The Productivity Gambit: Can AI and Automation Outrun Demographic Collapse?

AI, Automation & Energy Independence: The Last Escape Hatch

A quantitative assessment of whether artificial intelligence and automation can generate enough per-worker output to sustain aging welfare states — and what the vast energy demands of that AI actually require, from small modular nuclear reactors to orbiting solar power stations.

▶  Bottom Line Up Front (BLUF)

• AI and automation are real and measurable productivity forces — controlled studies show 15–55% task-level efficiency gains, and PwC data shows wages rising twice as fast in AI-exposed industries. But "task productivity" and "macro fiscal rescue" are radically different scales.
• The honest productivity estimates range from modest to transformative: Daron Acemoglu (MIT/Nobel) projects 0.9–1.8% cumulative GDP gains over 10 years; Penn Wharton Budget Model projects 1.5% GDP gain by 2035 and 3.7% by 2075; Goldman Sachs projects 7% global GDP gain over a decade. None of these, by themselves, close the demographic fiscal gap.
• The math problem is fundamental: Pension systems need not just more output, but more taxable output distributed through wages. If AI concentrates gains in capital (corporate profits) and not labor income, tax revenues from workers do not grow proportionally — creating a political crisis around how to tax AI's owners to fund retirees.
• AI's energy hunger is genuine and enormous: Global data center electricity consumption will more than double by 2030 to ~945 TWh — roughly Japan's entire annual electricity demand. In the U.S., data centers will account for nearly half of all new electricity demand growth through 2030.
• Green energy cannot meet baseload AI demand alone. Wind and solar are intermittent; AI inference runs 24/7. The IEA projects coal and natural gas will supply the majority of new data center electricity through 2030. Nuclear — both conventional and small modular reactors (SMRs) — is the only clean, dispatchable baseload solution.
• SMR deployment is real but slow: Google, Microsoft, Amazon, and Meta have committed over $10 billion to nuclear partnerships. First SMRs come online around 2030; meaningful scale arrives 2033–2038. The HALEU fuel supply chain, regulatory queue, and first-of-a-kind engineering challenges are genuine bottlenecks.
• Space-based solar power is a compelling long-term concept — geosynchronous arrays deliver 24/7 energy with no weather dependence — but commercial viability is a 2040s+ proposition. Current demonstrations transmit milliwatts, not megawatts. SpaceX Starship economics are improving the launch cost calculus but have not yet crossed the commercial threshold.
• The synthesis verdict: AI and automation can materially slow demographic fiscal deterioration, but cannot by themselves prevent it. Energy supply is the binding constraint on AI's deployment speed. Nuclear — not green renewables — is the bridge technology. A full rescue requires AI productivity + nuclear energy + structural pension reform simultaneously, not sequentially.

The argument for technology as a demographic salvation has an appealing internal logic. If each worker becomes dramatically more productive — if one person, augmented by AI, can do the economic work of two or three — then the ratio of workers to retirees becomes less relevant to fiscal sustainability than the ratio of output to obligations. The welfare state was designed around headcount, the argument goes, not productivity. Reframe the problem and the crisis disappears.

This argument is not wrong. It is, however, far more complicated, narrower in its fiscal application, and slower in its timeline than the political version of it usually suggests. The actual data on AI productivity gains, the specific mechanisms by which those gains translate (or fail to translate) into pension revenue, the massive energy infrastructure required to run the AI, and the realistic timeline for deploying that infrastructure form a chain of dependencies that is only as strong as its weakest link. This report assembles that chain in quantitative terms.

Part I: The Productivity Evidence — What the Research Actually Shows

There are now dozens of serious empirical studies on AI's productivity impact, ranging from micro-level controlled experiments to macro-level GDP modeling. The findings vary widely — not because researchers disagree on methodology, but because the honest answer is highly context-dependent. Understanding this range is essential.

Micro-Level: Task Productivity is Real and Large

At the level of individual tasks and workers, the evidence is genuinely impressive. A randomized controlled study by Shakked Noy and Whitney Zhang found that ChatGPT use among professionals reduced task-completion time by roughly 40% and increased output quality by 18%. [1] Erik Brynjolfsson's research on a Fortune 500 customer-service deployment found a 15% average productivity increase — measured as issues resolved per hour — with the bottom skill quintile gaining 36%, the single most striking finding in the micro literature. [2] PwC's 2025 Global AI Jobs Barometer, analyzing close to one billion job advertisements from six continents, found wages rising twice as fast in AI-exposed industries, with revenue growth in the most AI-capable industries having nearly quadrupled since ChatGPT's release in 2022. [3]

Deloitte's February 2025 survey of over 11,000 workers found that agentic AI systems — which chain multiple AI models into autonomous workflows — are beginning to boost productivity across job levels. [4] And the EY US AI Pulse Survey of October 2025 found that 78% of organizations now use AI in at least one business function, up from 55% in 2023, with workers using AI reporting an average 40% productivity boost. [5]

Macro-Level: The "Productivity Paradox" Persists

The picture at the macroeconomic level is far more sobering — and economists have seen this movie before. In 1987, Robert Solow famously observed: "You can see the computer age everywhere but in the productivity statistics." Four decades of IT investment produced enormous task-level improvements but strikingly modest aggregate productivity gains through most of that period. The Richmond Fed's October 2025 analysis drew the explicit historical parallel, noting that "the estimates of the boom in productivity from AI are perhaps more likely to follow along the lines of Acemoglu's work rather than that of Goldman Sachs or McKinsey." [6]

The forecasts from serious economists span three orders of magnitude in their optimism:

Daron Acemoglu (MIT)

0.9–1.8% GDP
over 10 yrs

Penn Wharton PWBM

1.5% by 2035
3.7% by 2075

McKinsey Global Institute

+1.5–3.4 pp
annual GDP growth

Goldman Sachs

+$7T GDP
over 10 yrs globally

Yale Budget Lab (2026)

Near-zero
measured so far

Yale's Budget Lab found that through 2025, AI measures of exposure, automation, and augmentation showed "no sign of being related to changes in employment or unemployment" at the aggregate level, consistent with what Erik Brynjolfsson calls the "productivity J-curve" — the idea that general-purpose technologies require years of organizational restructuring before their gains appear in statistics. [7]

The International Center for Law and Economics' February 2026 comprehensive literature review summarized the current state: the macroeconomic evidence is too limited to resolve the debate, "but it provides a structured framework for evaluating future research." The three key variables — what share of GDP tasks AI can affect, how fast adoption occurs, and whether complementary organizational investments happen at scale — determine whether we are in an Acemoglu world or a Goldman Sachs world. [8]

Part II: The Fiscal Translation Problem — Output vs. Pension Revenue

Even accepting the most optimistic AI productivity projections, there is a crucial step that most discussions of "AI solves demographics" skip entirely: the mechanism by which AI-generated productivity gains flow into government pension revenues. This is where the demographic math meets the political economy — and it is deeply problematic.

Pension systems in most developed countries are funded primarily through payroll taxes — taxes on wages. Social Security in the United States, for instance, is funded by a 12.4% payroll tax on earned income up to a cap. When AI increases productivity by making each worker produce more, two very different outcomes are possible:

Scenario A (Labor-Complementary AI): Workers become more productive, employers compete for skilled AI-augmented workers, wages rise, and payroll tax revenue grows proportionally. The pension system benefits directly. This is the path suggested by the PwC data showing wages rising twice as fast in AI-exposed industries.

Scenario B (Capital-Concentrating AI): AI replaces enough worker tasks that labor's share of income falls, corporate profits rise, wages stagnate or fall for routine workers, and payroll tax revenue grows much more slowly than total output. The gains accrue to shareholders — who are not subject to payroll taxes — while the pension funding gap persists or widens. ScienceDirect's life-cycle model (2020) found explicitly that both population aging and automation "reduce the labor share of income" as robots substitute for routine workers, with gains concentrated in non-routine labor and capital. [9]

The Penn Wharton Budget Model was blunt about this constraint: "There's this belief among policymakers that in this new era of AI, we don't have to be fiscally responsible because AI is going to solve everything," said PWBM faculty director Kent Smetters — explicitly warning against that assumption. [10] The fiscal frontier requires not just more output but more taxable output in the forms governments can actually reach.

AI can generate enormous wealth without generating the payroll tax revenue that funds pensions. This is not a hypothetical risk — it is already happening in the tech sector.

The resolution to this problem would require either taxing AI-generated capital income at rates comparable to payroll taxes — a radical restructuring of the tax system — or implementing something like a robot tax, which has been periodically proposed but never enacted in any major economy. Without this fiscal bridge, even dramatic AI productivity gains may leave Social Security, public pensions, and healthcare funding structurally starved of revenue as the labor share of income declines.

Part III: The Energy Reckoning — AI's Enormous and Growing Appetite

If productivity is the first constraint on AI's demographic rescue potential, energy is the second — and it is arguably more tractable but equally urgent. The IEA's landmark April 2025 report Energy and AI provided the most comprehensive quantitative picture to date. [11]

Global data center electricity consumption stood at approximately 415 TWh in 2024. By 2030, the IEA's base case projects it will reach 945 TWh — more than doubling in six years, and equivalent to Japan's entire annual electricity consumption. In the IEA's accelerated "Lift-Off" scenario, the figure reaches 1,700 TWh by 2035 — roughly the combined electricity demand of Germany and France. In the United States alone, data centers will account for nearly half of all new electricity demand growth through 2030, consuming more power by 2030 for data processing than for manufacturing all energy-intensive goods combined — aluminum, steel, cement, chemicals. [12]

A single AI inference query already consumes roughly 10 times the electricity of a conventional web search. Gartner estimates AI-optimized servers will represent 44% of total data center power usage by 2030, up from 21% in 2025, with AI server electricity rising nearly fivefold from 93 TWh in 2025 to 432 TWh in 2030. [13] A Carnegie Mellon study estimated that data centers and cryptocurrency mining could cause an 8% increase in the average U.S. electricity bill by 2030, potentially exceeding 25% in the highest-demand markets of Virginia and Ohio.

The Intermittency Problem: Why Green Power Cannot Solo

Solar and wind are the cheapest new electricity sources on the planet. But AI data centers have a fundamental operational requirement that neither can meet: continuous, uninterruptible 24/7 baseload power. Training runs cannot pause for cloudy days. Inference servers cannot tolerate brownouts. Battery storage at the scale required to buffer intermittent renewables for large hyperscale campuses is not currently cost-effective, though costs are falling.

The IEA's own projections reflect this reality. In its base case, coal remains the largest single source of additional electricity for data centers through 2030, with natural gas close behind. Renewables supply approximately half of all new data center electricity demand by 2035 — but only because coal and gas plug the baseload gap in the meantime. "Fossil fuels remain important for meeting the near-term surge in demand up to 2030," the IEA stated plainly. [11] A system that runs AI on fossil fuels creates a painful irony: the technology meant to enable sustainable economies is itself accelerating carbon emissions in the short run.

Year	Global DC Demand	Primary New Supply (Base Case)	Nuclear Share	Notable Context
2024	415 TWh	Natural gas 40%, Renewables 27%, Nuclear 15%	~62 TWh	Baseline — U.S. 183 TWh (4% of total consumption)
2030	945 TWh (base)	Renewables ~50%, Gas growth, Coal growth	Growing; first SMRs online	U.S. DCs hit 426 TWh — nearly half of all new U.S. power demand growth
2035	1,200–1,700 TWh	Renewables 60% of new supply; SMRs scaling	~175 TWh additional	AI servers = 44% of DC power; ratio flips to 60% clean

Part IV: The Nuclear Solution — SMRs as AI's Power Source

Nuclear power — uniquely among clean energy sources — provides continuous, weather-independent, high-density baseload electricity exactly matched to AI's operational requirements. It emits no carbon in operation. It occupies a fraction of the land required by equivalent solar or wind installations. And since 2023, it has attracted the most serious private-sector investment in the technology's history, driven entirely by AI demand.

The commitments are substantial. Google made history in October 2024 with the world's first corporate SMR purchase agreement, partnering with Kairos Power for 500 MW across six to seven molten salt reactors, with the first unit targeted for 2030 and full deployment by 2035. [14] Microsoft secured 837 MW by reviving Three Mile Island's Unit 1 — renamed the Crane Clean Energy Center — through a 20-year agreement with Constellation Energy, with power expected by 2028. [15] Amazon committed over $20 billion to the Susquehanna nuclear campus and backed X-Energy's SMR program for 5 GW by 2039. Meta issued an RFP targeting 1–4 GW of new nuclear. Oracle announced plans for a gigawatt-scale campus powered by three SMRs. [16]

Collectively, tech companies signed contracts for over 10 GW of possible new nuclear capacity in the United States in 2024 alone — a transformative signal to an industry that had been in slow decline. The Trump administration's May 2025 nuclear strategy set a goal to quadruple U.S. nuclear capacity to 400 GW by 2050. The UK announced a £2.5 billion SMR deployment package in June 2025, targeting first deployments in the mid-2030s, including Rolls-Royce SMR units at Wylfa in Wales. [17]

TerraPower Natrium — Wyoming

Groundbreaking June 2024. Sodium-cooled fast reactor, 345 MW base / 500 MW peak via molten salt storage. Bill Gates-backed; DOE-supported. Target: 2030 commercial operation. First advanced reactor under construction in the U.S.

Online ~2030 HALEU dependency Flexible output

Kairos Power / Google — Southeast U.S.

Fluoride salt-cooled, pebble fuel (no meltdown risk). 75 MW per unit, deployed in pairs for 150 MW. Demonstration reactor permitted in Tennessee. First unit for Google: 2030. Full 500 MW fleet: 2035.

First unit 2030 Meltdown-proof design First-of-kind risk

China Linglong One — Already Operating

World's first operational commercial land-based SMR. 210 MW at Hainan Province, online 2023. China has allocated $25–35 billion for domestic SMR deployment and is positioned to dominate the global SMR export market. By 2030 China plans SMRs specifically co-located with data centers in western provinces.

Already operational Significant lead

Rolls-Royce SMR — UK/Europe

470 MW light water reactor. UK government selected for Wylfa deployment. Joint venture includes Qatar sovereign wealth fund, BNF Resources France, Constellation U.S., and Czech utility CEZ. Target: mid-2030s first unit. Czech and Polish deployment planned. UK-Atlantic partnership with U.S. for joint regulatory fast-track.

Mid-2030s Regulatory risk Government backed

The Hard Engineering Constraints on SMR Speed

The SMR momentum is genuine, but so are the bottlenecks. Every SMR project is a first-of-a-kind (FOAK) deployment of an unproven commercial design. Gen IV reactors operating at 700–950°C using molten salt or liquid sodium require approximately 10 years of material testing data to certify structural components for 30-year lifespans — a timeline that cannot be compressed. [18]

The HALEU (High-Assay Low-Enriched Uranium) fuel supply chain is perhaps the most acute bottleneck. Most advanced SMR designs require HALEU enriched to 5–20% U-235, compared to the 3–5% used in conventional reactors. Current U.S. licensed production stands at approximately 900 kilograms per year; projected demand by 2035 is 50+ metric tons annually. The DOE has allocated HALEU to five companies in 2025, but the gap between current supply and projected demand is severe. [19] TerraPower's Natrium reactor, backed by $4 billion in DOE support, has already been delayed from 2028 to 2030 specifically due to HALEU fuel scarcity.

Grid interconnection queues stretch to 2035 in many U.S. regions. The IEA warns that unless transmission infrastructure is significantly upgraded, up to 20% of planned data center projects could face delays. Transformer delivery times have doubled in the past three years. The conclusion is that the nuclear renaissance is real — but its meaningful contribution to AI power supply arrives 2033–2040, not 2026–2029.

Part V: Space-Based Solar Power — The Long-Term Wild Card

Your instinct about orbiting solar-powered servers points to one of the most genuinely transformative long-term concepts in energy. The physics is compelling: satellites in geostationary orbit (GEO) 36,000 km above Earth receive sunlight 99.7% of the time — essentially 24/7 with only brief eclipses around the equinoxes. Solar intensity in GEO is 144% of the maximum achievable on Earth's surface, without atmospheric attenuation, weather, or night. Energy is converted to microwave radiation and beamed to large ground-based receiving antennas (rectennas), which convert it back to DC electricity at approximately 85% efficiency. The beam itself, at 230 W/m² peak intensity, is roughly a quarter the strength of midday sunlight — not harmful to humans or wildlife if wandering off target. [20]

The concept was first proposed by Peter Glaser in 1968, and the fundamental physics has been validated. In 2023, Caltech's MAPLE experiment became the first device to wirelessly transmit power in space and send a detectable signal to Earth. The Air Force Research Laboratory's Arachne experiment demonstrated the modular sandwich-tile approach to space-to-ground transmission. Japan's OHISAMA program is testing a 180-kilogram LEO satellite prototype. Several commercial startups — Aetherflux, Overview Energy, Emrod — have raised early capital. [21]

What Still Stands Between Concept and Commercial Reality

The gap between a Caltech demo transmitting milliwatts and a gigawatt commercial power station is measured in decades, not years. The obstacles are not primarily scientific but engineering-economic:

Mass and launch costs. A gigawatt-scale GEO power satellite would span multiple kilometers and weigh thousands of tonnes. The ISS — the largest structure ever built in orbit — weighs 420 tonnes and produces less than 100 kW from its 164 solar panels. SpaceX Starship has reduced LEO launch costs toward $100–200/kg from the Shuttle-era $50,000/kg, which fundamentally improves the calculus. But a 1 GW power station may still require 50–100 Starship-class launches for its components alone. [22]

In-orbit assembly. Structures of the required size cannot be launched fully assembled — they require autonomous robot assembly in space, a capability that does not yet exist at commercial scale. The IEEE Spectrum noted that "there is no comparable incremental path to a robot-assembled power satellite in GEO." [23]

Efficiency losses. End-to-end efficiency from solar panel through microwave transmission to grid-ready DC electricity is approximately 50–60% for optimistic systems — compared to roughly 20% for terrestrial photovoltaics (accounting for nighttime and weather). The constant solar exposure and high intensity partially compensate, but the conversion chain losses are not trivial.

Timeline consensus: The European Space Agency's Solaris program, launched in 2022 with €60 million for feasibility assessment, projected commercial viability no earlier than the 2040s with sustained investment. A well-capitalized 2026 analysis projected orbital demos through 2030 preceding gigawatt-scale deployment by 2035 — optimistic by most independent assessments. Most mainstream analysts place commercial SBSP in the 2040–2060 window. [24]

One highly relevant emerging concept: space-based data centers rather than space-based power transmission. Aetherflux pivoted in December 2025 from power-beaming to space-based data center operations — compute infrastructure in orbit powered directly by space solar, eliminating the transmission-to-Earth step entirely. If orbital manufacturing and servicing mature through the 2030s, this may be a more achievable path than GEO power stations for Earth grids. It is, however, an even longer-horizon proposition.

Space-based solar is not a 2030 answer. It is not a 2035 answer. It is a 2045–2060 answer — assuming sustained investment, SpaceX-class launch economics, and autonomous orbital assembly. The demographic crisis peaks before it arrives.

Part VI: The Synthesis — Can Technology Close the Gap?

The Integrated Assessment: Conditions for a Technological Rescue

For AI and automation to stave off demographic fiscal collapse, three conditions must be met simultaneously — not sequentially. They are interdependent, and the failure of any one collapses the chain.

Condition 1 — Sufficient productivity gains, quickly enough. The Penn Wharton model projects 1.5% GDP gain by 2035. The demographic fiscal gap in the United States is approximately 3.82% of taxable payroll over 75 years; the EU faces a 40-percentage-point excess over sustainable debt by 2040. These do not close each other. The Goldman Sachs optimistic scenario gets closer, but Goldman itself estimates 15% productivity improvement "when fully adopted" — a process taking decades, while the insolvency timeline is years. The productivity gains need to arrive on a 10-year horizon, not a 30-year one, to prevent the Phase 2 forced adjustment events (2033–2040).

Condition 2 — Gains must flow through wages, not just capital. If AI concentrates value in corporate profits and capital income while labor's share falls, pension and Social Security systems — funded by payroll taxes — receive diminishing benefit. This requires either a structural tax reform (taxing AI/automation capital income at rates comparable to payroll taxes) or an explicit AI dividend/robot tax redistributed to fund entitlements. Neither has been enacted anywhere. Without this, even strong productivity growth fails to rescue pension systems.

Condition 3 — Sufficient clean energy to run the AI at scale. If AI infrastructure runs on coal and gas (the IEA's near-term projection), its productivity gains come with carbon emissions — creating a political constraint on deployment. If it runs on nuclear, which is clean and dispatchable, the deployment can scale unconstrained. SMRs are the key bridge technology: the first meaningful deployments arrive 2030–2033, scaling through 2038. This creates a 7–12 year gap between now and the point where nuclear can cleanly power AI at demographic-rescue scale. That gap is bridged by natural gas — acceptably, if imperfectly.

The most likely outcome, integrating all three conditions: AI and automation provide a meaningful partial offset to demographic fiscal stress — perhaps reducing the required Social Security reform package by 15–25%, and allowing some developed nations that are close to the margin (the U.S., Germany) to make more modest structural adjustments than would otherwise be needed. They do not prevent the need for structural adjustment in Southern Europe or Japan, which face gaps too large to be technology-bridged. The productivity gains arrive too slowly and are captured too much in capital rather than labor to substitute for the pension reforms that were needed in the 2010s and still need to happen in the 2030s.

Space-based solar is not a factor in the demographic crisis window (2026–2050). It is potentially transformative for the second half of the 21st century — providing unlimited clean baseload power that could support both an aging population's care needs and whatever AI systems are deployed to serve them. It is a reason for long-run optimism, not a reason to delay near-term structural reform.

The Aging-Automation Feedback Loop: One Genuine Positive

There is one mechanism the models may be underestimating: the NBER and ScienceDirect research showing that aging populations themselves drive automation adoption, as the labor scarcity they create raises wages and makes automation economically attractive to firms. [25] Japan — with its extreme demographic pressure — has the world's highest robot density in manufacturing and has been an automation pioneer precisely because of its workforce shortage. South Korea, with fertility at 0.72 and shrinking fast, has the world's highest overall robot density at 1,012 per 10,000 manufacturing workers. [26]

The irony is that the crisis itself may accelerate the solution. As Western labor markets tighten due to demographic decline, wages for remaining workers rise, making AI and robot substitution more economically compelling for firms. This self-reinforcing cycle may produce faster AI adoption than purely demand-driven projections suggest. It will not arrive fast enough to prevent Social Security insolvency in 2033, but it may meaningfully reduce the reform burden required in the 2040s and 2050s.

The Crucial Missing Variable: Political Will to Capture the Gains

The deepest constraint is political, not technical. Even if AI generates sufficient productivity gains, the current tax architecture of most developed nations is poorly designed to capture them for public pension systems. Capital gains taxes are lower than payroll taxes in virtually every OECD country. Corporate income taxes have been declining for decades. Wealth taxes face legal and practical constraints. The productivity gains from AI will flow into the financial accounts of Amazon, Microsoft, Google, Nvidia, and their shareholders — not automatically into the Social Security Trust Fund.

Solving this requires a deliberate political choice to restructure how AI-generated value is taxed and distributed. That political choice requires the same courage as directly raising retirement ages or cutting benefits — and it faces the same opposition from concentrated economic interests. Technology does not bypass politics. It merely changes which political fights must be won.

This report is Part III of a three-part series on the demographic crisis in developed nations. Part I examined the mathematical limits of immigration as a solution. Part II assessed country-by-country collapse timelines. This report addresses the technology offset question. All three draw on primary and peer-reviewed sources as cited; full citations follow. The analysis represents this publication's assessment of quantitative evidence and should not be taken as investment advice or policy advocacy.

Verified Sources & Formal Citations

Noy, S. & Zhang, W. "Experimental Evidence on the Productivity Effects of Generative AI." Science 381(6654), 2023. Summarized in: International Center for Law & Economics. "AI, Productivity, and Labor Markets: A Review of the Empirical Evidence." February 5, 2026.
https://laweconcenter.org/resources/ai-productivity-and-labor-markets-a-review-of-the-empirical-evidence/

Brynjolfsson, E., Li, D., & Raymond, L. "Generative AI at Work." NBER Working Paper, updated 2025. Summarized in ICLE review (see [1] above).

PwC. 2025 Global AI Jobs Barometer: The Fearless Future. PricewaterhouseCoopers, 2025. Analysis of ~1 billion job advertisements globally.
https://www.pwc.com/gx/en/services/ai/ai-jobs-barometer.html

Deloitte. "AI, Demographic Shifts, and Agility: Preparing for the Next Workforce Evolution." Deloitte Insights, December 24, 2025. Based on survey of 11,387 workers, February 2025.
https://www.deloitte.com/us/en/insights/topics/talent/strategies-for-workforce-evolution.html

Ernst & Young. "AI-Driven Productivity Is Fueling Reinvestment Over Workforce Reductions." EY US AI Pulse Survey, 4th wave (Oct. 2025 fieldwork). December 9, 2025.
https://www.ey.com/en_us/newsroom/2025/12/ai-driven-productivity-is-fueling-reinvestment-over-workforce-reductions

Federal Reserve Bank of Richmond. "The Productivity Puzzle: AI, Technology Adoption and the Workforce." Economic Brief EB 24-25, October 2, 2025. Referencing Acemoglu, Solow, and historical IT adoption comparisons.
https://www.richmondfed.org/publications/research/economic_brief/2024/eb_24-25

Gimbel, M. et al. (Budget Lab at Yale). "Evaluating the Impact of AI on the Labor Market: Current State of Affairs." The Budget Lab, Yale University, 2025.
https://budgetlab.yale.edu/research/evaluating-impact-ai-labor-market-current-state-affairs
See also: Brynjolfsson, E., Rock, D., & Syverson, C. (2021). "The Productivity J-Curve." American Economic Journal: Macroeconomics.

International Center for Law & Economics (ICLE). "AI, Productivity, and Labor Markets: A Review of the Empirical Evidence." February 5, 2026. Comprehensive literature synthesis.
https://laweconcenter.org/resources/ai-productivity-and-labor-markets-a-review-of-the-empirical-evidence/

Abeliansky, A.L. & Prettner, K. "The Impact of Aging and Automation on the Macroeconomy and Inequality." Journal of Macroeconomics (ScienceDirect), November 2020.
https://www.sciencedirect.com/science/article/abs/pii/S0164070420302020

Arnon, A. & Smetters, K. "The Projected Impact of Generative AI on Future Productivity Growth." Penn Wharton Budget Model, September 8, 2025. Analysis of 784 occupations.
https://budgetmodel.wharton.upenn.edu/issues/2025/9/8/projected-impact-of-generative-ai-on-future-productivity-growth

International Energy Agency. Energy and AI. IEA Special Report, April 2025. Comprehensive analysis of data center energy demand, supply mix, and projections to 2035.
https://www.iea.org/reports/energy-and-ai

Pew Research Center. "What We Know About Energy Use at U.S. Data Centers Amid the AI Boom." October 24, 2025. Based on IEA Energy and AI report.
https://www.pewresearch.org/short-reads/2025/10/24/what-we-know-about-energy-use-at-us-data-centers-amid-the-ai-boom/

Data Center Dynamics. "IEA: Data Center Energy Consumption Set to Double by 2030 to 945TWh." April 11, 2025. Citing Gartner projections on AI server power share.
https://www.datacenterdynamics.com/en/news/iea-data-center-energy-consumption-set-to-double-by-2030-to-945twh/

CNBC. "These Nuclear Companies Are Leading the Race to Build Advanced Small Reactors in the U.S." March 30, 2025. Covering Kairos Power, TerraPower, and Google/Amazon/Microsoft commitments.
https://www.cnbc.com/2025/03/29/these-nuclear-companies-lead-the-race-to-build-small-reactors-in-us.html

Introl. "Nuclear Power for AI: Inside the Data Center Energy Deals." January 8, 2026. Comprehensive overview of Microsoft/Three Mile Island, Google/Kairos, Amazon/X-Energy, Meta, Oracle.
https://introl.com/blog/nuclear-power-ai-data-centers-microsoft-google-amazon-2025

WWT (World Wide Technology). "Big Tech's Nuclear Bet: Key Small Modular Reactors for Cloud Power." December 2, 2025. Analysis of all major hyperscaler nuclear commitments and realistic timelines.
https://www.wwt.com/blog/big-techs-nuclear-bet-key-small-modular-reactors-for-cloud-power

GIS Reports Online. "What Is Holding Up Progress on Small Modular Reactors?" November 25, 2025. Covering U.S. nuclear strategy (Trump administration, May 2025), UK £2.5B SMR package, US-UK Atlantic Partnership for Advanced Nuclear Energy.
https://www.gisreportsonline.com/r/smrs/

Grayson, T. "Nuclear for Data Centers: Why the Gen IV SMR Timeline Is 2035." December 11, 2025. Engineering analysis of ASME material certification requirements, HALEU constraints, and FOAK risks.
https://www.tonygrayson.ai/post/nuclear-for-data-centers-smr-timeline

WWT (see [16]). On HALEU: DOE estimates 50+ metric tons/yr demand by 2035; Centrus (only U.S. licensed producer) output: ~545 kg delivered as of 2025. See also: TerraPower HALEU delay documentation in same source.

World Economic Forum. "Why We Need Space-Based Solar Power (SBSP)." October 2025. Covering beam safety, rectenna design, and economic case for SBSP in energy transition.
https://www.weforum.org/stories/2025/10/space-based-solar-power-energy-transition/

Wikipedia (continuously updated). "Space-Based Solar Power." Covering Caltech MAPLE (2023), Aetherflux (pivot to space data centers, Dec. 2025), Japan OHISAMA, ESA Solaris, Overview Energy, US DoD OECIF support for Aetherflux. https://en.wikipedia.org/wiki/Space-based_solar_power
TechCrunch. "Overview Energy Wants to Beam Energy from Space to Existing Solar Farms." December 10, 2025. https://techcrunch.com/2025/12/10/overview-energy-wants-to-beam-energy-from-space-to-existing-solar-farms/

Harvard Technology Review. "The Future of Energy: Unlocking the Potential of Space-Based Solar Power." September 5, 2025. On launch cost requirements and NASA analysis of financial barriers.
https://harvardtechnologyreview.com/2025/09/05/the-future-of-energy-unlocking-the-potential-of-space-based-solar-power/

IEEE Spectrum. "A Skeptic's Take on Beaming Power to Earth from Space." December 12, 2024. Analysis of GEO assembly requirements vs. ISS precedent; ESA Solaris program scope.
https://spectrum.ieee.org/space-based-solar-power-2667878868

ESA (European Space Agency). Solaris Programme, 2022 launch. CFC Solutions. "A Glimpse at the Future of Space-Based Solar Power." Citing ESA 2021 report projecting viability in the 2040s. https://www.nrucfc.coop/content/solutions/en/stories/energy-tech/...
NASA Office of Technology, Policy, and Strategy. Space-Based Solar Power Study. January 2024. https://www.nasa.gov/wp-content/uploads/2024/01/otps-sbsp-report-final...

National Bureau of Economic Research (NBER) Digest. "Automation Can Be a Response to an Aging Workforce." July 2018. On demographic-driven automation adoption in aging economies.
https://www.nber.org/digest/jul18/automation-can-be-response-aging-workforce

Digital Watch Observatory. "AI and Robotics Could Offset Impact of Aging Populations in Asia." March 2026. Citing Bank of Korea GDP impact study, International Federation of Robotics 2024 robot density data. https://dig.watch/updates/ai-and-robotics-aging-population-asia

Federal Reserve Bank of Dallas. "Advances in AI Will Boost Productivity, Living Standards Over Time." November 10, 2025. GDP per capita scenarios 2024–2050 including AI scenarios. https://www.dallasfed.org/research/economics/2025/0624

Carbon Brief. "AI: Five Charts That Put Data-Centre Energy Use and Emissions Into Context." September 17, 2025. Citing IEA 2025 Energy and AI report projections. https://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context/

IAEA. "Data Centres, Artificial Intelligence and Cryptocurrencies Eye Advanced Nuclear to Meet Growing Power Needs." IAEA Bulletin, 2024. https://www.iaea.org/bulletin/data-centres-artificial-intelligence-and-cryptocurrencies-eye-advanced-nuclear-to-meet-growing-power-needs

Underhyped AI. "Data Center Power Demand: Global Electricity Consumption, Supply Gaps, and the Nuclear Bet." February 16, 2026. https://underhyped.ai/articles/data-center-power-demand

AI Productivity Estimates — Key Range

0.9–1.8% GDP gain over 10 years (Acemoglu/MIT — most conservative, task-model based)

1.5% GDP gain by 2035; 3.7% by 2075 (Penn Wharton Budget Model, Sept. 2025)

+$7T Global GDP over 10 years (Goldman Sachs — most optimistic)

40% Worker productivity boost reported by AI users (EY Survey, Oct. 2025)

~0 Measurable macro-level labor market impact so far (Yale Budget Lab, 2025)

The gap between task-level and macro-level is the "productivity J-curve" — organizational restructuring takes years to show in statistics.

AI Energy Demand — Key Numbers

415 TWh — Global data center electricity demand, 2024

945 TWh — Projected 2030 (IEA Base Case) = Japan's entire grid

1,700 TWh — IEA "Lift-Off" case by 2035 (= Germany + France)

10× More electricity per query: AI vs. conventional web search

50% Of all new U.S. electricity demand growth 2024–2030: data centers

8% Estimated increase in average U.S. electricity bill by 2030 (Carnegie Mellon)

Big Tech Nuclear Commitments (2024–2025)

Microsoft + Constellation: 20-yr PPA, 837 MW, Three Mile Island restart by 2028. $1B DOE guarantee.

Google + Kairos Power: 500 MW across 6–7 molten salt SMRs. First unit 2030, full fleet 2035. World's first corporate SMR purchase agreement.

Amazon + X-Energy: $500M investment; 5 GW by 2039. Also $20B+ Susquehanna nuclear campus.

Meta: RFP for 1–4 GW new nuclear, targeting early 2030s.

Oracle: Plans for 1 GW campus powered by 3 SMRs (timeline TBD, likely mid-2030s).

Total: 10+ GW in U.S. nuclear commitments signed in 2024 alone. Industry-transforming demand signal.

Space Solar Power — Status Report

Concept origin: Peter Glaser, 1968.

2023 milestone: Caltech MAPLE — first wireless power transmission in space, signal detected on Earth. Milliwatts, not megawatts.

Current ventures: Aetherflux (pivoted to space data centers Dec. 2025), Overview Energy (2030 GEO target, $20M raised), ESA Solaris (€60M feasibility study).

SpaceX factor: Starship reduces LEO launch cost from ~$50,000/kg to ~$100–200/kg — fundamentally changes economics. Still requires 50–100 launches for 1 GW station.

Commercial viability: ESA projects 2040s with sustained investment. Most mainstream analysts: 2045–2060.

Key unresolved: Autonomous in-orbit assembly (no precedent at scale); HALEU equivalent doesn't apply but GEO mass requirements are enormous; orbital debris management.

Not a solution for the 2026–2050 demographic window.

The Fiscal Translation Problem

Payroll taxes — which fund Social Security, Medicare, and EU pension equivalents — are taxes on wages, not on output or profit.

If AI increases GDP by concentrating gains in corporate profits (capital income), payroll tax revenues do not grow proportionally.

Resolution requires either: (a) a robot/AI tax redistributing capital gains to pension systems, or (b) AI augmenting wages rather than replacing them.

Evidence: PwC finds wages rising 2× faster in AI-exposed industries (positive). But ScienceDirect life-cycle model finds labor's income share falls as automation scales (negative). The outcome is contested.

Without the fiscal bridge, even strong AI productivity gains do not rescue pension systems.