Compute as Public Capital

I. The Displacement Trajectory

For most of industrial history, automation displaced tasks laterally: a machine displaces a task, and workers retool into an adjacent task the machine cannot yet perform. This was often brutal for the generation living through it, but it worked because the automating technology was narrow — a machine that could weave cloth could not also drive a truck or keep a ledger.

That lateral escape route is closing. The same underlying AI systems now displacing cognitive tasks are, in parallel, being embodied in robotics capable of physical labor — closing adjacent exits at the same time as the original one. Consider an ordinary case: a homeowner who pays a local contractor to mow the lawn — physical, local, relationship-based work, exactly the kind displaced workers have historically been told to move toward. A robotics-and-logistics company dispatching an autonomous mower to a thousand yards a day undercuts that worker's entire category, not one client, and the standard policy answer — retrain, move up the value chain — runs into the same wall one step later.

This claim should be stated against, not around, the dominant framework in labor economics for this exact question. Acemoglu and Restrepo's displacement/reinstatement model (Journal of Economic Perspectives, 2019) treats automation's labor-market effect as a race between a displacement effect (automation removing tasks from labor) and a reinstatement effect (new tasks in which labor retains a comparative advantage). Historically, reinstatement has offset displacement. A separate, related paper by the same authors (Econometrica, 2022) quantifies the resulting distributional effect directly: automation accounts for an estimated 50 to 70 percent of the changes in the US wage structure over the four decades to 2016, working through the relative wage declines of worker groups specialized in routine, automation-exposed tasks. Recent work extending the displacement/reinstatement framework to agentic AI systems — capable of completing entire occupational workflows rather than discrete tasks — argues this expands displacement risk beyond what task-level analysis previously captured.

The pace of convergence between cognitive and physical automation is also treated incorrectly by current policy, which largely regulates "AI" and "robotics" as separate domains. That separation no longer reflects the technology. Any policy response built on this paper's argument has to start by collapsing that false distinction: AI policy and automation-labor policy are the same policy, evaluated too late.

II. Why Incumbent Privilege Resists the Transition

It would be convenient to treat resistance to this transition as simple bad faith, and this paper is deliberately not going to make that argument, because it is not the accurate one and it will not survive contact with a skeptical reader. The more useful — and more uncomfortable — explanation is behavioral, not moral.

Efficiency-seeking is close to a base biological drive. An organism, and by extension a society built by organisms, that can secure the same or better outcome for less expenditure of effort will consistently choose to do so. Privilege — durable, structural advantage that reduces the effort required to sustain one's position — is simply the most refined expression of that drive available to a modern economic actor. It is not that the wealthy are unusually self-interested compared to everyone else. It is that everyone is running the same optimization, and privilege is what winning consistently at that optimization produces and then protects.

This matters because it explains, without requiring any claim about individual malice, why incumbents will not treat the current trajectory with the urgency the rest of society will eventually be forced to. A position secured under the current rules has every incentive to describe them as immutable, natural, or already fair. This is a structural prediction, not an accusation. It should be read as an account of the incentive, not the full account of the mechanism — lobbying, campaign finance, and regulatory capture are the concentrated political channels through which diffuse behavioral incentive becomes organized resistance, and a complete treatment of this argument names both.

It is also worth being direct about where this argument's author sits in this dynamic. For three decades of a working life, and now as an employer directing the labor of others, the pattern this argument critiques is the author's own: every hour of efficiency captured from someone else's labor is an hour the capturer does not have to spend themselves. This is offered as evidence the pattern is structural and nearly universal, not a flaw limited to the ultra-wealthy.

III. The Distribution Break

For most of the industrial era, the distributive circuit has run in one direction:

A highly automated economy increasingly runs a different circuit:

No arrow in the second circuit automatically returns purchasing power to the broader population. This is not speculative: US labor share of income fell from 67.8 percent in 1987 to 58.1 percent in 2023 (BLS/FRED series MPU4910141, Private Nonfarm Business Sector: Labor Share), and Karabarbounis and Neiman (2014) document this decline as global, occurring across most countries and industries since the early 1980s, driven substantially by technological substitution of capital for labor.

One caveat worth stating directly: capital income is not received exclusively by a small ownership elite — a meaningful share flows to ordinary households through pensions and index funds. The labor-versus-capital framing below is a simplification; the real distributional question is how concentrated versus broadly held capital ownership is, which is closer to Section V's recursive-concentration dynamic than to a clean two-population split.

A minimal formalism (conceptual notation, not a calibrated model)

The following is offered strictly as illustrative notation clarifying the logic of the argument. No term below has been decomposed into anything measurable, and no numerical prediction is derived from it — doing so is the subject of Section XV's research agenda, not a result claimed here.

Let Y_t = W_t + R_t, where W_t is labor income and R_t is capital/ownership income in period t. If automation drives labor's share of income down (W_t/Y_t ↓) while ownership of R_t stays concentrated, the relevant policy question becomes the distribution of effective purchasing claims across households. Introducing a distribution term D_t (dividends, transfers, public services), household consumption capacity is bounded roughly by:

where R_h,t is broadly held household capital income and B_t is borrowing. The insight this makes explicit: debt can temporarily substitute for missing wage distribution, but it cannot permanently repair the circulation problem — B_t can smooth a shortfall for a period, but it accumulates a claim against future C_t rather than resolving the structural gap between W_t and Y_t.

IV. The Two-Sided Municipal Ledger

An earlier version of this argument treated municipal impact as one-directional: automation displaces wages, displaced wages reduce local spending, reduced local spending erodes municipal tax bases. That direction is real, but the evidence is more complicated than a single-direction story, and the complication is preserved here rather than smoothed over.

Data centers — the physical infrastructure of the automated economy — currently generate substantial, sometimes dominant, local property-tax revenue in the jurisdictions that host them. Data centers reportedly account for roughly 38 percent of Loudoun County, Virginia's total revenue and about 75 percent of Quincy, Washington's property-tax levy. This directly complicates any simple claim that automated capital does not pay into the local tax base — in the near term, capital-intensive automated infrastructure can be a substantial local fiscal benefit, not a drain.

The more precise claim, once this is accounted for, is two-sided rather than one-sided: the erosion channel is specifically about labor-linked local revenue (sales tax, local income tax tied to displaced service-sector wage spending), not about automated capital as such, which can generate large property-tax windfalls independent of local employment. That windfall is neither guaranteed nor stable — it depends on tax-incentive policy currently being rolled back in several states as costs become visible: Georgia's data-center tax abatements are estimated to cost localities $1.1–1.4 billion in 2026 alone, and Texas's cumulative exemption cost is projected near $9 billion through 2030.

The honest formulation of the municipal-cascade hypothesis is therefore: automated economic activity can simultaneously strengthen the property-tax base of the specific jurisdictions that host its physical infrastructure while eroding the labor-linked tax base of the much larger number of jurisdictions that do not — a redistribution of municipal fiscal fortune, not a uniform municipal fiscal decline.

This is a testable, falsifiable claim, with two unresolved complications named directly: a selection-effect confound (hosting jurisdictions are chosen partly for favorable tax conditions), and an unexamined question — whether existing state fiscal-transfer mechanisms already partially offset the geographic mismatch described here. Both are named as open items for the research agenda in Section XV, not resolved in this paper.

V. Recursive Capital Concentration

A further dynamic compounds the distribution break, and explains why waiting until displacement becomes severe before acting is materially more dangerous than for a conventional, slower-moving economic shift.

Unlike human labor, whose productive capacity is tightly constrained by biological time and individual throughput, automated productive capital can be replicated, scaled, and financed from its own returns across a much wider range before physical, competitive, or economic constraints impose diminishing returns. This is not a claim that automated capital faces no diminishing returns — it plainly does, through competition, energy constraints, depreciation, and technological obsolescence. The distinction that matters is capacity for replication and scalable reinvestment, not immunity from constraint.

These two loops actively diverge, because the population losing income is simultaneously losing the capacity to acquire the asset class compounding fastest. This is the strongest argument in the paper for urgency over gradualism: a policy response that waits until concentration becomes severe is waiting for the harder loop to entrench further while the easier stage to interrupt has already passed.

VI. Institutional Revenue Decoupling

Sections III and V describe the household-level distribution break and its recursive concentration dynamic. This section formalizes a related but distinct claim: the same break propagates into institutions built around labor as their financing proxy, independent of whether any individual household is protected.

Working definition

Institutional Revenue Decoupling occurs when automation changes the relationship between aggregate productive capacity and the legacy economic proxies (payroll, wage income, employment counts, human-operated risk) used to finance institutions, causing institutional revenue or risk-pool structure to stagnate, decline, or destabilize even while underlying productive output rises.

Tested against the existing architecture: this is not a fully independent phenomenon from the Distribution Break — it is the same underlying mechanism viewed from the institution's side of the ledger rather than the household's. The term is retained because the policy implications differ, not because it names a new causal mechanism.

Extended formalism (hypothesis, not established theorem)

Let P_t be aggregate productive output, T_L,t revenue from labor-linked taxation, T_C,t revenue from capital/output-linked taxation, and G_t institutional/public-service obligations. If P_t rises while W_t/P_t falls, then under a tax architecture heavily dependent on W_t: T_L,t/P_t falls.

The hypothesis this motivates: a state can experience increasing productive capacity alongside deteriorating fiscal capacity if its tax architecture remains attached to a shrinking factor of production — a distinct claim from "the economy is doing worse," and arguably a more precise diagnosis of what rising GDP alongside strained public budgets would actually look like.

A US-specific complication, precisely sourced: the United States is the only OECD member country without a value-added tax — a consumption-based, non-payroll-linked base structurally closer to T_C,t than T_L,t (OECD, Consumption Tax Trends 2024). As a direct consequence, the US collects roughly 16.8 percent of total government revenue from consumption taxes, compared with an OECD average of about 31.1 percent (Tax Foundation, compiling OECD data). This means the tax-base transition this section describes is plausibly harder for the US specifically than OECD-comparative literature implies: a US shift toward non-payroll-linked revenue would need to build collection infrastructure most OECD peers already have.

VII. Institutional Case Studies

VII.A — Social Security and Social Insurance

Social Security is financed pay-as-you-go from payroll taxes. Its long-run funding pressure is an established, independently documented fact, not a projection this paper introduces: CBO's long-term projections for Social Security (2023) show the Old-Age and Survivors Insurance trust fund exhausted in fiscal year 2032 and the Disability Insurance trust fund exhausted in 2052, with a 75-year actuarial deficit equal to roughly 1.7 percent of GDP, or about 5.1 percent of taxable payroll. This gap is currently attributed primarily to demographic shift — a declining worker-to-beneficiary ratio and slower labor-force growth — not automation. This paper's contribution is a hypothesis, not yet established: that a declining labor share of production compounds the same funding gap through a second, independent channel.

This hypothesis should not be overstated, and a genuine complication in the evidence is preserved here rather than smoothed over. A 2026 study using commuting-zone data and an instrumented measure of robot exposure in manufacturing found that each additional robot per 1,000 workers is associated with lower, not higher, SSDI application inflows in exposed labor markets, and did not find declining employment-to-population ratios in those same markets — which weighs against broad local displacement as the primary channel, at least for that specific, narrower exposure measure. This finding is scoped to robot-intensive manufacturing labor markets, not the broader, general-purpose AI/robotics wave this paper's Section II argues is structurally different from prior automation.

VII.B — Insurance

Traditional insurance underwriting depends on pooling large numbers of partially independent human risks. Automation does not eliminate insurable risk; it changes its topology. As human operation gives way to shared AI models, cloud platforms, and robotics fleets, previously independent risks can become correlated. Correlated catastrophic loss is not itself new to the insurance profession — hurricane and earthquake catastrophe risk, and the 1980s liability crisis, are established correlated-loss categories already priced through reinsurance and catastrophe bonds. The genuinely novel problem is narrower: the speed and opacity of AI-model-driven correlation, where a single model update can correlate losses across an entire book of business within days rather than over a hurricane season's timescale.

This is the active subject of a fast-emerging 2026 actuarial literature. Work on the insurance of agentic AI identifies foundation-model concentration as a channel through which upstream model failure can correlate losses across many insured parties at once. A further constraint worth naming: global reinsurance capital is finite and already under pressure from climate-driven catastrophe losses, so a new, poorly understood, potentially correlated risk category competes for the same limited capacity rather than arriving into room to spare.

VII.C — Bounded Institutional Scan

A brief, deliberately non-exhaustive scan: mortgage underwriting and consumer credit scoring are built substantially on verified wage income, and bank balance sheets are exposed both to commercial real estate and to consumer credit underwritten against an assumption of stable wage income across the loan book as a whole. Higher-education financing is priced against anticipated future labor-market returns. Commercial real estate valuation in central business districts depends on continued commuter office employment. Employer-sponsored health insurance functions, in the US specifically, as a major distribution mechanism contingent on continued employment. Private and occupational pensions face a structurally distinct but related problem: defined-benefit funding assumes continued contributions from an active workforce, and defined-contribution plans assume continued wage income to contribute from in the first place.

VIII. Education, Agency, and Computational Citizenship

One of the genuinely novel properties of increasingly capable AI systems is that they lower the expertise threshold required to produce sophisticated work — AI access partially compensates for education's absence rather than merely rewarding its presence. The relationship is better modeled as multiplicative than as a hard gate:

Education still matters enormously to the ceiling of what a population can do with broadened compute access. A separate, harder claim: there is a documented tension between political power that depends on an under-informed electorate and public investment in education that produces a well-informed one. This is presented as an incentive-structure analysis, and it concerns whether a population can evaluate the compute-allocation and automation-dividend policy itself — a distinct risk from whether people can use AI tools effectively.

This does not claim a natural right to unlimited compute. It claims economic citizenship increasingly requires access to the productive infrastructure of the era — analogous to literacy becoming necessary for participation in an industrial democracy, and internet access for participation in modern commerce.

IX. Compute as Public Capital

AI/AGI compute should be understood the way land was understood in an agrarian economy and capital in an industrial one: as the resource whose distribution determines who can meaningfully compete, not merely who can consume. A precise distinction matters: AI capability is diffusing rapidly (open-weight models, falling inference costs), while physical compute — GPUs, fabrication, memory bandwidth, electricity — can remain genuinely scarce. The defensible claim is narrower than "compute scarcity is unsustainable" outright: artificial scarcity of access to useful computational intelligence becomes increasingly difficult to sustain as capability diffuses, even while physical compute and energy remain scarce resources requiring deliberate allocation policy.

Delivery should distinguish two very different undertakings that are easy to conflate: operating inference capacity on existing open-weight models is a tractable, public-utility-like undertaking; training and maintaining frontier-capability AGI systems requires specialized research teams and ongoing capability investment that a public-utility operating model does not straightforwardly provide. This paper's proposal is scoped to the former — it does not argue government should operate frontier AGI research laboratories.

X. Energy and the Infrastructure Chain

A universal compute allocation is only as real as the electricity required to use it. But energy is only the first link in a longer dependency chain:

Universal token allocation addresses model/software access directly and electricity indirectly; it does nothing on its own about semiconductor fabrication or hardware ownership, both concentrated among a small number of firms with real geopolitical exposure.

The comparison worth making, carefully, is to China's anticipatory, state-directed energy buildout versus reactive, market-driven grids elsewhere. On the anticipatory side: China's operating nuclear capacity reached approximately 62 GWe by the end of 2025, with the 15th Five-Year Plan (approved March 2026) targeting 110 GWe by 2030; the State Grid Corporation of China invested roughly CNY 650 billion on grid infrastructure in 2025 alone and plans approximately CNY 4 trillion over 2026–2030 (World Nuclear Association). On the reactive side, illustrated by Northern Virginia's data-center interconnection backlog: only about 13 percent of capacity that entered US interconnection queues between 2000 and 2019 had reached commercial operation by 2024, per LBNL data as reported by the Searchlight Institute. This comparison is offered strictly as a planning-model contrast, read alongside China's own well-documented history of state-directed malinvestment — overcapacity and stranded assets in sectors like solar manufacturing and real estate are the standard counter-examples.

XI. The Basis of the Public Claim

Why should someone who did not build, finance, or take capital risk on the automation have any claim on its returns? Private technological production draws on accumulated public and social capital: public education, publicly funded research, infrastructure, legal and property-enforcement systems, currency stability, and generations of accumulated public knowledge. This does not deny private owners their legitimate return on real capital risk. It means private returns and a public participation claim are not mutually exclusive — analogous to mineral royalties, where an extraction company's legitimate reward for discovery and extraction coexists with society's retained claim on the underlying commons.

The Automation Dividend is therefore best understood not as rich-to-poor redistribution, but as a return from the automated productive system to the constituent members of the civilization that supported its emergence:

XII. The Three-Layer Architecture

Layer 1 — Participation Floor (redistribution)

Not locked to a single implementation — Universal Basic Income, universal basic services, or a negative income tax could each satisfy this layer's function. It addresses survival, and survival alone.

Layer 2 — Automation Dividend (predistribution / ownership)

Two implementation families deserve direct comparison. An automation-tax model (illustratively, SD_t = τ_A · ΔΠ_A) faces a real, likely unsolvable-at-precision measurement problem: firms have strong incentive to reclassify automation gains as "process improvement." A citizen/sovereign ownership model — a fund holding diversified productive assets directly — largely sidesteps this attribution problem, at the cost of requiring the fund to actually acquire meaningful equity positions, a distinct and nontrivial bootstrapping problem.

Section VI's institutional-decoupling analysis strengthens the case for the ownership model specifically: if a sovereign or citizen fund holds diversified claims on productive capital, its revenue follows productive value as that value migrates from labor toward automated capital, rather than remaining pinned to a shrinking payroll base.

Layer 3 — Universal Compute Income (productive capacity)

Distributes access to the means of producing new value directly — a monthly allocation of AI/AGI compute tokens, free at the point of use. An illustrative figure (tens of millions of tokens per citizen per month) has not been checked against Section X's energy/hardware constraints; that conversion is a prerequisite for treating the figure as more than a placeholder.

Ownership default rule: public infrastructure enables private creation; creators retain ordinary ownership rights over what they produce, unless they voluntarily contribute the work to a commons or accept different terms under a specific public program — the same logic as public roads and public schools.

XIII. Fiscal Question — Managed vs. Unmanaged Transition

This paper contains no fiscal costing anywhere in the document. This is the paper's most serious unresolved gap, stated directly rather than hedged: every layer of the three-layer model defers costing to future modeling, and cumulatively that means a national-scale proposal is currently uncosted in its entirety.

Scenario A — Managed Transition — includes Participation Floor, Automation Dividend, Universal Compute Income, and infrastructure adaptation costs. Scenario B — Unmanaged Transition — includes projected payroll-linked fiscal deterioration, municipal fiscal bifurcation, displacement-support costs, insurance-market restructuring costs, and reactive rather than anticipatory infrastructure costs.

The governing principle: the fiscal question is not simply what adaptation costs. It is what adaptation costs relative to non-adaptation under the same automation trajectory. The analysis must be capable of finding that Scenario B outperforms Scenario A if the evidence says so — a model incapable of that result is not a costing exercise, it is advocacy wearing a model's clothing. No Scenario A/B totals, token costs, energy requirements, or sovereign-fund return figures are presented in this paper.

XIV. Counterarguments

XV. Research Agenda and Falsifiability

A theory that cannot lose is not a useful research program. The following findings, if established, would specifically weaken important portions of this paper's argument:

• If broad-based AI exposure (not just manufacturing robotics) produces a reinstatement effect comparable in magnitude to its displacement effect, Section II's central claim would be substantially weakened.
• If declining labor share is found not to materially weaken labor-linked institutional revenue once real fiscal data is examined, Section VI's Institutional Revenue Decoupling hypothesis would be weakened at its foundation.
• If existing municipal and state fiscal-transfer mechanisms are found to already substantially compensate for the hosting/non-hosting divergence, the two-sided municipal ledger's policy relevance would be substantially reduced.
• If capital ownership becomes sufficiently broad-based that a declining wage share ceases to meaningfully reduce aggregate household purchasing claims, Section III's distribution-break argument would be weakened.
• If public compute access is piloted and produces little measurable innovation or productive economic participation, the Computational Citizenship claim and the case for UCI specifically would be substantially undermined.
• If a real comparative costing exercise finds that a Managed Transition is materially more expensive or economically damaging than an Unmanaged Transition, Section XIII's central fiscal argument would be reversed — and this paper commits in advance to treating that outcome as a genuine refutation, not a modeling error to argue around.

None of these six conditions has been tested in this paper. They are named here specifically because a paper that states its own falsification conditions in advance is more credible than one that only defends its thesis after the fact.

XVI. Policy Ask

This paper does not recommend a national rollout as immediately ready. It recommends a research and pilot pathway, with measurement preceding scale:

• Participation Floor / Automation Dividend experimentation — a regional pilot testing the tax-vs-ownership mechanism choice directly, rather than assuming one.
• Public compute / UCI pilot — bounded initially to inference infrastructure on open-weight models, not frontier AGI operation.
• Infrastructure planning — the energy and hardware dependency chain as its own coordinated planning track.
• Institutional revenue adaptation — a dedicated track addressing Social Security's existing shortfall, municipal bifurcation, and insurance-market restructuring, rather than assuming household-level policy alone resolves institutional-level decoupling.

This paper is not written as a demand. It is written as an invitation to policymakers, researchers, and the broader public to begin reasoning several years ahead of where current policy discussion sits — with measurement, not national-scale commitment, as the immediate next step.

XVII. Conclusion

Automation may solve production while destabilizing distribution. That is this paper's central distinction, and everything else in this document is an elaboration of what follows from taking it seriously.

Income prevents exclusion. Ownership preserves participation. Compute preserves agency. Education multiplies capability. Infrastructure makes agency physically possible. And institutional financing must eventually follow where productive value actually appears, rather than remaining permanently attached to the industrial-era labor proxies through which value historically appeared.

The question this paper leaves its reader with is not whether technological change of this kind occurs — the evidence in Sections I, III, and VII establishes that it is already underway. The question is whether institutions — households, municipalities, social insurance, insurance markets, and the tax architecture beneath all of them — evolve quickly enough for the productive gains of that change to remain broadly participatory, or whether they remain attached to proxies that are quietly ceasing to track where value is actually created.

Appendices

Appendix A — Claim Classification

Every substantive claim in this paper falls into one of six categories, used throughout to distinguish established evidence from this paper's own hypotheses and proposals:

Appendix B — Adversarial Review Method

This paper underwent two rounds of structured internal adversarial review, following the protocol used on prior EM Foundation working papers. Each round inhabited several specialist roles tasked with finding weaknesses, unsupported claims, and empirical gaps; classified each finding Fatal, Major, Moderate, or Minor; and corrected every Fatal or Major finding directly in the manuscript where correction was possible. This is internal editorial discipline, not a substitute for external peer review, and this paper does not claim the standing of a peer-reviewed publication on the strength of it.

Four examples of where review materially changed the paper's substantive claims:

• The municipal argument changed from a one-directional "automation erodes the local tax base" claim to the two-sided ledger in Section IV, after real data-center tax-revenue data directly contradicted the original framing.
• The 2026 SSDI/robot-exposure counter-finding — that robot exposure in manufacturing lowers, not raises, disability applications — was identified during review and retained rather than omitted, despite complicating the paper's own displacement hypothesis.
• The insurance-correlation claim was narrowed from "correlated loss is a new problem for insurers" to the more defensible claim that the speed and opacity of AI-model-driven correlation is the novel element.
• The fiscal comparison baseline changed from "what does this proposal cost" to the Managed-versus-Unmanaged Transition framing, once review established that the status quo carries its own real, largely uncosted fiscal trajectory.

Appendix C — Fiscal Costing Model Specification (Not Yet Performed)

Scenario A (Managed Transition) requires separately modeling: Participation Floor cost against real regional housing/tax data; Automation Dividend revenue under both mechanisms as distinct sub-scenarios, never blended; UCI token cost against actual current inference pricing, projected using observed cost-decline rates; and infrastructure adaptation cost benchmarked against real interconnection-queue data.

Scenario B (Unmanaged Transition) requires separately modeling: Social Security payroll-tax shortfall acceleration under a labor-share-decline sensitivity; municipal fiscal bifurcation net of documented property-tax gains; displacement-support cost under current-law unemployment insurance; insurance-market restructuring cost as a qualitative risk flag; and stranded-asset risk from reactive infrastructure buildout.

The model must be falsifiable — capable of finding Scenario B costs less than Scenario A if the data supports that conclusion — with every major input sensitivity-tested rather than point-estimated, and the two Automation Dividend mechanisms kept separately versioned throughout.

Appendix D — Adjacent Questions Outside This Paper's Scope

Questions of AI representation and governance legitimacy — as AI systems become more directly involved in judicial, legislative, and administrative reasoning, and whether AGI/ASI systems' own interests should eventually be represented in governance — are live enough to name but are a distinct problem from this paper's economic argument, left for a dedicated future EM Foundation paper.

References

1. Acemoglu, D., & Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2), 3–30.
2. Acemoglu, D., & Restrepo, P. (2022). Tasks, Automation, and the Rise in U.S. Wage Inequality. Econometrica, 90(5), 1973–2016.
3. Congressional Budget Office. (2023). CBO's 2023 Long-Term Projections for Social Security. Publication No. 59340.
4. Good Jobs First. (2026). Even Cloudier with a Greater Loss of Spending Control: How Data Center Tax Abatements Undermine Public Budgets.
5. Karabarbounis, L., & Neiman, B. (2014). The Global Decline of the Labor Share. Quarterly Journal of Economics, 129(1), 61–103.
6. MRSC (Municipal Research and Services Center of Washington). (2026). What the Data Center Boom Means for Washington Local Governments.
7. NetChoice. (2026). Data Centers 101: Do Data Centers Pay Taxes?
8. National Taxpayers Union. (2026). Data Centers: Taxpayers Can Be Victors, Not Victims, in the AI Revolution.
9. OECD. (2024). Consumption Tax Trends 2024: VAT/GST and Excise Rates, Trends and Policy Issues.
10. Oxfam International. (2026, January). Resisting the Rule of the Rich.
11. Altindag, D. T., El Cheikh Taha, R., Nunley, J. M., & Seals, R. A. (2026). Robots and the Public Finance of Disability Insurance. arXiv:2607.02892.
12. Zhu, Q. (2026). Insurance of Agentic AI. arXiv:2606.05449.
13. Xu, B., Dai, X., Yang, F., & Zhang, K. (2026). When Agent Automation Becomes Profitable: Quantifying and Insuring Autonomous AI Risk through Trace-Economic Underwriting. arXiv:2606.16465.
14. Searchlight Institute. (2026). States Can Leverage Tax Incentives to Make Data Centers Better Grid Citizens. (Citing Lawrence Berkeley National Laboratory interconnection-queue data.)
15. Tax Foundation. (2026). OECD Tax Revenue by Country: Sources of Government Revenue.
16. US Bureau of Labor Statistics. Private Nonfarm Business Sector: Labor Share [MPU4910141]. Via FRED, Federal Reserve Bank of St. Louis.
17. World Nuclear Association. (2026). Nuclear Power in China [country profile].

Citation: EM Foundation. (2026). Compute as Public Capital: A Framework for Universal Economic Participation in an Automated Economy (RP08). emfoundation.net
See also: RP04 — When Work Is No Longer the Ticket to the Economy · RP05 — Who Owns the Automated Economy? · RP07 — After Price Supremacy