Democracy for Sale | How Big Tech Could Capture the Economy, Drain Public Resources, and Put Democracy Under Private Control

Abstract

This paper argues that the current U.S. AI buildout should be understood as a machine-first political economy. The relevant problem is not artificial intelligence in the abstract. It is a buildout model in which dominant AI and cloud firms are consolidating control over the compute layer while public institutions are being asked to reorganize law, infrastructure, subsidy, and finance around the needs of machine-centered production. In economic effect, that model is anti-human: it reallocates public resources, public accommodation, and long-duration savings toward concentrated private infrastructure faster than it secures worker protection, household income stability, local control, or meaningful democratic bargaining power.

The factual record is already substantial. The Federal Trade Commission found that major cloud-AI partnerships include equity and revenue-sharing rights, consultation, control, and exclusivity provisions, cloud-spend commitments, access to training data, and chip co-development, and warned that these arrangements may affect access to computing resources and engineering talent, increase switching costs, and expose sensitive information to dominant cloud partners. The White House’s March 2026 legislative recommendations call for a “minimally burdensome” national framework and for Congress to preempt state AI laws that impose “undue burdens.” DOE and Berkeley Lab reported that data centers used about 4.4% of U.S. electricity in 2023 and could rise to roughly 6.7% to 12% by 2028, with total data-center electricity use projected to reach 325 to 580 TWh by 2028. GAO separately reported that generative AI uses significant energy and water resources while companies generally do not disclose the details.

The financing structure deepens the danger because the buildout no longer stops at public power, public water, and public subsidy; it extends into the savings architecture of the country. Total U.S. retirement assets reached $49.1 trillion at year-end 2025. Brookfield launched a $100 billion AI infrastructure program spanning energy, land, data centers, and compute, anchored by a fund targeting $10 billion in equity commitments. Reuters reported on June 10, 2026 that Morgan Stanley expects AI-related global debt issuance to exceed $570 billion in 2026. These facts do not establish that all retirement savings are being redirected into AI. They do establish that AI infrastructure is already being financed in forms capable of tying pensions, infrastructure funds, insurers, and credit markets to its continuation.

The resulting danger is best described as oligarchic drift through asymmetric adaptation: public systems are being adapted rapidly to the needs of compute and capital while human beings are left to adapt privately, locally, and belatedly to the consequences. Formal democratic institutions may remain intact under those conditions. The narrower and more operational risk is that practical control over infrastructure, industrial policy, and the terms of economic life shifts toward a small private bloc because public institutions become increasingly dependent on that bloc for compute, infrastructure buildout, financing continuity, and policy tempo. The unresolved question is no longer whether AI will be powerful. It is whether dominant firms will be permitted to convert technical dominance into durable political privilege while the public absorbs the burdens, the risks, and an expanding share of the dependence required to sustain it.

Executive Summary

This paper addresses a political-economic question, not a generalized fear of innovation. The issue is whether the United States is already reorganizing law, infrastructure, subsidy, and long-duration capital around AI infrastructure faster than it is specifying how workers, households, communities, and democratic institutions will remain secure, empowered, and meaningfully included in the order being built. The central finding is that the current model creates a credible risk of oligarchic drift because it combines concentrated control over core inputs, increasing public dependency on the sector for competitiveness and security, outward transfer of fiscal and infrastructure burdens, pressure to narrow distributed oversight, and financial embedding through pension-linked and institutional capital. That combination makes the buildout machine-first in effect: public institutions are adapting themselves to the needs of compute and capital before they have secured a comparably serious settlement for people.

The concentration record is already severe. FTC staff found that the leading cloud-AI partnerships include equity and revenue-sharing rights, consultation, control, and exclusivity provisions, cloud-spend commitments, access to training data, and chip co-development, while warning that these arrangements may increase switching costs and affect access to computing resources and engineering talent. The buildout therefore does not rest on a dispersed market with ordinary substitutability. It rests on a small number of firms occupying chokepoint positions at the infrastructure layer and then building commercial and policy leverage on top of those positions.

The policy environment is reinforcing that structure rather than constraining it. The White House’s March 2026 legislative recommendations urge Congress to create a “minimally burdensome” national framework and to preempt state AI laws that impose “undue burdens.” The same document states that American workers “must benefit” from AI-driven growth rather than merely from its outputs. Taken together, those positions reveal an asymmetry in the current policy posture: the legal architecture for faster buildout and reduced friction is being specified more concretely than the social architecture for protecting workers, households, and communities from the costs of transition. In institutional terms, the state is being asked to solve for deployment first and for human stability later.

The public-system burden is no longer speculative. DOE and Berkeley Lab reported that data centers used about 4.4% of U.S. electricity in 2023 and could rise to roughly 6.7% to 12% by 2028, with total data-center electricity use projected to reach 325 to 580 TWh by 2028. GAO separately reported that generative AI uses significant energy and water resources while companies generally do not disclose the details. Brookings has reported that large data centers can use around 5 million gallons of water per day, roughly the needs of a town of up to 50,000 residents. These facts support a finding that the buildout is increasingly reliant on public electricity systems, public water systems, and local resource tolerance even where durable local human return remains weak or unresolved.

The fiscal and financial architecture deepens the lock-in. Public records show states already rethinking whether the public return justifies the concessions. At the same time, retirement assets reached $49.1 trillion at year-end 2025, and Brookfield’s AI infrastructure platform demonstrates that energy, land, data centers, and compute are already being packaged as an institutional asset class. Reuters’ reporting on projected AI-related debt issuance above $570 billion in 2026 shows that the buildout is also increasingly embedded in major credit markets. The significance of these facts is not that all household savings are being redirected into AI. It is that AI infrastructure is being embedded in the mandates, expected returns, and balance sheets of pensions, infrastructure funds, insurers, and credit markets in ways that raise the cost of later democratic correction. The public is being positioned not only as host and ratepayer, but increasingly as involuntary financial participant in a system whose continuation becomes harder to challenge because so many balance sheets have been tied to it.

The labor side of the transition is not specified at remotely comparable scale. The White House’s own framework says workers “must benefit” from AI-driven growth, but that remains a policy objective rather than a defined replacement-income architecture. The country has already learned how to finance the machine answer. It has not learned, at anything like the same scale, how it will replace household income, bargaining power, or local human return where AI deepening weakens labor demand or thins the number of stable livelihoods directly supported. That asymmetry is one of the clearest reasons the buildout should be understood as machine-first in effect.

The formal judgment of this paper is limited, specific, and severe. The evidence does not justify a universal claim that AI as such is illegitimate. It does justify a finding that the current U.S. AI buildout model is anti-human in economic effect because it institutionalizes asymmetric adaptation: public systems are being rapidly adapted to the needs of compute and concentrated capital while human beings are left to absorb the transition without comparably concrete guarantees of protection, bargaining power, meaningful inclusion, or replacement income. If that model continues without stronger structural limits, stricter cost allocation, fuller disclosure, preserved democratic veto points, and enforceable proof that workers and households will capture a proportionate share of the gains, the United States will not need to lose elections to lose practical control over part of its economic future. It will have financed, housed, powered, watered, and increasingly capitalized a privately governed strategic regime before proving that the public remains first in the bargain. That is the formal judgment of this paper.

I. Introduction: The Human Priority Problem

The central question is no longer whether artificial intelligence will matter. The capital has moved. The infrastructure is being built. The law is being rewritten. The relevant question now is who the country is reorganizing itself for. This paper argues that the United States is beginning to reorganize public life around machine infrastructure faster than it is specifying how human beings will remain economically secure, politically first, and meaningfully included in the order being built. That is the problem addressed here. Not AI in the abstract. A planning regime in which power systems, water systems, fiscal concessions, legal standards, and long-duration capital are being adapted with extraordinary speed to the requirements of concentrated compute, while the conditions of human stability remain comparatively underdefined, underfunded, and unsecured.

The machine side of the transition is already concrete. The FTC found that major cloud-AI partnerships include equity and revenue-sharing rights, consultation, control, and exclusivity provisions, cloud-spend commitments, access to training data, and chip co-development, while warning that these arrangements may increase switching costs and affect access to key inputs such as computing resources and engineering talent. Synergy Research’s late-2025 market data place Amazon, Microsoft, and Google at 28%, 21%, and 14% of worldwide cloud-infrastructure spending, respectively, indicating that frontier AI is being built on a substrate already dominated by a small number of firms. Federal policy is moving in parallel: the White House’s March 2026 legislative recommendations call for a “minimally burdensome” national framework and for Congress to preempt state AI laws that impose “undue burdens.” Together, these facts show a concentrated sector and a federal policy posture moving in the same direction—toward faster buildout, lower friction, and weaker competing centers of control.

The physical and fiscal dimensions of that shift are no longer speculative. DOE and Berkeley Lab report that data centers used about 4.4% of total U.S. electricity in 2023 and could rise to roughly 6.7% to 12% by 2028, with total data-center electricity use projected to reach 325 to 580 TWh by 2028. GAO reports that generative AI uses significant energy and water resources while companies generally do not disclose the details. Brookings reports that a large data center can use an estimated 5 million gallons of water per day, equivalent to the needs of a town of up to 50,000 residents. State records show that the fiscal side is also material: Georgia’s official evaluation found that roughly 70% of recent data-center construction likely would have occurred without the tax break, and Ohio suspended new applicants after reported tax-expenditure costs rose to about $554 million in 2024 and nearly $1.6 billion in 2025. These facts show that the buildout is not merely advancing through private risk capital. It is already leaning on public electricity systems, public water systems, and public fiscal capacity.

The labor side of the transition is not specified at remotely comparable scale. Reuters/Ipsos reported on June 10, 2026 that 53% of Americans fear AI could put themselves or someone in their household out of work, and 73% are uneasy about AI’s growing role. The White House’s own framework says American workers “must benefit” from AI-driven growth, but that statement remains an objective rather than a defined replacement-income architecture. The St. Louis Fed found that AI-related investment categories contributed 0.97 percentage points to real GDP growth in the first three quarters of 2025, or about 39% of total GDP growth over that period, showing that AI-related capital deepening is already materially shaping macroeconomic outcomes. Yet the FY 2027 federal budget requests $9.9 billion in discretionary budget authority for the entire Department of Labor. The problem is therefore not simply that jobs may be lost. It is that the country has learned how to finance the machine answer at industrial scale while it has not built a comparably serious answer for household income security.

This asymmetry extends into finance. Total U.S. retirement assets reached $49.1 trillion at year-end 2025. Brookfield launched a $100 billion AI infrastructure program spanning energy, land, data centers, and compute, anchored by a fund targeting $10 billion in equity commitments and already securing $5 billion. S&P Global reported that private equity accounted for 72% of U.S. data-center investment value in 2025. Reuters reported that Morgan Stanley expects AI-related global debt issuance to exceed $570 billion in 2026. These facts do not establish that all household savings are being redirected into AI. They do establish that AI infrastructure is already being embedded in forms capable of tying pensions, infrastructure funds, insurers, and credit markets to its continuation. Once that process matures, the public is no longer engaged only as citizen, worker, ratepayer, or local resident. It is increasingly engaged as an involuntary financial participant in a system whose continuation becomes harder to challenge because so many balance sheets have been tied to it.

The argument that follows proceeds from a limited but serious proposition. The current AI buildout is not merely concentrated and risky; it is institutionalizing asymmetric adaptation. Public systems are being rapidly adapted to the needs of compute and concentrated capital while human beings are left to absorb the transition without comparably concrete guarantees of protection, bargaining power, local control, or replacement income. In that sense, the present buildout is anti-human in economic effect not because it announces hostility to people, but because it systematically ranks machine infrastructure, capital formation, and deployment speed ahead of the conditions required for a broadly human future. The task of this paper is to define that structure, identify the revealed priorities through which it is advancing, and determine whether democratic institutions still retain the practical capacity to impose terms before dependence hardens into governing fact.

II. Governing Framework: Asymmetric Adaptation and the Human Priority Standard

This paper uses the term asymmetric adaptation to describe a specific political-economic condition: institutions adapt law, infrastructure, subsidy, and finance to the needs of machine infrastructure faster than they adapt protections for the people expected to live through the transition. The key analytic claim is not that AI leaders have openly declared a reduced role for human beings. It is that the planning regime is already visible in the allocations. Capital is moving at industrial scale. Electricity systems are being reoriented. Water burdens are being accepted. State friction is being treated as a policy problem. Long-duration finance is being recruited. The human side remains comparatively vague. That is a revealed-priority argument, not a motive claim.

Within this framework, a buildout becomes anti-human in economic effect when four conditions converge. First, household welfare is subordinated to infrastructure expansion, as when electricity, water, land, or tax expenditures are reorganized around hyperscale facilities before public burdens are clearly assigned back to those facilities. Second, labor is subordinated to capital deepening, as when the financing, permitting, and protection of machine infrastructure are specified in detail while replacement-income architecture for households remains thin or undefined. Third, community consent is subordinated to speed and preemption, as when state or local veto points are treated as “undue burdens” on a nationally preferred buildout. Fourth, democratic correction is subordinated to financial lock-in, as when the buildout becomes increasingly embedded in pensions, infrastructure funds, insurers, utilities, and credit markets, raising the cost of later restraint.

The framework therefore rejects two weaker approaches. It rejects the view that the issue is merely one of market concentration, because concentration alone does not explain how public systems and household exposure become part of the same buildout. It also rejects the view that the issue is purely constitutional in the narrow sense, because the decisive question is not simply whether formal institutions remain in place. The decisive question is whether practical bargaining power remains with the public once capital, infrastructure, and finance have been reorganized around the continuity requirements of a concentrated private sector. In that sense, oligarchic drift is not defined here as the abolition of democracy. It is defined as the reduction of practical public sovereignty over the terms of economic life under conditions of concentration, dependency, cost transfer, and financial embedding.

To make that judgment operational, the paper applies a Human Priority Standard. Under this standard, the buildout must be evaluated not only by output, investment, or national competitiveness claims, but by whether ordinary people remain first in the bargain. The standard asks four questions. Does the buildout leave workers with more secure income and bargaining power, or with less? Does it reduce household exposure to utility, tax, and affordability burdens, or increase it? Does it expand local and state capacity to shape the conditions of buildout, or narrow it? Does it keep long-duration household savings at arm’s length unless broad public benefit is demonstrated, or does it increasingly draw those savings into the financing machine? If the answer to these questions is negative or unresolved, then the burden of proof shifts to those demanding faster deployment, broader subsidy, lighter oversight, or wider preemption.

The Human Priority Standard also clarifies what this paper is not claiming. It does not claim that every AI deployment is socially harmful, that every data center is economically irrational, or that every infrastructure investment linked to AI is improper. It does claim that the present model has crossed a threshold at which private firms are no longer merely selling tools. They are building a strategic layer whose continuation increasingly depends on public systems and public forbearance. Once that threshold is crossed, proof of technological promise is no longer sufficient. The buildout must also prove proportionate human return. On the present record, that proof remains incomplete.

The sections that follow use this framework to test whether the United States is still planning around human beings as the organizing subject of development, or whether it is already teaching its institutions to plan around machine infrastructure first. The analysis proceeds by examining revealed priorities in capital mobilization, infrastructure burden, labor asymmetry, financial lock-in, and distributed democratic control. The governing question throughout is the same: who is being planned for first, and who is being asked to adapt later?

III. Revealed Priorities: Evidence the Planning Regime Is Already Underway

The central analytical claim of this paper is that the current AI buildout should be evaluated through revealed priorities rather than declared intent. The decisive question is not whether policymakers or industry leaders have openly stated that human beings will matter less in an AI-centered economy. The decisive question is whether enough capital, infrastructure, legal adaptation, fiscal accommodation, and financial exposure have already been reallocated toward the buildout to show that institutional planning is proceeding faster than any comparably serious plan to preserve household income, worker bargaining power, local control, and durable human return. On the present record, that threshold has already been crossed.

The first revealed-priority threshold is capital mobilization. A sector is no longer operating at the level of experimentation when its financing requirements become large enough to reshape balance sheets, debt markets, and strategic planning. Reuters reported on June 10, 2026 that Morgan Stanley expects AI-related global debt issuance to exceed $570 billion in 2026, and on the same day Reuters reported that the largest technology firms’ AI and cloud spending is expected to approach $700 billion this year. Reuters also reported that Amazon secured a $17.5 billion loan facility amid its AI-driven capital-expenditure ramp, while Oracle told investors it plans $70 billion in net capital expenditures for fiscal 2026 and intends to raise $40 billion through debt and equity. These are not the financing patterns of a tentative emerging sector. They are the financing patterns of an industrial buildout whose scale already exceeds the ordinary boundaries of corporate experimentation.

The second threshold is infrastructure allocation. DOE and Berkeley Lab report that data centers used about 4.4% of total U.S. electricity in 2023 and could rise to roughly 6.7% to 12% by 2028, with total data-center electricity use projected to increase from 176 TWh in 2023 to between 325 and 580 TWh by 2028. Those numbers establish that the AI buildout is not a marginal load issue. It is already large enough to reshape utility planning, transmission needs, reserve-margin assumptions, and regional siting conflicts. The same is true of water. Brookings reports that a large data center can use an estimated 5 million gallons of water per day, roughly the needs of a town of up to 50,000 residents. Once a sector reaches this level of physical demand, the relevant issue is no longer whether it is innovative. The relevant issue is whether public systems are being reorganized around it.

The third threshold is policy adaptation. The White House’s March 2026 legislative recommendations call for a “minimally burdensome” national framework and urge Congress to preempt state AI laws that impose “undue burdens.” That same framework treats a fragmented state-law landscape as a threat to national competitiveness. In practical terms, this means the legal order is being asked to adapt itself to the buildout’s need for speed and uniformity before the public has secured comparably concrete guarantees on cost allocation, labor protection, or local consent. A planning regime is already underway when legal friction is being cleared for machine infrastructure more quickly than social protection is being specified for the people expected to live through the transition.

The fourth threshold is fiscal and public-system accommodation. State records now show that the public is not simply observing the buildout; it is increasingly being asked to carry part of it. Georgia’s official evaluation found that roughly 70% of recent data-center construction likely would have occurred without the tax break. Ohio suspended new applicants after reported tax-expenditure costs rose to about $554 million in 2024 and nearly $1.6 billion in 2025. Those facts do not prove that every project lacks value. They do show that public subsidy, public utility tolerance, and public fiscal capacity are already being mobilized as inputs into the buildout. Once that occurs, the burden of proof should shift. The relevant question is no longer whether a project is privately profitable. It is whether the public return is proportionate to the public burden.

The fifth threshold is financial embedding. Total U.S. retirement assets reached $49.1 trillion at year-end 2025. Brookfield’s AI infrastructure program spans energy, land, data centers, and compute, and was launched with a fund targeting $10 billion in equity commitments and already securing $5 billion. S&P Global reported that private equity accounted for $45.7 billion, or 72%, of total U.S. data-center investment value in 2025. These facts do not establish that household savings are being centrally directed into AI. They do establish that the buildout is being packaged in forms capable of recruiting pensions, insurers, infrastructure funds, and credit markets at scale. When that threshold is crossed, opposition becomes more expensive not only politically but financially, because the continuation of the buildout becomes linked to the expected returns and balance-sheet assumptions of a widening set of institutions.

The sixth threshold is human-transition underdevelopment. A revealed-priority analysis must compare not only what is being built, but what is missing. The White House’s own framework states that American workers “must benefit” from AI-driven growth rather than merely from its outputs. That language is significant because it describes a requirement, not an accomplished fact. Reuters/Ipsos reported on June 10, 2026 that 53% of Americans fear AI could put themselves or someone in their household out of work, and 73% are uneasy about AI’s growing role. Yet the institutional answer to that anxiety remains comparatively thin. The President’s FY 2027 budget requests $9.9 billion in discretionary budget authority for the entire Department of Labor, while Reuters reports that AI-related debt issuance alone is expected to exceed $570 billion in 2026. The mismatch is not subtle. The machine side of the transition is specified in debt facilities, capex plans, power projections, and legal recommendations. The human side is still framed largely in terms of future benefit, study, and adaptation.

Taken together, these thresholds support a more exact conclusion than general warnings about technology or even general warnings about concentration. They show that the United States is already reallocating capital, infrastructure, law, and public-system tolerance toward AI infrastructure at a speed and scale that materially outpace the specificity of its plans for worker security, household income stability, local consent, and democratic control. This is why the present paper treats the buildout as a revealed-priority regime. The system is already showing what it is prepared to finance, what it is prepared to protect, and what it expects the public to absorb.

What follows from that finding is not yet a claim that democratic institutions have failed in formal terms. It is a narrower and more operational judgment: the country is already teaching its institutions to solve for machine continuity before they have solved for human continuity. That is the point at which concentration becomes governance, governance becomes dependence, and dependence becomes the material basis of oligarchic drift.

IV. Resource Burden and Human Return

A machine-first buildout should not be evaluated only by its scale, novelty, or contribution to national output. It should also be evaluated by the ratio between the public resources it consumes and the durable human return it delivers. That is the relevant local economic test. A project that requires large and continuing claims on electricity systems, water systems, land use, and public fiscal capacity must show more than technological importance or future optionality. It must show that the communities asked to host it receive a proportionate and durable human benefit in jobs, wages, tax value, or ecosystem development. Where that showing is weak, the buildout begins to look less like development and more like extraction.

The resource side of the ratio is already substantial. Brookings reports that a typical data center uses about 300,000 gallons of water each day, equivalent to the demands of about 1,000 households, while large data centers can use an estimated 5 million gallons per day, equivalent to the needs of a town of up to 50,000 residents. DOE and Berkeley Lab report that data centers used about 4.4% of total U.S. electricity in 2023 and could rise to roughly 6.7% to 12% by 2028. WRI similarly notes that the data-center boom is reshaping local energy grids, water systems, and land use and that the pace of expansion is often occurring with limited public information about long-term impacts and benefits. These facts establish that the AI-era data-center buildout is not light-touch infrastructure. It is resource-intensive infrastructure whose burdens are imposed locally and continuously.

The human-return side of the ratio is more mixed and often less impressive than the promotional narrative suggests. Brookings’ May 2026 analysis of county-level labor-market effects found that counties receiving their first large data center saw total private employment rise by 4% to 5% over five to six years, construction employment jump 11%, and information-sector employment grow 22%, with wages rising by 3% to 4% for both existing workers and new hires. Those are real effects and they matter. But Brookings also found that the employment gains depend heavily on facility type and local scale. Counties with a single data center saw modest total employment effects but no significant information-sector growth, while the largest information-sector gains appeared only in counties with four or more facilities. The same article notes that critics are right to question the long-term job case because large data-center projects often promise only dozens to a few hundred permanent workers and because construction employment is temporary.

That distinction is economically decisive. A county may receive short-run construction activity, modest total job gains, or some information-sector spillovers, especially where multiple facilities cluster. But the permanent operating footprint of a single hyperscale facility can still be thin relative to its continuing claims on power, water, land, and tax concessions. In other words, the right question is not whether data centers create any jobs. The right question is whether the direct and durable human return is proportionate to the public burden being imposed. Brookings’ evidence suggests that this cannot be assumed. It varies by facility type, by whether a true ecosystem develops, and by whether multiple facilities accumulate. The burden of proof should therefore rest on project proponents, not on local communities expected to trust generalized promises of future prosperity.

The same logic applies to fiscal claims. Georgia’s official tax-incentive evaluation found that roughly 70% of recent data-center construction likely would have occurred without the exemption. Ohio suspended new applicants after reported tax-expenditure costs rose to about $554 million in 2024 and nearly $1.6 billion in 2025. These records matter because they show that the public return cannot be inferred from the presence of investment alone. If a large share of construction would have happened anyway, then tax concessions represent less an inducement of new local welfare than a transfer of fiscal capacity away from other public purposes. In that setting, even modest job gains or construction activity may not justify the total package of subsidy, utility accommodation, and community burden.

This is why the paper treats resource burden versus human return as a core analytic test rather than a subsidiary concern. A buildout fails that test when it requires town-scale water, grid-scale electricity, land-use concessions, and material tax expenditures while offering only thin permanent employment and underdefined local spillovers. It also fails when the public case rests on broad future gains rather than demonstrated local human return. WRI’s formulation is especially instructive: whether data-center growth strengthens local economies or instead shifts risks to residents depends on rules governing energy procurement, water use, land siting, community engagement, and cost recovery. In other words, the public value of the buildout is not self-executing. It depends on governance choices that are currently too weak, too opaque, or too accommodating.

The broader implication is that the current model is not simply under pressure because people are uneasy with technological change. It is under pressure because its local economic case is often structurally thin relative to its resource intensity. Where projects demand large and durable claims on shared systems while leaving direct human return limited, uncertain, or dependent on future clustering that may never arrive, the machine-first character of the buildout becomes easier to see. The community is asked to reorganize itself around the facility. The facility is not yet required to prove that it reorganizes itself around the community. That inversion is one of the clearest economic signatures of asymmetric adaptation.

V. The Missing-Income Problem: Why the Machine Answer Is Funded and the Human Answer Is Not

The central economic danger in the current AI buildout is not simply that jobs may change, or even that some jobs may disappear. It is that the United States is already specifying the financing, power demand, legal accommodation, and capital structure for machine expansion at industrial scale while it has not specified, at remotely comparable scale, how households will retain income, bargaining power, and durable economic participation where that same expansion weakens labor demand or thins the number of stable livelihoods directly supported. In that sense, the problem is not “missing money” in the national accounts. It is missing replacement-income architecture for people. The machine answer is already funded. The human answer is not.

The machine side of the ledger is concrete. Reuters reported on June 10, 2026 that Morgan Stanley expects AI-related global debt issuance to exceed $570 billion in 2026, while the largest technology companies’ AI and cloud capital expenditures are expected to approach $700 billion this year. Reuters also reported that Amazon secured a $17.5 billion loan facility tied to its AI-driven capital-expenditure ramp. Brookfield has launched a $100 billion AI infrastructure program spanning energy, land, data centers, and compute, structured specifically for institutional-scale participation. These are not the financial signatures of a tentative technological transition. They are the financial signatures of a buildout whose expansion is being planned, funded, and protected in detail.

The labor side of the ledger is not comparably specified. The White House’s March 2026 legislative recommendations state that American workers “must benefit” from AI-driven growth rather than merely from its outputs and call for expanded federal attention to “task-level workforce realignment” driven by AI. That language is significant precisely because it is aspirational. It describes an outcome that has not yet been secured. The FY 2027 federal budget requests $9.9 billion in discretionary budget authority for the entire Department of Labor. Even without forcing a simplistic one-to-one comparison, the revealed-priority contrast is stark: the country is mobilizing hundreds of billions for machine infrastructure, while the institutional architecture intended to protect or replace household income remains measured in low single-digit billions and framed largely in terms of training, study, and adaptation.

Macroeconomic data make the asymmetry even clearer. The St. Louis Fed found that four AI-related investment categories contributed 0.97 percentage points to real GDP growth in the first three quarters of 2025, amounting to roughly 39% of total GDP growth over that period. That is a meaningful contribution. It shows that AI-related capital deepening is already shaping national growth. But growth contribution is not the same thing as broad-based income security. GDP can rise while labor income is pressured, bargaining power weakens, and local human return remains thin. A buildout that contributes materially to growth while leaving unresolved where equivalent household income will come from for those displaced, downgraded, or bypassed by the same transition does not solve the human problem. It obscures it.

Public anxiety is already tracking that gap. Reuters/Ipsos reported on June 10, 2026 that 53% of Americans fear AI could put themselves or someone in their household out of work, and 73% are uneasy about AI’s growing role. Those numbers do not prove that mass displacement is imminent. They do show that the labor side of the transition is already experienced by the public as a household-security problem. When a population sees the machine answer being financed and accelerated at enormous scale while the human answer remains diffuse and underdefined, the resulting anxiety is not irrational. It is an accurate reading of institutional asymmetry.

Recommended by LinkedIn

The Week Ahead | 22-29 March 2026

The Tech Capital 6 months ago

Time to Break Free from Big Tech!

Stéphane Garneau 1 year ago

Open AIs Pitch to Power Australia

Samantha McLean 1 year ago

The local-development record reinforces the same conclusion. Brookings has shown that large data-center projects can produce real gains in construction employment and, under some conditions, in total county-level employment. But Brookings also notes that many large projects promise only dozens to a few hundred permanent workers, and that the most substantial information-sector gains tend to appear only when multiple facilities cluster and a broader ecosystem develops. At the same time, Brookings reports that a large data center can use an estimated 5 million gallons of water per day, roughly the needs of a town of up to 50,000 residents. This means the buildout can make town-scale claims on shared resources while offering relatively thin long-run employment in the host community unless additional development arrives later. Where projects demand large and durable public burdens without comparably durable human return, the missing-income problem is not theoretical. It is embedded in the local economics of the model itself.

This is why the present paper treats the labor issue as more than a conventional “workforce development” concern. The relevant question is not only whether workers can be retrained. The relevant question is whether the country has a serious plan for income continuity where machine-centered investment expands faster than human-centered compensation, ownership, or transfer mechanisms. On the present record, that plan is not comparably concrete. The buildout has debt facilities, capex schedules, load forecasts, siting fights, water claims, institutional funds, and preemption proposals. It does not yet have an equally specified architecture for wage replacement, bargaining-power preservation, regional income stabilization, or broad ownership claims that would allow ordinary households to remain secure if labor demand is weakened by the same transition.

The policy consequence is straightforward. Once the buildout is seen through the lens of missing-income architecture, the burden of proof shifts. It is no longer sufficient to say that AI will boost productivity, contribute to GDP, or strengthen national competitiveness. Those claims may be true and still leave the central human question unanswered. The relevant burden now falls on those demanding faster deployment, broader subsidy, lighter oversight, or narrower state authority. They should be required to demonstrate, with specificity, where replacement household income will come from if labor demand weakens, how workers will retain bargaining power, how local communities will receive durable human return, and why the public should continue financing and stabilizing a machine-centered regime before the human answer exists at comparable scale. On the present record, that showing has not been made.

VI. Financial Lock-In: How the Buildout Reaches Beyond Power and Water into the Savings Architecture of the Country

The current AI buildout no longer stops at electricity, water, land, and public subsidy. It extends into the balance sheet of the country. That extension matters because it changes the politics of correction. A buildout financed only by founders, retained earnings, or a narrow set of private investors can, in principle, be disciplined without broad social spillover. A buildout financed through institutional vehicles, private equity, credit markets, insurers, utilities, and retirement-linked capital becomes harder to challenge because more balance sheets begin to depend on its continuation. The issue is therefore not whether capital formation is legitimate. It is whether the present capital structure is broadening public dependence faster than it is broadening public control.

The first mechanism is direct institutional packaging. Total U.S. retirement assets reached $49.1 trillion at year-end 2025, equal to 34% of household financial assets. That figure does not mean retirement savings are being centrally directed into AI. It does mean that any infrastructure strategy of sufficient scale will operate within a capital environment shaped by pensions, retirement mandates, insurance liabilities, and long-duration savings. Brookfield’s November 2025 launch of a $100 billion AI infrastructure program is a clear example of how the sector is being structured for that environment. Brookfield said the program would be deployed across the full AI value chain, including energy, land, data centers, and compute, and that the vehicle would be anchored by a dedicated AI infrastructure fund targeting $10 billion in equity commitments. Its later shareholder letter said the fund had already secured $5 billion of commitments. This is not the language of speculative technology finance. It is the language of institutional infrastructure formation.

The second mechanism is private-capital dominance in the asset class itself. S&P Global reported that private equity accounted for $45.7 billion, or 72%, of the $63.35 billion invested in U.S. data centers in 2025. That fact is important not simply because private equity is active. It is important because it shows that AI-era data-center expansion is increasingly being organized through concentrated capital structures whose business model depends on scaling assets, securing favorable infrastructure conditions, and protecting valuation assumptions over time. Once that pattern becomes dominant, the buildout is no longer governed primarily by diffuse market signals. It is increasingly governed by actors whose returns depend on the continuation of rapid expansion under conditions of regulatory and fiscal accommodation.

The third mechanism is debt-market finance. Reuters reported on June 10, 2026 that Morgan Stanley expects AI-related global debt issuance to exceed $570 billion in 2026. Reuters also reported that Amazon secured a $17.5 billion loan facility as part of its AI-driven capital-expenditure ramp, and that major technology companies’ combined AI and cloud spending is expected to approach $700 billion this year. Oracle separately told investors it plans $70 billion in net capital expenditures for fiscal 2026 and intends to raise $40 billion through debt and equity to sustain its AI expansion. These are not isolated financing events. They show that the continuation of the buildout is already being tied to very large flows through credit markets and corporate balance sheets. Once that happens, restraint becomes harder because the question is no longer only whether a firm should be slowed. It becomes whether slowing the buildout disrupts the expectations embedded in lenders, investors, utilities, and a widening set of financial intermediaries.

The fourth mechanism is passive household exposure. Households do not need to make an explicit “AI bet” for the buildout to begin shaping their economic position. If retirement systems, insurers, infrastructure funds, bond markets, and utilities all become increasingly exposed to the continuation of AI infrastructure expansion, then ordinary savers can become indirectly dependent on the same regime even where they never consciously chose it. This is the deeper meaning of financial lock-in. The public is not only asked to host the buildout and absorb its utility burdens. It is increasingly positioned as an involuntary financial participant in a system whose continuity may begin to matter for pension returns, insurance portfolios, fund mandates, municipal expectations, and broader credit conditions. The public becomes entangled not because it voted for the bargain, but because the bargain is being financed through the systems on which it already depends.

The fifth mechanism is macro-financial significance. When a sector begins to influence the wider supply of corporate debt and long-duration capital demand, it becomes harder to treat as an ordinary industry. Reuters’ reporting on Morgan Stanley’s forecast makes clear that AI-related issuance is no longer marginal. It is large enough to matter to credit markets. That scale does not prove that AI is the dominant driver of rates or asset prices. It does prove that the buildout is increasingly embedded in the broader financial environment within which households save, firms borrow, insurers allocate, and governments plan. A sector that reaches that level of financial significance gains an additional shield against discipline because the cost of disruption is no longer borne only by its managers or equity holders. It begins to diffuse outward into the financial architecture itself.

This is why the financial question is inseparable from the democratic question. A buildout that reaches into public systems can be contested through politics. A buildout that also reaches into pensions, infrastructure funds, insurers, private equity, and debt markets becomes harder to contest because democratic correction now threatens a larger coalition of actors with material exposure to its continuation. At that point, the state is no longer simply asked to permit the buildout. It is pressured to stabilize it. That is the point at which financial embedding becomes a governance mechanism. The buildout is no longer defended only as innovation. It is defended as infrastructure, as portfolio strategy, as credit necessity, and as macroeconomic continuity. That is how financial lock-in becomes part of oligarchic drift.

The practical implication is straightforward. The public cannot be said to govern a buildout whose continuation is being written simultaneously into electricity systems, subsidy regimes, and the savings architecture of the country. Once the machine answer is embedded that deeply, saying no becomes more expensive not because the public agreed, but because too many powerful institutions have already learned how to profit from the continuation of the same system. That is the financial core of the present danger.

VII. Practical Sovereignty: How Asymmetric Adaptation Becomes Oligarchic Drift

The central constitutional question raised by the current buildout is not whether elections will formally disappear. It is whether public institutions retain the practical capacity to say no, slow down, impose terms, or reverse course once enough law, infrastructure, subsidy, and finance have been reorganized around the continuity requirements of a concentrated private sector. This paper uses the term practical sovereignty to refer to that capacity. A political order can retain ballots, legislatures, and courts while losing practical sovereignty over major economic decisions if the cost of public resistance becomes sufficiently high. In that setting, democracy remains in form while bargaining power weakens in substance. That is the narrower and more operational danger described here as oligarchic drift.

The mechanism through which that drift occurs is cumulative. The first stage is concentration. The FTC found that major cloud-AI partnerships include equity and revenue-sharing rights, consultation, control, and exclusivity provisions, cloud-spend commitments, access to training data, and chip co-development, while warning that these arrangements may increase switching costs and affect access to computing resources and engineering talent. Once that structure is in place, the public is not dealing with a loosely organized field of competitors. It is dealing with firms occupying chokepoint positions at the infrastructure layer. The second stage is policy reinforcement. The White House’s March 2026 legislative recommendations urge Congress to create a “minimally burdensome” national framework and to preempt state AI laws that impose “undue burdens.” In combination, those two stages mean that market concentration and legal accommodation begin to move together rather than offset one another.

The third stage is public-system dependence. DOE and Berkeley Lab report that data centers used about 4.4% of U.S. electricity in 2023 and could rise to roughly 6.7% to 12% by 2028. Brookings reports that a large data center can use an estimated 5 million gallons of water per day, equivalent to the needs of a town of up to 50,000 residents. Once facilities at that scale become central to utility planning, transmission buildout, land use, and water allocation, public systems begin to adapt themselves around the needs of the buildout. The state is no longer only regulating a sector. It is increasingly provisioning for its continuity. That distinction matters because a public institution that has already reorganized its infrastructure assumptions around the continuation of a sector faces a much higher price for later restraint.

The fourth stage is financial embedding. ICI reports that total U.S. retirement assets reached $49.1 trillion at year-end 2025. Brookfield’s AI infrastructure program spans energy, land, data centers, and compute and is publicly structured as an institutional-scale vehicle. Reuters reported that Morgan Stanley expects AI-related global debt issuance to exceed $570 billion in 2026. Once a buildout is tied not only to corporate strategy but also to pensions, infrastructure funds, insurers, utilities, and credit markets, public correction becomes more difficult because more institutions now have something to lose from the exercise of public authority. At that point, the question confronting policymakers is no longer simply whether a project or company should be slowed. It is whether slowing the buildout disrupts a wider architecture of returns, liabilities, and market expectations.

The fifth stage is the narrowing of democratic veto points. Public institutions do not need to be abolished for this narrowing to occur. They need only to operate under conditions in which the practical price of saying no rises at every level. Legislators confront competitiveness narratives and investment threats. Regulators confront reliability and infrastructure claims. Local governments confront subsidy competition and siting pressure. Fiduciaries confront portfolio exposure. Households confront rate impacts and shrinking alternatives. Under those conditions, public authority remains formally intact but increasingly constrained by the dependence that earlier decisions have already built. This is the point at which practical sovereignty begins to erode. The state still acts, but inside a narrowing range of politically and economically available choices.

This is why the present paper treats the current AI buildout as more than a regulatory challenge or more than a conventional antitrust problem. The issue is whether the country is permitting a sector to become too infrastructurally necessary, too financially embedded, and too legally protected to discipline without systemic pain. If that threshold is crossed, democratic institutions can survive while losing part of their practical authority over the terms of economic life. The public is then left with a thinner form of sovereignty: it may still vote, debate, and protest, but the cost of materially redirecting the buildout has been engineered upward through concentration, dependence, cost transfer, and financial lock-in. That is the institutional logic of oligarchic drift.

The implications of that logic are not only constitutional. They are distributive. A society that loses practical sovereignty over the terms of its strategic infrastructure also loses practical sovereignty over who bears the burdens and who captures the gains. In the present case, the public is already being asked to host the facilities, absorb the utility strain, tolerate the water burden, subsidize the siting, and increasingly finance the continuation of the model. If it cannot later impose terms because too much has already been reorganized around the buildout, then the underlying transfer is no longer merely economic. It becomes political in the deepest sense: the country has taught its institutions to plan around machine continuity before proving that human continuity remains first.

VIII. Conditions of Falsification: What Would Have to Be True for the Machine-First Thesis to Fail

A serious analysis must identify the conditions under which its central thesis would be materially weakened. The argument advanced in this paper is not that machine-first political economy is already total or irreversible. It is that the present U.S. AI buildout has already crossed enough thresholds of concentration, public-system burden, financial embedding, and human under-provision to justify a presumption of democratic and economic risk. That presumption would be substantially weaker if the underlying record showed that the buildout was being governed on terms that left human beings, rather than machine infrastructure, first in allocation, law, and finance. On the present record, that showing has not been made.

The first falsification condition would be full and transparent cost internalization. The machine-first thesis would weaken materially if the record showed that hyperscale AI and data-center operators were consistently bearing the full costs they impose on electricity systems, water systems, transmission buildout, reliability measures, and related public infrastructure. DOE and Berkeley Lab’s projections of sharply rising data-center electricity demand, together with GAO’s finding GAO’s finding that companies generally do not disclose the details of generative AI’s energy and water use, point in the opposite direction: the burden is large, while visibility remains incomplete. So long as the public cannot see the burden clearly and cannot assign it back reliably to the facilities creating it, the presumption that the buildout is socializing material costs remains intact.

The second falsification condition would be robust and durable distributed oversight. The thesis would weaken if state and local governments retained broad practical authority to condition, delay, or refuse projects whose burdens fall on their communities, and if federal policy were neutral or supportive toward those dispersed veto points. The White House’s March 2026 legislative recommendations instead call for a “minimally burdensome” national framework and for Congress to preempt state AI laws that impose “undue burdens.” That posture does not prove that all local authority has vanished. It does show that the legal trajectory is not toward stronger distributed control. A machine-first thesis becomes less plausible when the buildout remains meaningfully governable by the communities asked to host it; it becomes more plausible when those communities are increasingly treated as obstacles to be overcome.

The third falsification condition would be declining concentration and lower switching dependence. The thesis would weaken if frontier AI were clearly moving toward a less concentrated compute structure, if major cloud-model arrangements were losing their governance significance, and if switching costs and access barriers were visibly falling. FTC staff documented the opposite pattern: partnership structures involving equity and revenue-sharing rights, consultation, control, exclusivity, cloud-spend commitments, access to training data, and chip co-development, together with explicit concerns about switching costs and access to key inputs. Where the infrastructure layer remains highly concentrated and the commercial relationships remain sticky, the buildout cannot plausibly be described as one in which market decentralization is likely to solve the problem on its own.

The fourth falsification condition would be high and durable human return relative to public burden. The thesis would weaken if projects making large claims on power, water, land, and tax expenditures also demonstrated strong long-run local benefits in stable employment, wage growth, tax value, and ecosystem development. Brookings’ county-level evidence does show real gains in construction employment and, under some conditions, in broader private employment. But Brookings also shows that the largest and most durable information-sector gains tend to emerge only in counties with multiple facilities and broader ecosystem formation, while large projects often promise only dozens to a few hundred permanent workers. At the same time, Brookings reports that large facilities can use water equivalent to the needs of a town of up to 50,000 residents. Those facts do not falsify the machine-first thesis; they qualify it by showing that human return is possible in some settings but not sufficiently robust or automatic to rebut the broader concern.

The fifth falsification condition would be a serious replacement-income architecture for households. The thesis would weaken substantially if the country had already built a social architecture proportionate to the scale of AI capital mobilization—one capable of protecting household income, worker bargaining power, and regional human return where labor demand is weakened or the number of stable livelihoods directly supported by production is thinned. The current record instead shows the opposite asymmetry. Reuters reports AI-related debt issuance above $570 billion in 2026 and AI/cloud capital expenditures approaching $700 billion this year, while the White House’s own framework still states that workers “must benefit” from AI-driven growth and the FY 2027 budget requests $9.9 billion in discretionary budget authority for the entire Department of Labor. The machine answer is financed with industrial specificity; the human answer remains comparatively modest, aspirational, and underbuilt. That is one of the central reasons the thesis remains standing.

The sixth falsification condition would be limited and fully disclosed financial embedding. The thesis would weaken if retirement-linked and institutional exposure to AI infrastructure remained narrow, transparent, and genuinely optional in ways that prevented broad lock-in. Instead, total U.S. retirement assets reached $49.1 trillion at year-end 2025; Brookfield’s AI infrastructure vehicle spans energy, land, data centers, and compute at institutional scale; and S&P Global reported that private equity accounted for 72% of U.S. data-center investment value in 2025. Those facts do not prove that household savings have been commandeered. They do show that the capital structure of the buildout is already compatible with the widening participation of pensions, insurers, infrastructure funds, and credit markets. As that embedding deepens, the cost of democratic correction rises.

Taken together, these falsification conditions clarify the paper’s standard. The thesis would weaken if the buildout were less concentrated, less subsidized, less legally protected from local friction, less demanding of shared infrastructure, less financially embedded, and more clearly paired with a serious architecture for household income continuity and durable local human return. The present record points in the opposite direction often enough, and at large enough scale, that the machine-first thesis remains warranted. The burden therefore remains with those advocating faster deployment, broader subsidy, lighter oversight, or wider preemption to show that the buildout will remain publicly governable and human-centered in effect. On the evidence now available, that showing has not been made.

IX. Counterarguments and Their Limits

A serious analysis must address the strongest rival positions in their best form. The current buildout has defenders who do not merely deny the costs described in this paper. They argue that those costs are justified by strategic necessity, economic development, capital formation, and the eventual self-correcting power of markets. Some of those arguments identify real considerations. None of them, on the present record, is sufficient to rebut the machine-first thesis. The problem is not that every pro-buildout claim is false. The problem is that the public burden is already concrete while the human guarantee remains conditional, delayed, or undefined.

The first counterargument is geopolitical urgency. On this view, the United States must build quickly because advanced AI capability now bears directly on national competitiveness and strategic power. The White House’s March 2026 recommendations reflect that position explicitly by calling for a “minimally burdensome” national framework and preemption of state laws that impose “undue burdens,” tying that posture to the national strategy for “global AI dominance.” That concern is real, but it does not resolve the central problem identified in this paper. Strategic importance is not a reason to weaken public leverage over a concentrated sector. It is a reason to strengthen it. A country that responds to strategic dependence by lowering friction for dominant firms before securing cost allocation, human protection, and durable public control is not solving a sovereignty problem. It is deepening one.

The second counterargument is local economic development. On this view, data centers create jobs, raise wages, and expand local tax bases, making current concessions and infrastructure accommodation economically rational. Brookings’ May 2026 employment analysis gives that argument some support: counties receiving their first large data center saw total private employment rise by 4% to 5% over five to six years, construction employment jump 11%, information-sector employment grow 22%, and wages rise 3% to 4%. Those are real gains. But the same Brookings analysis also shows why the developmental argument remains incomplete. The biggest information-sector gains tend to appear only in counties with multiple facilities and broader clustering, while critics remain right that large projects often promise only dozens to a few hundred permanent workers. Brookings’ water analysis further reports that large facilities can use an estimated 5 million gallons of water per day, roughly the needs of a town of up to 50,000 residents. Taken together, those findings do not refute the paper’s argument. They sharpen it. Human return is possible, but it is not automatic, and it is not always proportionate to the resource burden being imposed.

The third counterargument is that large-scale capital formation is itself a public good. According to this view, if institutional investors, infrastructure funds, and debt markets are willing to finance AI infrastructure, that willingness should be treated as evidence of economic promise rather than democratic risk. The scale is indeed large. Total U.S. retirement assets reached $49.1 trillion at year-end 2025. Brookfield’s AI infrastructure platform is expressly structured as an institutional vehicle spanning energy, land, data centers, and compute. Reuters reported that Morgan Stanley expects AI-related global debt issuance to exceed $570 billion in 2026. But none of those facts establishes that the resulting buildout is publicly legitimate simply because it is financeable. Financial usefulness to large allocators does not answer the underlying question of who bears the burden and who captures the gains. It can deepen the problem by making later correction more expensive as more balance sheets become exposed to the continuation of the buildout. Capital formation can be rational for investors and still be anti-human in economic effect for workers, households, and host communities.

The fourth counterargument is that markets will self-correct if the buildout overshoots. If valuations become excessive, if capacity is overbuilt, or if debt becomes too large, market discipline will eventually force a repricing. That proposition may be true in a narrow financial sense, but it is inadequate as a public-governance answer. By the time markets correct, public burdens can already be entrenched. DOE and Berkeley Lab show that data-center electricity demand is already large enough to reshape utility planning. GAO shows that the sector’s energy and water burdens remain incompletely disclosed. Those are not merely asset-pricing issues. They are public-system facts. A late financial correction does not automatically return forgone tax capacity, unwind local siting decisions, reverse utility investments, or restore public bargaining power once legal and financial dependence have hardened. Markets can correct price. They do not necessarily correct governance.

The fifth counterargument is that ordinary democratic politics remains intact and therefore the buildout remains publicly legitimate. Elections continue, legislatures debate, courts remain open, and local opposition is visible. Those points are true, but they do not rebut the narrower concern advanced here. Reuters/Ipsos reported on June 11, 2026 that only one-third of Americans support the rapid construction of AI-driven data centers, that 64% oppose the pace, that 57% oppose having such a facility in their own community, and that 77% worry AI developments could increase electricity costs. The public is already registering the burden as a household-security issue. The mere existence of elections and dissent therefore does not settle the matter. Oligarchic drift, as used in this paper, does not require the disappearance of formal democratic procedure. It requires only that the effective cost of using public authority against the dominant buildout model rise high enough that institutions increasingly adapt to it rather than govern it. The present record remains consistent with that concern.

The limits of these counterarguments are therefore cumulative. Strategic urgency does not answer the governance problem. Local gains do not erase thin direct human return relative to resource burden. Capital formation does not substitute for a human-income architecture. Market correction does not undo public-system dependence. And the continued existence of elections does not guarantee practical sovereignty once enough law, infrastructure, and finance have already been reorganized around the continuity of the buildout. The strongest defenses of the current model identify real benefits or constraints. They do not yet show that the public remains first in the bargain. On that question, the burden of proof still lies with the buildout’s defenders, and on the present record, that burden remains unmet.

X. Reversal Program: Restoring Human Priority to AI Governance

If the core failure of the current buildout is asymmetric adaptation, then the corrective program cannot be symbolic. It must reverse the order of priority. The United States does not need to abandon AI development. It does need to stop granting the machine side of the transition a level of legal speed, infrastructure accommodation, fiscal support, and financial embedding that the human side has not yet earned. The objective of the program that follows is therefore not to slow innovation for its own sake. It is to restore human priority: workers, households, communities, and democratic institutions should not be treated as the adjustment variable of a machine-centered regime. That recommendation follows from the record developed above: concentrated compute, rapidly rising public-system burden, incomplete disclosure, widening financial lock-in, and an underbuilt replacement-income architecture.

The first requirement is to end unconditional public support for AI-linked infrastructure. States and the federal government should not continue granting open-ended data-center incentives, tax abatements, or comparable concessions without short sunset periods, project-level disclosure, enforceable performance conditions, and automatic clawbacks. Georgia’s official review found that roughly 70% of recent data-center construction likely would have occurred without the tax break, while Ohio suspended new applicants after reported tax-expenditure costs rose sharply above prior projections. Those facts support a presumption against blanket accommodation. Public support should be treated as justified only where proponents can prove durable local human return, not merely the presence of large capital expenditure.

The second requirement is strict cost allocation for large-load facilities. The public should not serve as the insurer of first resort for hyperscale AI expansion. Senator Chris Van Hollen’s January 2026 proposal rests on precisely this principle, stating that Americans should not have to foot the bill for large data centers and that corporations, not consumers, should bear the costs of expansion and associated grid-reliability measures. That principle should be generalized. Where a facility drives transmission upgrades, generation additions, reserve-margin needs, distribution expansion, water-system stress, or other public burdens, those costs should be assigned back to the facility wherever law and regulation permit. A buildout that remains viable only by displacing its infrastructure costs onto households is not a neutral market outcome; it is a publicly subsidized transfer.

The third requirement is full burden disclosure. Congress and the states should require standardized, project-level reporting on electricity demand, peak load, water withdrawals, water consumption, cooling method, backup generation, all public incentives received, and material grid or water infrastructure required to support the facility. GAO’s April 2025 assessment states that generative AI uses significant energy and water resources while companies generally are not reporting the details. That is already enough to justify mandatory disclosure. A republic cannot govern what it is not allowed to see, and no serious cost-allocation regime can function while the most important burden variables remain opaque.

The fourth requirement is structural discipline at the compute layer. The FTC’s January 2025 staff report documented that major cloud-AI partnerships combine equity and revenue-sharing rights, consultation, control, exclusivity, cloud-spend commitments, access to training data, and chip co-development, while raising concerns about switching costs, access to key inputs, and information asymmetries. That record does not, by itself, establish that every such arrangement is unlawful. It does establish that concentration at the infrastructure layer is sufficiently severe to justify structural remedies or stringent conditions. The public should not accept a future in which the same firms control the chokepoints of compute, the preferred partnership terms of frontier developers, and the effective tempo of the buildout. Public enforcement should therefore move from generalized concern to presumptive limits on arrangements that consolidate both infrastructure power and governance influence in the same narrow set of firms.

The fifth requirement is to reject blanket preemption without prior public safeguards. The White House’s March 2026 legislative recommendations urge Congress to preempt state AI laws that impose “undue burdens” and to establish a “minimally burdensome” national framework. That sequence is backwards. A federal framework that arrives before robust cost-allocation rules, disclosure obligations, labor protections, and meaningful local recourse would not protect democracy from fragmentation; it would strip away one of the few remaining public counterweights while the sector is still consolidating. Federal uniformity is not inherently illegitimate, but it must follow, not precede, the construction of a genuinely protective human-priority framework. Until such protections exist, distributed state and local authority should be treated as a democratic asset rather than a regulatory inconvenience.

The sixth requirement is fiduciary transparency and stress testing for retirement-linked exposure. Total U.S. retirement assets reached $49.1 trillion at year-end 2025, and AI infrastructure is already being packaged in forms suitable for institutional mandates. Beneficiaries should therefore know their direct and indirect exposure to digital-infrastructure funds, data-center debt, AI-linked private credit, utility expansion dependent on hyperscale load, and concentration in firms whose valuations depend heavily on AI capex assumptions. This recommendation is an inference from the scale of retirement assets, the institutional structure of AI infrastructure vehicles, and the rapid growth of AI-related debt issuance. It is necessary because the public should not discover only after the fact that its savings architecture was silently enlisted in a buildout whose burdens it was already being asked to host.

The seventh requirement is a replacement-income architecture proportionate to the machine answer. The White House’s own framework says workers “must benefit” from AI-driven growth, but the institutional response remains centered on training, study, and adaptation rather than on a fully specified architecture for income continuity. That gap should no longer be tolerated. Any serious national AI strategy should be required to identify, in advance and at commensurate scale, the mechanisms through which workers and communities will retain income, bargaining power, and durable participation if AI deepening weakens labor demand or thins the number of stable livelihoods directly supported by production. This is not a call for a single predetermined policy instrument. It is a demand that the human side of the transition be specified with the same seriousness now devoted to power provisioning, capex schedules, debt facilities, and legal preemption. Without that, the machine answer remains funded while the human answer remains rhetorical.

The eighth requirement is public-interest compute capacity. This recommendation is partly inferential, but it follows directly from the documented concentration of the compute layer and the federal government’s increasing willingness to treat rapid access to privately controlled frontier systems as a national priority. A republic that wishes to preserve leverage over strategically important infrastructure cannot rely exclusively on privately governed compute. Universities, national laboratories, research consortia, and mission agencies should therefore be supported in building public or nonprofit compute capacity sufficient to reduce dependence on a handful of commercial vendors. The purpose is not to displace private firms entirely. It is to ensure that the public retains some independent technical footing from which to govern, experiment, and negotiate.

The final requirement is to reverse the burden of proof. The current buildout continues to receive a presumption of legitimacy that the present record does not justify. Firms seeking subsidy, accelerated permitting, reduced oversight, or preemption of state authority should bear the burden of showing that the associated costs will not be shifted to households and communities, that the financing will not create opaque public dependence, that the gains will not be captured primarily by concentrated private actors, and that the public will retain the practical capacity to impose terms later if necessary. On the evidence reviewed in this paper, that showing has not yet been made. Until it is, the default policy posture should be conditionality, not accommodation.

XI. Conclusion: The Country Is Learning How to Fund the Machine Answer Before It Has Built the Human Answer

The central finding of this paper is limited, specific, and severe. The present U.S. AI buildout is not best understood as a neutral wave of innovation moving through an otherwise intact democratic political economy. It is better understood as a machine-first regime of asymmetric adaptation. Law, infrastructure, subsidy, and long-duration capital are being specified rapidly for the needs of compute and concentrated private expansion, while the conditions of human continuity—household income security, worker bargaining power, local control, and durable public return—remain comparatively underdefined, underprotected, and underbuilt. That is the governing asymmetry. It is the reason the current model is anti-human in economic effect even without requiring any claim about hidden motive.

The evidence supporting that judgment is already substantial. The compute layer is highly concentrated, and the FTC has documented partnership structures marked by equity and revenue-sharing rights, consultation, control, exclusivity, cloud dependence, and switching-cost concerns. Federal policy is moving toward a “minimally burdensome” national framework and preemption of state AI laws deemed too restrictive. Data-center electricity demand is already large enough to reshape utility planning, and the energy and water burdens of generative AI remain incompletely disclosed. These are not isolated facts. Together they show that the country is already learning how to power, permit, and protect the machine answer.

The same pattern appears in finance. Total U.S. retirement assets reached $49.1 trillion at year-end 2025, and AI infrastructure is already being packaged in forms capable of drawing in pensions, infrastructure funds, insurers, and credit markets. That fact does not prove that all household savings are being redirected into AI. It does prove that the buildout is no longer confined to founders, hyperscalers, or venture capital. It is extending into the savings architecture of the country. Once that process matures, democratic correction becomes more difficult because the public is no longer engaged only as citizen, worker, ratepayer, or local resident. It is increasingly engaged as an involuntary financial participant in the continuation of the same regime.

What remains missing is the human answer. The United States has learned how to mobilize capital for machine infrastructure. It has not built a comparably serious architecture for replacement income, bargaining-power preservation, or durable local human return where that same buildout weakens labor demand or thins the number of stable livelihoods directly supported by production. The White House’s own framework states that workers “must benefit” from AI-driven growth. That formulation is revealing. It describes a requirement, not an accomplished fact. The machine answer is being financed and accelerated with institutional precision. The human answer remains an objective.

That is why the constitutional danger described here is narrower than the collapse of formal democracy and more serious than ordinary policy error. The risk is loss of practical sovereignty. Public institutions may retain ballots, courts, legislatures, and agencies while steadily losing the practical capacity to impose terms on a sector that has become too infrastructurally necessary, too financially embedded, and too legally accommodated to challenge without systemic pain. Under those conditions, the public can still speak. It becomes less able to decide. That is what this paper means by oligarchic drift.

The policy consequence follows directly. The United States does not need to reject AI development. It does need to reject a buildout model in which concentrated private actors receive public legal accommodation, public infrastructure support, public fiscal tolerance, and growing access to public savings before proving proportionate human return. Until the burden of proof is reversed—until the buildout is required to show full cost internalization, durable local benefit, a serious replacement-income architecture, robust disclosure, preserved democratic veto points, and limited financial lock-in—the default posture should not be acceleration. It should be conditionality.

The decisive question is therefore no longer whether AI will be powerful. The decisive question is whether the United States will continue teaching its institutions to plan around machines first and people later. On the present record, that shift is already underway. The country is learning how to fund the machine answer before it has built the human answer. That is the formal judgment of this paper.