, not mature. Stage 3, automated production and replacement of capital, projected: robotic manufacture and installation of membranes, pumps, and pipe, approaching self-replicating capital for the water subsystem, bounded by closure (C3), with high-complexity components, membranes above all, the realistic imported residue. Established: the floor (WP2), the source-dependence (WP2), the experience curve (WP1), and stage-1 automation. Projected along known robotics and P1 trajectories: stages 2 and 3. So the proof demonstrates the mechanism is applicable and underway in the friendly domain, with the labour-to-zero endpoint as a supported projection, not an accomplished fact. The endpoint is not claimed as achieved.
from the residual term in WC1; C4
Even water leaks at the residual. The locational and positional component: watershed geography, aquifer access, transboundary rivers, and water rights are scarce and rivalrous and do not ride an experience curve. Desalination removes the source constraint for coastal regions but not for the interior, where the lifting and transport energy is itself substantial, so the residue is smaller for the coast and real for inland and high-elevation demand. The materials component: brine disposal, intake ecology, and membrane replacement (P2c, J1). The risk component: drought, climate variability, and contamination events. These exist and cap how completely water migrates; they do not dominate the cost, and that non-domination, more than anything, is why water is the demonstrator and food and care are not.
Why water is the friendly case, and the boundary
Water is the demonstrator precisely because its irreducible input is energy and materials, the one condition under which C1 runs cleanly. Its relational input is near zero, its provisioning is infrastructure-dominated, its source is substitutable through desalination, and its locational residue, while real, is manageable. These are exactly the properties the hard cases lack.
So the result does not generalise by itself, and this is where the proof stops. Food is land- and biology-bound: arable land is positional, photosynthesis does not ride an experience curve, and the migration route through vertical or cellular agriculture carries a high energy floor. Care, eldercare, and early teaching are relational in Baumol's sense, where the irreducible input may be human presence and judgement automation has been worst at, and whether AI moves that boundary is open (C4). Novel pharmaceuticals are bound by discovery and trials rather than production, so the experience curve may not apply at all. Water is silent on every one of these, because water is the case with none of their difficulties. The clean demonstration here is a beachhead, not a conquest, and the critic's correct next words are "now show me eldercare."
Water: falsification and open questions
The water claim is wrong, or narrower than stated, if SWRO specific energy stalls well above the thermodynamic floor as capacity scales (against WP2 and WP1); or the labour content of water provision proves irreducible under genuine automation pressure (against WC1); or the locational residue (WC4) is shown to dominate delivered cost rather than sit beside it, which would reclassify water from a friendly case to a resistant one. Each is measurable.
The realised-versus-projected gap in WC3 is the central open question for the pathway. Whether self-repair and self-replicating water capital arrive, or stall at partial closure with a persistent labour and high-complexity-import residue (membranes especially), decides whether the labour term reaches near zero or settles at a low but nonzero plateau.
WC4's locational residue binds to the degree demand is locationally captive rather than reachable by desalination plus affordable transport. For what fraction of demand, and at what added energy, water is source-flexible sets the size of the residue and is not estimated here.
WP1's learning rate was measured where cost sat well above the floor. As SWRO energy approaches the thermodynamic limit (WP2), the curve must flatten, since the floor cannot be crossed. Where the flattening sets in, and how close to ~1.06 kWh per cubic metre practical plants get, is the P2(a) question for water specifically.
What this part is
Parts I and II establish feasibility-in-principle, and N1 fences it off from distribution. This part is what sits behind that firewall. Its job is to map the fork, not to close it: a structured account of why the default vector points one way, what would have to be done to bend it, and why each corrective faces the same obstacle. It does not hand down a solution, because the ownership question is genuinely unsolved and heavily contested, and a part that pretended to settle it would be overclaiming on exactly the side where the evidential standard is weakest. It is descriptive and predictive-political throughout; its single normative input, that a universal floor is worth wanting, is an agent-relative commitment grounded in N1 and in TR Module 2, not a result, and the firewall is restated at the end so the map is never read as a verdict.
One distinction first: the floor is not the gap
The literature this part draws on is overwhelmingly about inequality, the size of the gap. Parts I and II ask about a floor, the absolute minimum everyone is guaranteed. These come apart, and conflating them is the first error to remove, because it is easy to import a pessimistic inequality result and mistake it for a pessimistic floor result.
Question A, the floor. Does the cheap abundance reach everyone as a guaranteed minimum? A generous universal floor is fully compatible with extreme inequality above it: a society can guarantee everyone a comfortable baseline while a few own astronomically more. This high-floor-high-gap outcome is close to what the optimistic version of the original ambition actually wants, and it does not require equality.
Question B, the gap. Is the resulting distribution egalitarian? This is what Scheidel, Piketty, and Acemoglu-Restrepo measure.
The findings below bear hardest on Question B. They bear on Question A more weakly and more interestingly, through a separate channel, whether the owners of automated capacity have any reason to fund a universal floor once they no longer need the people it would serve. Keeping the two apart is what stops this part from being uniformly gloomy where the evidence is mixed.
Imported results
Acemoglu and Restrepo 2018, 2022
In the standard task-based model, if the rental rate of capital is low enough relative to wages, the long-run equilibrium automates all automatable tasks, the labour share of income falls, and the wage-rental ratio falls with it. Empirically, automation-driven decline in routine-task wages is assigned a large share of four decades of US wage-structure change. The point for this part: the same cheapening of capital that drives C1's migration also lowers labour's share of the gains. The mechanism that makes the floor cheap is the mechanism that erodes the bargaining position of the people who would need it distributed.
Acemoglu and Robinson 2006; Boix 2003; Aidt and Jensen 2014
On the dominant political-economy account, the great extensions of the franchise and the construction of redistributive states were not gifts. Elites conceded political power and redistribution when non-elites could credibly threaten disorder or revolution and when the cost of repressing them was high relative to the cost of conceding; democratisation functioned as a credible commitment to future redistribution, made under duress. The empirical record bears this out: across Europe from 1820 to 1938, suffrage extensions track the threat of revolution, with some role for war, while modernisation accounts fare poorly. The mechanism is leverage: the people had to be reckoned with because they could impose costs.
Scheidel 2017
Across recorded history, the large compressions of inequality were triggered by violent shocks: mass-mobilisation warfare, transformative revolution, state collapse, and lethal pandemics. In stable, peaceful conditions, inequality tends to rise toward the maximum the social structure permits. The twentieth-century "great compression" is, on this account, substantially a product of the World Wars rather than of autonomous benevolence, and inequality resumed rising once those pressures faded. Contested in scope, but no compelling counter-model of large-scale peaceful leveling has displaced it. The import: spontaneous, non-coerced equalisation is the exception, not the baseline.
Frantz, Kendall-Taylor and Wright 2020; Svolik 2012; counter in Yang 2026
The second historical brake on elite predation, alongside the threat from below in DP2, was the principal-agent problem of repression: a regime that depends on police and soldiers is vulnerable to those agents defecting or refusing, so coercion was labour-intensive and politically fragile. Automated surveillance and AI-driven monitoring cut that cost sharply and reduce dependence on a large human enforcement class that could defect. This is the same automation dynamic as C1, now applied to coercion rather than provisioning.
Core results
from DP1 through DP4
Why did concession ever happen? On DP2, when non-elites could compel it. Their power to compel had two distinct sources, worth keeping apart because automation attacks them by different routes. The first is economic leverage: the productive system needed their labour, so withholding or disrupting it imposed real cost on elites (the strike, and the Black Death labour shock). The second is the threat of disorder: labour aside, a population can riot, sabotage, and revolt, and historically this threat compelled concession only because repressing it was expensive and fragile, since repression required human enforcers who could refuse or defect (DP4's principal-agent constraint).
Self-replicating automation neutralises both sources, but not the same way, and the asymmetry is the more defensible claim. The economic source it severs directly: labour becomes dispensable (DP1), so withholding it stops costing the owner anything. The disorder source it does not remove, a dispossessed population can still threaten disorder and can innovate in resistance that does not depend on labour at all, sabotage, evasion, and its own use of cheap tools for asymmetric disruption; what automation does is defuse the threat by collapsing the cost of repression below the cost of conceding (DP4), so the threat compels less. Run through the Acemoglu-Robinson model, the threat-from-below input and the cost-of-repression input fall together, and those are exactly the two inputs that forced elites to concede. So the default vector of the technology, under unchanged ownership, points toward concentration. This is the central finding and the uncomfortable one: the technology that makes a universal floor physically cheap simultaneously dismantles the machinery by which floors have historically been made universal.
from DC1 and C2
DC1 bears on the gap (Question B). The floor (Question A) runs through a different question: given the people are no longer needed, does whoever owns the self-replicating capacity have any reason to fund a universal floor? Historically there were three instrumental reasons to keep a population provisioned: as labour, as soldiers, and as consumers. Automation removes each. Labour: removed by DP1. Soldiers: removed, since automated force needs no mass conscription. Consumers: removed in the limit by self-provision, since an owner whose capacity produces anything they want does not need a market to sell into.
That last severance is conditional, and the condition is strong: it holds only if the owner's wants are satisfiable by closed-loop self-provision. If elite utility includes culturally dense goods, positional goods, novelty, or services a broad human ecosystem alone can generate, then a residual reason to sustain a population survives, because autarky cannot manufacture a culture to be highest in or a market to be richest within. So the consumer reason severs fully only for an owner whose definition of the good has decoupled from any human network, and partially otherwise. What remains once the instrumental reasons are gone is two things only: the owner's own values (an agent-relative commitment to universal flourishing, which some hold and some do not, which the framework cannot manufacture), and a residual coercive threat (which DP4 is eroding). So the universality of the floor reduces, in the abundance limit, to the owners' preferences plus whatever threat survives automation of enforcement, a thinner and more contingent foundation than the hard instrumental need that built the twentieth-century floor. The residual-threat term carries one consequence worth drawing out for Question A: managing that threat can mean provisioning rather than repressing, since provisioning a surplus population is a pacification cost, the bread of bread and circuses, and where the floor is near-free (C2) feeding people may be cheaper than containing them. That is a thin instrumental reason to provision even absent labour, soldiers, and consumers, and it is the owner's prudential choice rather than restored leverage, so it does not contradict DC1. It cuts toward Question-A optimism. DP4 cuts the other way on the same ledger by cheapening the alternative of repression, so pacification-by-provisioning prevails only where it is the cheaper of two cheap options, a contingent fact about relative costs and not a guarantee.
from DC1; Gauthier 1986; Hobbes
There is a thermodynamically respectable, fully agent-relative route to cooperation that needs no global optimisation target: conflict is dissipative, so agents with conflicting persistence-oughts can each, from their own standpoint, prefer a low-friction cooperative arrangement to costly conflict. This is the contractarian convergence, legal within the framework because each agent endorses the equilibrium for its own reasons, with no agent-neutral "ought" smuggled in. But the convergence is conditional on rough power symmetry. Rational bargaining yields cooperation among parties who can each impose costs on the other; those who can be excluded or coerced at negligible cost fall outside the cooperative surplus, the standing objection to contractarianism, that it under-protects those who cannot reciprocate or retaliate. Post-scarcity automation is precisely a symmetry-breaking event: by DC1 it moves the dispossessed toward the costlessly-excluded category. So contractarianism, run honestly, predicts cooperation fails exactly in the world this program theorises, because the condition the convergence needs is the condition the technology removes. Descriptive and predictive, not normative. DC3 says what a power-symmetry account predicts; it does not say the predicted outcome is good, nor the egalitarian one bad. The ought stays on the far side of the firewall.
from DC1 applied to the state; Acemoglu and Robinson 2006; Piketty 2014
The lever menu that follows presupposes a state independent and capable enough to wield the levers. But the state is not outside the dynamic DC1 describes; it is inside it, and two of its load-bearing properties erode under the same automation. Its fiscal base: a state that funds itself by taxing labour and consumption loses revenue as labour is automated out, so its capacity to fund redistribution or public provision contracts precisely as the need for them rises, unless it can tax capital, which is mobile and concentrated (DO4). Its independence: as capital concentrates (DP1; Piketty's secular tendency of the capital share to rise), the practical distinction between the state and the largest owners erodes through capture, until "the state regulates the owners" shades into "the owners regulate themselves." So the levers face not only a closing window but a wielder whose own capacity and independence are being severed by the same process the levers are meant to counter. The instrument is part of the problem, and that is why the shared failure mode below is structural rather than incidental.
The lever menu
If the default is concentration, what bends it? Four classes of corrective appear in the literature, presented as a map of options with their characteristic failure modes, not as recommendations and not ranked. The unifying observation is the final row: every lever requires, to be enacted, the very power symmetry that DC1 is eroding, and a wielder that DC4 shows is itself eroding. The map is the content; no single row is endorsed.
| Lever | Mechanism | Characteristic failure mode |
|---|---|---|
| L1. Predistribution of capital Meade's Citizens' Trust; Atkinson's capital endowment and public sovereign wealth fund; partial instances (Alaska, Norway) |
Spread ownership of the productive capital itself, so that as the labour share falls the broadly-held capital share rises to compensate. Attacks the problem at its source: who owns the seed. | The stake must be acquired before concentration locks in, and acquisition itself faces the default: buying or taxing one's way to broad ownership requires the political leverage DC1 removes. Real instances are mostly resource-windfall-dependent, not generalisable on demand. |
| L2. Redistribution by tax and transfer capital taxation funding a basic income; Piketty-style wealth taxation |
Leave ownership concentrated but tax the returns and transfer them, funding a cash floor from the output of automated capital. | Requires sustained capacity to tax mobile, concentrated capital, against owners who by DC1 need not concede and a state that by DC4 is losing its labour-tax base and its independence. DP3 supplies the base rate. The one mercy is that capital's exit is not free (DO4): it still needs markets, infrastructure, and location, so some fiscal leverage survives, by an amount that is open. |
| L3. Public provision of the floor Universal Basic Services; the de-commodified baseline (Coote and Percy 2020; Gough 2019; Portes et al. 2017) |
Provide the floor directly in kind (water, energy, shelter, transport, information) rather than as cash, exploiting the result that the material floor is cheap. | Still requires the state to own or fund the productive capacity, so it reduces to a version of L1 or L2 upstream and inherits DC4. Adds quality, capture, and provisioning-monopoly risks of its own. |
| L4. Structural limits on capability monopoly antitrust on the seed factory; mandated openness of designs (P4: software replicates near-free, so the control layer is the cheap part to share) |
Prevent the concentration rather than redistribute after it. Keep the self-replicating capability plural and non-monopolised, so no single owner reaches the symmetry-breaking position of DC1. | Enforcement runs against entities that own autonomous capability, raising the "who guards the guards" problem, sharpened by DC4. Openness of designs disperses capability but not the energy and materials floor (C2), which can still concentrate. |
| The shared failure mode. All four require enactment under power symmetry, and DC1 is the erosion of power symmetry; all four are wielded by a state whose own capacity and independence DC4 shows to be eroding under the same process. The levers are therefore time-bounded twice over: more installable the earlier they are attempted, and dependent on an instrument that is itself decaying. This double bind, not any individual lever's weakness, is the load-bearing difficulty of the distributive problem. | ||
Counters, taken at full strength
Symmetric vigilance requires the strongest case against the concentration default, not a strawman. Four serious counters, each with the honest extent of its force.
The cheap-floor counter, which is the strongest
C2 says the material floor can be driven near its thermodynamic minimum and held there by abundant energy. If providing everyone the floor costs the owner almost nothing, then withholding it is no longer resource conservation; it is nearly pure control. So in the abundance limit the owner's cheapest move might be to grant the floor and reserve scarcity-competition for what stays scarce. This is a real argument, and it cuts toward Question-A optimism more strongly than the rest of this part might suggest.
Its limit is positional, and that is the firmer objection. Status is relative: a floor that everyone has stops signalling status, so what an elite secures by withholding a cheap good is not the good's resources but the hierarchy the scarcity sustains. Elites have historically manufactured and defended scarcity to preserve rank, not to conserve resources, which is C4's positional mode applied here, the floor's material cost can fall to nothing while its positional meaning keeps it contested. The Black Death has to be stated carefully to make this point rather than miss it: post-plague wage-fixing was elites spending real effort to suppress a labour-scarcity gain, but it was about preserving a labour relationship and relative position, and the labour relationship is exactly what automation removes, so the case transfers only on its positional half, not "elites need wages kept down" but "elites act to preserve relative position even at a cost." On that half it transfers cleanly. So cheapness genuinely lowers the barrier to a universal floor, more than a pessimist would grant, and the residual obstacle is the positional and control value of withholding, which DC2 already located as the owner's values plus residual threat. The honest verdict: C2 makes a universal floor easier and more likely, especially for Question A, and it does not make it automatic, because what blocks it was never mainly resource cost.
Institutions can outlast their enabling conditions
The post-war welfare states and mass democracies delivered broad distribution for roughly three generations. Even granting DP3 that the World Wars triggered them, the institutions persisted after the triggering violence passed, so distribution-sustaining institutions can run on inertia for a long time. The limit: persistence is not robustness against a new shock that removes the original supports, and DC1 with DC4 is exactly such a shock, aimed at both the leverage and the fiscal base that underwrote those institutions. The counter establishes a floor can survive for a while without its founding leverage, not that it survives this specific erosion; the interaction of inertia and shock is itself unsettled, taken up in DO1.
The default is an attractor, not a destiny
DP1 is explicitly conditional, and the direction of technical change is a policy variable. Subsidising labour-complementary rather than labour-replacing automation, and pricing capital differently, can move the parameter regime out of the all-automating equilibrium. This is correct and is the strongest reason DC1 is written as a default rather than a prediction. Its limit is DC1's own timing point: redirecting the technology is itself a political act requiring the leverage the technology erodes, so the counter is real but inherits the window problem.
Automated repression may be self-undermining
Per DP4's clamp, AI-enabled coercion has inherent and partly self-inflicted weaknesses. Replacing human judgement with rigid thresholds trades flexibility for brittleness; once an algorithm is seen to err, deference falls; and the authoritarian data problem means the strategic silence repression induces degrades the data the surveillance system learns from, so accuracy falls as repression rises. Recent events fit: AI street-camera enforcement failed to quell mass protest, and the regimes in question survived through old-fashioned violence rather than seamless control. The dispossessed also retain forms of costly resistance independent of labour leverage. If automated enforcement is this brittle, the coercive blade of DC1 is duller than its clean version claims, and the residual-threat term in DC2 survives longer. Its limit, against over-correction: this weakens one of the two blades and leaves the economic blade (DP1) fully cut, so it makes concession less compelled, not compelled. The default still points toward concentration, just less steeply and with a longer residual-threat tail.
Open questions
The central open question, stated so a pessimistic answer is not mistaken for doom. Suppose the honest answer is that no distributive regime is a costless stable equilibrium once labour leverage is gone, that each lever requires continuous enforcement against a default that reasserts whenever vigilance lapses (the pattern DP3 reports for inequality in peaceful conditions). That would not entail collapse. It would mean a universal floor is like any institution: maintainable, but only by permanent political work. The real question becomes whether that work can be routinised, made low-cost, or constitutionally embedded in structures hard to unwind, below the cost at which the concentration default and the state-erosion of DC4 would overwhelm it. This is also where the inertia counter and the DC1 shock meet, and which force wins is settled by neither alone. The framework leans toward "no costless equilibrium, permanent maintenance required," and treats the routinisation question as open.
C4 leaves open whether automation eats the relational domains (care, judgement, trust). If it does not, some human roles, and so some residual leverage and some residual need for a provisioned population, survive, partially restoring the conditions DC1 removes. If it does, DC2's severance is more complete. Shared with C4 and not resolved here.
Every lever is time-bounded by DC1 and DC4. The duration of the window in which power symmetry and state independence are still sufficient to install a corrective is the practically decisive unknown, and it is not estimated here. Naming it as the operative variable is the contribution; quantifying it is not attempted.
This part is geographically single-framed, and that framing hides structure rather than removing it. The likely real outcome is not one timeline but a patchwork: some polities installing predistribution or sovereign-fund models, others attempting capital-funded transfers, others sliding to concentration, with the mix shaped by cross-border dynamics the single-polity frame omits, capital flight, regulatory arbitrage, and migration pressure. One sub-question matters for L2 in particular: capital's exit is not free. Self-replicating capital still needs markets, infrastructure, legal protection, and physical location, so the threat "tax me and I leave" is bounded by what leaving costs the owner. How bounded, and therefore how much fiscal leverage survives capital mobility, is open, and is the difference between L2 being hard and L2 being impossible.
The firewall, restated
What this document does not claim
On feasibility (Parts I and II). It does not claim post-scarcity in the strict sense; the floor is nonzero, only small (C2, WC2). It does not claim all needs migrate; it claims it for material cases and works only water, naming the resistant domains it excludes (C4). It does not claim water's labour-to-zero endpoint is built; stages 2 and 3 are projected, not demonstrated (WC3). It does not claim energy abundance is solved (D1, a dependency). It does not claim directed provision matches market provision in riding the curve (O1, open). It does not define "a need" or the adequacy level as a physical fact; that basket is partly a value choice, discharged for water by a stated and defended choice. It is a claim about the asymptotic cost floor and the direction of migration, not about the transition path, its duration, ordering, or transitional cost.
On distribution (Part III). It does not claim concentration is certain; DC1 is a default and an attractor, defeasible by deliberate action taken in time. It does not claim that building the technology delivers universal abundance; it claims the opposite, that the technology resists universality and that universality is a choice made against the grain. It does not claim a universal floor is impossible; the cheap-floor counter and DC2 leave Question A more open than Question B. It does not derive any ought; the firewall governs. It does not recommend a lever; the menu is a map of options and failure modes. It does not settle the ownership question; DO1 is open.
Throughout. It is geographically single-framed: the cross-border patchwork and the limits of capital mobility are logged as DO4 rather than resolved, and cost figures and feasibility are likely country-specific.
What is canonical here
The load-bearing feasibility claim is C1, with C2 as its consequence, P2 as the floor that keeps both honest, and D1 as the dependency the payoff reduces to. C3 and C4 bound it: C3 says full labour removal is gated by closure, C4 says migration may not reach the relational, positional, or no-curve domains. O1 flags that the policy premise (direction works) is weaker than the physics. Part II is the existence proof in the single friendliest domain: water's binding cost is migrating to an energy floor that is negligibly small (WC1, WC2), the pathway realised through automated operation and honestly projected beyond it (WC3), with a residue that does not migrate (WC4), and its weight is capped by being the easy case.
N1 is the firewall, and feasibility-in-practice is conceded to be entangled with distribution. The load-bearing distributive finding is DC1, the two-blade severance: directed automation removes both the economic and the coercive leverage by which floors have historically been made universal, so the default vector under unchanged ownership is concentration. DC2 specialises this to the floor; DC3 supplies the mechanism in contractarian terms; DC4 closes the loop by showing the state that would wield the correctives is subject to the same severance, which makes the lever menu's shared failure mode structural. The cheap-floor counter is the strongest reason Question A may resolve more hopefully than DC1 alone suggests, limited by positional rather than resource scarcity. The open crux of the whole document is double: C4 and O1 on the feasibility side, DO1 on the distributive side. The firewall keeps the two layers from contaminating each other: the physics is empirical and falsifiable, the distributive map is value-laden and the value is named and owned, never derived.
References
Verification status per entry, on the TR Module 10 scheme. Verified checked against the primary or an authoritative source by search. Standard a canonical work cited from bibliographic knowledge; confirm formatting before submission.
One-sentence collapse: An ultra-low-bandwidth condensation of the paper above:
ReplyDelete"Synthesizing technology economics with thermodynamic invariants, the framework posits that directed automation drives the human-labor content of provisioning toward zero, migrating the binding cost constraint to an irreducible physical floor of energy and materials—demonstrated by the negligible energetic requirement of universal water adequacy—while warning that under default ownership dynamics, this automated abundance drives severe wealth concentration by systematically dismantling labor's distributive leverage."