The J-Curve: LEROIS Through the AI Energy Transition, 2009–2035

Daniel Goodwin (with help from Claude Code 🤖) | May 2026

---

Abstract

Western civilization has not taken energy seriously enough. The US grid has been deteriorating for over a decade, and the artificial intelligence transition is imposing the largest single demand shock on US electricity since rural electrification. As Doomberg says, "Energy is Life"; as Vaclav Smil writes, "Energy is the only universal currency—one of its many forms must be transformed to get anything done" (Energy and Civilization: A History, 2017). The Terawatt-hours (TWh) needed are best understood through their historical relationship to functioning or failing societies, for which the Energy Return on Investment is an imperfect but best single metric from which to analyze the decade ahead.

We model the resulting trajectory of Lambert EROI Social (LEROIS) using the Lambert et al. (2014) financial-proxy method, extending a replicated 16-year historical series (2009–2024) with forward projections through 2035 under five scenarios. LEROIS matters because its values are historically correlated with civilizational function: the Human Development Index, female literacy, health expenditure, public infrastructure quality, and the Gini coefficient all degrade below specific LEROIS thresholds. To our knowledge, this is the first paper to combine EROI analysis with the AI transition forecast: how the US traverses this curve over the next decade will likely decide national growth or collapse, and the degree of social liberty the country can afford.

Our model ingests a bottom-up supply ledger of 3,283 power generation projects and 1,500 data center facilities, applies probability-weighted capacity projections, and feeds the resulting net energy balance into a price-responsive LEROIS calculation using the Lambert formula directly—no calibration factors applied. We derive our own empirical thresholds from observed macroeconomic outcomes during the 2022 European energy crisis rather than adopting Lambert's original thresholds verbatim. Data and code are available on request.

We find that the US faces a structural LEROIS J-Curve during the AI transition. From a 2024 baseline of 14.81:1, LEROIS declines under every scenario as massive energy investment collides with surging data center demand. The depth and duration of the trough depend critically on the scale and speed of new supply:

• Manhattan Project-scale intervention: trough of 12.92:1 in 2029, recovery to 20.10:1 by 2035—surpassing all historical LEROIS values; 100 GW datacenter megacluster achieved by ~2033
• Mobilization Case (enhanced-geothermal buildout): holds LEROIS within the functional-economy band throughout, with a shallow trough of 12.49:1 in 2033 and 12.94:1 by 2035
• AI efficiency gains only: monotonic decline to 9.11:1 by 2035 (no recovery within window)
• Baseline (no intervention): monotonic decline to 8.69:1 by 2035, entering the zone of industrial strain
• Disruption (2026 geopolitical shock): collapse to 6.30:1 by 2035—into the severe-distress territory observed in France and Germany during the 2022 European energy crisis

Two pathways hold LEROIS within the 12:1 functional economy band through the projection window: the full Manhattan Project-scale portfolio—combining accelerated nuclear, enhanced geothermal, fusion pilot plants, massive grid expansion, war-speed permitting, and AI-driven efficiency/productivity gains—and the more focused Mobilization Case, which bets on enhanced geothermal at scale plus AI gains. The Manhattan path additionally drives a dramatic recovery: by 2035, Manhattan LEROIS reaches 20.10, surpassing all historical values and exceeding the 2015–16 peak of ~17. The J-Curve shape emerges naturally from the physics: LEROIS must first fall as society invests enormous energy building new infrastructure, before rising as that infrastructure delivers returns. The Disruption scenario shows the downside tail—a compound geopolitical shock beginning in 2026 pushes LEROIS to 6.30, below the deindustrialization threshold entirely.

The policy implication is unambiguous: a mobilization-scale approach may be the only way to avoid spiralling chaos. The bill for the AI transition is coming due regardless—it will either be paid by the masses through rising energy costs and declining quality of life, or through a concentrated effort of Infrastructure-Led Development that creates national assets and jobs. The speed of the energy buildout determines whether the trough is a temporary valley or a permanent collapse.

Methodology is described in full; input is greatly welcome.

---

1. Introduction

1.1 The Problem: Western LEROIS Has Been Declining

Energy Return on Investment (EROI) measures the ratio of energy delivered to society versus energy expended to obtain it. When EROI is high, a small fraction of total energy output is reinvested in the energy sector, leaving the bulk available for manufacturing, transport, healthcare, education, and defense. When EROI declines, the energy sector consumes an increasing share of economic output, leaving less for everything else. The relationship is non-linear: as EROI drops from 10:1 toward 5:1, the fraction of output consumed by energy extraction doubles from 10% to 20%, and the discretionary surplus available to society shrinks disproportionately. Murphy and Hall (2010) termed this the "net-energy cliff."

Lambert, Hall, Balogh, Gupta, and Arnold (2014) established a financial-proxy method for estimating societal EROI across nations, demonstrating strong correlations between what we term LEROIS (Lambert EROI Social) and quality-of-life indicators including the Human Development Index, female literacy, health expenditure per capita, and the Gini coefficient. Their central finding: above a certain LEROIS threshold (approximately 11–14:1 on Lambert's original scale; approximately 12–15:1 on our directly-computed scale — see Section 1.4 for discussion of the divergence), societies sustain industrial civilization, including advanced manufacturing, higher education, and universal healthcare. Below it, these functions degrade.

The problem is that Western LEROIS has been going down for over a decade and will very likely tank if we do nothing. The United States currently operates in the functional economy band. Our replicated Lambert series places US LEROIS at 14.81:1 in 2024—within the 12–15:1 range we empirically associate with a functional but pressured economy, and well below the 15+:1 level associated with comfortable surplus. This is not a sudden crisis; it is a slow decay that the AI demand shock threatens to accelerate catastrophically.

1.2 The AI Energy Demand Shock

Into this already-constrained energy landscape, the AI revolution is arriving as the largest new electricity demand source in decades. Our bottom-up pipeline places US data center electricity consumption at approximately 175 TWh in 2024 (1,500 tracked facilities, loaded at a 0.85 capacity factor reflecting near-continuous AI inference). Industry projections from EPRI, Goldman Sachs, US DOE, McKinsey, and BCG span a range from 290 to 1,050 TWh by 2030 (the DOE-commissioned LBNL estimate runs to 2028), with our bottom-up pipeline analysis suggesting approximately 1,000 TWh. Leopold Aschenbrenner's Situational Awareness projects a single 100 GW training cluster by 2030—run continuously for a year, that one cluster alone would draw 876 TWh.

Figure 1 shows the supply-demand picture in stark terms.

This demand arrives against a backdrop of constrained conventional supply growth. Coal plants are retiring faster than they can be replaced. Nuclear restarts are proceeding slowly. Natural gas expansion faces permitting barriers averaging 4.5 years for major projects. Wind and solar, while growing rapidly in nameplate capacity, contribute firm-equivalent power at only 20–30% of their rated output due to intermittency and limited storage. Net new firm-equivalent supply from the entire US generation portfolio is modest—roughly 18 TWh added in 2025, rising to 150–250 TWh per year later in the decade—while data center demand alone grows by 150–600 TWh per year over the same period. Supply grows; it simply cannot grow fast enough.

The result is a widening gap between electricity supply and demand that drives energy prices higher, shifts the fuel mix toward more expensive marginal sources, and—through the Lambert formula—pushes LEROIS downward.

1.3 The J-Curve Hypothesis

We propose that the AI energy transition will produce a characteristic J-Curve in LEROIS:

1. Descent phase (2025–2029): LEROIS falls as massive construction energy is invested in new generation capacity, data center infrastructure absorbs grid supply, and scarcity drives energy prices upward. The energy sector consumes an increasing share of GDP, directly depressing the Lambert ratio.

2. Trough (2028–2032): The point of maximum pain, where net energy on the grid is most negative and prices peak. The depth and timing of the trough vary by scenario.

3. Recovery phase (2030–2035+): New capacity comes online, supply catches up with demand, prices moderate, and LEROIS begins to recover. Whether recovery occurs within the projection window—and whether it returns LEROIS above the 12:1 functional economy threshold—depends entirely on the scale of intervention.

This J-Curve is structurally analogous to an investment curve: society must spend energy now to earn energy later. The risk is that the trough is deep enough, or lasts long enough, to trigger cascading failures in energy-dependent systems before recovery arrives.

1.4 The LEROIS Metric

We adopt the notation LEROIS (Lambert EROI Social) to distinguish the Lambert et al. financial-proxy method from other EROI variants in the literature (Hall et al.'s bottom-up EROI, Brockway et al.'s exergy-based estimates, Cleveland's primary-stage EROI). The Lambert method uses macroeconomic variables—GDP, total primary energy supply, end-user delivered prices, and fuel mix shares—to compute a top-down societal EROI that captures the full cost of energy to the economy.

A note on divergence from Lambert's original values: Our replicated Lambert series does not match the original Lambert et al. (2014) numbers exactly. Lambert reported US EROI_soc of approximately 11:1 for 2009; our pipeline produces 13.81:1 for the same year. This discrepancy arises from (a) differences in delivered price data sources (Lambert used IEA data; we use EIA end-user prices), (b) GDP data vintage (Lambert used values available circa 2012; we use revised World Bank series), and (c) potential differences in the primary electricity accounting convention. The divergence is not uniform across countries (US: 1.25×, China: 2.59×, UK: 1.92×), which rules out a simple unit-conversion explanation and means that force-fitting a scalar calibration factor to match Lambert's numbers would be methodologically unsound.

Rather than calibrate, we use our raw (direct) values throughout this paper. They are internally consistent, fully transparent, and reproducible from public data. We then derive our own empirical thresholds from observed macroeconomic outcomes rather than adopting Lambert's original thresholds verbatim. The 2022 European energy crisis provides an ideal natural experiment: Germany's LEROIS of 7.29 on our scale coincided with active deindustrialization (BASF closures); the UK at 8.43 required £40 billion in emergency energy subsidies and saw pension fund near-collapse; France at 6.79 experienced fiscal stress and social unrest; the US at 12.72 experienced 9% inflation but no structural breakdown. From these observations we derive empirically-grounded thresholds on our scale (see Section 5.1).

---

2. Results

2.1 Historical LEROIS Trajectory (2009–2024)

Our replicated Lambert series shows US LEROIS oscillating within a band of 11.55–17.21:1 over the 16-year historical period, with a mean of approximately 14.1:1.

| Year | LEROIS | Notable Events | |——|——–|—————-| | 2009 | 13.81 | Post-financial crisis; energy prices depressed | | 2011 | 11.55 | Series minimum; high oil prices ($107/bbl Brent) | | 2015 | 16.00 | Oil price collapse; shale revolution lowers costs | | 2016 | 17.21 | Series maximum; peak energy surplus | | 2020 | 16.40 | COVID demand destruction lowers prices | | 2022 | 12.72 | Post-Ukraine price shock; European energy crisis | | 2024 | 14.81 | Current baseline; functional economy |

The historical series demonstrates LEROIS sensitivity to energy price movements. The 2015–2016 peak coincides with the oil price collapse that made delivered energy substantially cheaper. The 2022 dip reflects the post-Ukraine price shock. These are not merely financial artifacts: the 2022 European energy crisis produced observable deindustrialization in Germany (BASF announced permanent closure of European operations), fiscal crisis in the UK (pension fund near-collapse, emergency energy price cap), and infrastructure strain across the continent—all at LEROIS levels of 7–8.5:1 on our scale. The US weathered the same shock at 12.72—experiencing 9% inflation but no structural breakdown—providing a natural experiment for threshold derivation.

2.2 The J-Curve: LEROIS Projections Through 2035

Figure 2 presents the central result of this paper: the projected trajectory of US LEROIS under five scenarios, overlaid on the historical series.

The five scenarios produce sharply divergent outcomes:

Manhattan Project-scale (blue): Combines accelerated nuclear deployment (restarts, new builds, SMRs), enhanced geothermal at scale, fusion pilot plants (2033–2035), massive grid expansion, streamlined permitting, and AI-driven efficiency/productivity gains (grid optimization, industrial efficiency, GDP amplification). LEROIS drops from 14.81 to a trough of 12.92 in 2029, then recovers sharply—reaching 20.10 by 2035, surpassing all historical values. The recovery is driven by the compounding effects of massive new supply (2,000 TWh by 2035 including fusion), which drives electricity prices from a peak of ~$195/MWh down to ~$95/MWh, combined with AI-amplified GDP growth. LEROIS remains within the functional economy band throughout the trough. The cumulative net energy deficit is -299 TWh, the smallest of any scenario. Under this scenario, a 100 GW datacenter megacluster—the rough threshold at which a "country of geniuses in a datacenter" (Amodei 2024) becomes feasible—is achieved by approximately 2033. The figure plots the aggressive end of the Manhattan band; a conservative variant (no fusion, slower manufacturing ramp) holds LEROIS at 12.69 by 2035—still within the functional band but without the dramatic recovery.

Mobilization Case (green): A more focused intervention than the full Manhattan portfolio—it bets on a single modality, enhanced geothermal (EGS), scaled aggressively to 1,202 TWh per year of firm baseload by 2035, composed with the same AI efficiency and productivity overlay. LEROIS dips only shallowly, to a trough of 12.49 in 2033, and holds at 12.94 by 2035—remaining within the functional economy band for the entire projection window. It does not produce Manhattan's explosive recovery (no fusion, no broad nuclear portfolio), but it demonstrates that a determined buildout of one high-EROI firm resource is, on its own, enough to keep the US economy functional through the AI transition. Cumulative deficit: -2,822 TWh.

AI improvements only (orange): AI-driven efficiency gains are applied to all components of the energy system, with no new supply mobilization. AI can improve building energy efficiency by 1–3% (Google DeepMind's data center cooling system achieved 40% reduction in cooling energy; scaled nationally, similar systems could reduce total building energy by 15–25 TWh/yr). AI optimizes grid dispatch and reduces transmission losses by 2–5% (DOE-documented voltage/VAR optimization yields roughly 2–4% distribution energy savings in field demonstrations). AI accelerates materials discovery (GNoME discovered 2.2 million stable crystal structures in 2023, potentially reducing battery costs and solar cell efficiency timelines by years). AI boosts GDP productivity by 1–15% (Goldman Sachs 2023 estimates 7% global GDP increase; McKinsey 2023 estimates $2.6–4.4 trillion annually). Yet despite these real and quantifiable gains, LEROIS still declines monotonically from 14.81 to 9.11 by 2035—falling into the industrial-strain zone (the same band where the UK required emergency fiscal intervention in 2022). The reason is scale: AI-driven efficiency operates on the margin while the AI demand shock operates on the base. A 3% reduction in building energy saves ~10 TWh/yr; data center demand growth adds 100–200 TWh/yr. Cumulative deficit: -5,609 TWh.

Baseline / No intervention (crimson): Only announced and permitted projects, with probability weighting based on development status. No accelerated deployment, no policy intervention. LEROIS falls steadily from 14.81 to 8.69 by 2035, entering the industrial strain zone (8–10:1) where the UK experienced pension fund near-collapse and required emergency fiscal intervention in 2022. This is the path of current policy. Cumulative deficit: -5,609 TWh.

Disruption (2026 geopolitical shock) (dark red): The baseline pipeline subjected to a compound geopolitical and infrastructure shock beginning in 2026—a Taiwan/Strait of Hormuz crisis, Chinese export controls on critical components (gas turbines, batteries, solar), and a coordinated grid cyberattack. The shock cuts baseline supply growth by 25%, raises delivered energy prices sharply (oil +40%, gas +30%, electricity +20%), and imposes a 4% GDP penalty, all persisting through 2035 with no recovery dynamic modeled. LEROIS tracks the baseline through 2025, then steps down sharply in 2026 and declines to 6.30 by 2035—below the 7:1 deindustrialization threshold entirely, into the severe-distress territory observed in France (6.79) during the 2022 European energy crisis. Cumulative deficit: -6,405 TWh, the worst of any scenario. The Disruption case is not a forecast but a stress test: it shows how little additional shock is required to push the baseline trajectory from "industrial strain" into outright collapse.

2.3 The 100 GW Superintelligence Threshold

A 100 GW dedicated compute megacluster represents the approximate power requirement for what Amodei (2024) describes as the "country of geniuses in a datacenter" moment—the threshold at which AI systems have sufficient compute density to achieve capabilities qualitatively beyond current models. At an assumed 0.85 capacity factor, 100 GW produces approximately 745 TWh annually—roughly 17% of current US electricity generation dedicated to AI compute.

Only the Manhattan scenario clears this threshold cleanly within the projection window (approximately 2033), when its annual net energy balance turns durably positive. The Mobilization Case narrows the gap substantially—its cumulative deficit by 2035 (-2,822 TWh) is roughly half the baseline's—but does not generate enough surplus to power a full 100 GW dedicated megacluster within the window. Under the baseline and disruption scenarios, 100 GW of datacenter capacity may be announced by 2030 but will not be powered due to the supply-demand gap.

The geopolitical implications are severe. If the United States does not become a "Superintelligence Power" (a state capable of deploying ASI-scale compute infrastructure), others will. China's AI capabilities are advancing rapidly—Alibaba's Qwen alone has surpassed 700 million cumulative downloads on Hugging Face, and China's energy system, while lower LEROIS (8.79:1 in 2024 on our raw scale, versus 14.81 for the US), is expanding generation capacity at a rate the US cannot currently match. A world in which no democratic nation achieves superintelligence-scale compute is one in which the technology's governance defaults to authoritarian regimes. As Amodei and Aschenbrenner have argued, the alignment of superintelligence with democratic values requires that democracies be the ones to build it—which requires the energy to power it.

2.4 The Supply-Demand Gap

The driver of LEROIS decline is the widening gap between electricity supply growth and data center demand growth. Figure 3 shows the net energy added to the grid per year under each scenario.

All scenarios begin with net-negative energy in 2024–2026, as data center demand growth outpaces new generation. The scenarios diverge from 2027 onward:

• Manhattan crosses into positive annual net energy in 2033, driven by combined output of nuclear restarts, new builds, enhanced geothermal at scale, and fusion pilot plants—ending with the smallest cumulative deficit (-299 TWh)
• The Mobilization Case bends its annual deficit back toward zero through the enhanced-geothermal ramp, ending at a cumulative deficit of -2,822 TWh—roughly half the baseline's
• Baseline and AI-only never recover, both reaching a cumulative deficit of -5,609 TWh by 2035 (the two share an identical net energy balance—AI-only adds no new supply—and diverge only through price and GDP effects)
• The Disruption scenario fares worst, with the 2026 supply shock deepening the cumulative deficit to -6,405 TWh

2.5 Energy Price Trajectories

The supply-demand gap translates directly into energy prices through basic supply-demand economics. Figure 4 shows projected price trajectories.

Price dynamics differ markedly across scenarios:

• Manhattan: Electricity peaks at ~$195/MWh around 2027–2029, then falls dramatically to ~$95/MWh by 2035 as massive supply surplus (2,000 TWh including fusion) floods the market
• Mobilization Case: Electricity peaks near $192/MWh around 2029 and then holds roughly flat, ending at ~$190/MWh by 2035 as the enhanced-geothermal ramp keeps pace with demand
• AI-only: Electricity rises from $154/MWh (2025) to ~$296/MWh (2035)
• Baseline: Electricity rises continuously from $154/MWh (2025) to ~$302/MWh (2035)
• Disruption: Electricity jumps on the 2026 shock and climbs to ~$378/MWh by 2035, the highest of any scenario

These prices emerge from a supply-demand elasticity model calibrated to historical price sensitivity (short-run elasticity of -0.3, GDP-energy price sensitivity of -3% GDP per 100% energy price increase), bounded by observed crisis peaks.

2.6 Cross-Scenario Summary

Table 1 summarizes the key metrics across all scenarios, ordered best to worst by 2035 outcome.

Table 1. Cross-Scenario Comparison

| Scenario | Trough Year | Trough LEROIS | 2030 LEROIS | 2035 LEROIS | Cumulative Deficit (TWh) | 100 GW Achieved? | |———-|————-|—————|————-|————-|————————–|——————| | Manhattan Project-scale | 2029 | 12.92 | 13.67 | 20.10 | -299 | ~2033 | | Mobilization Case | 2033 | 12.49 | 12.92 | 12.94 | -2,822 | No (approaches) | | Manhattan (conservative) | 2029 | 12.57 | 13.12 | 12.69 | -2,651 | No | | AI-only | 2035 | 9.11 | 12.84 | 9.11 | -5,609 | No | | Baseline | 2035 | 8.69 | 12.37 | 8.69 | -5,609 | No | | Disruption (2026 shock) | 2035 | 6.30 | 8.94 | 6.30 | -6,405 | No |

The two Manhattan rows bracket an uncertainty band: the aggressive variant assumes fusion pilot plants come online 2033–2035 and a fast manufacturing ramp; the conservative variant assumes neither. Notably, the conservative Manhattan portfolio and the single-modality Mobilization Case land within ~0.3 LEROIS points of each other by 2035—a determined enhanced-geothermal buildout matches a diversified war-speed portfolio that stops short of fusion.

The threshold markers are empirically grounded in observed macroeconomic outcomes during the 2022 European energy crisis:

| LEROIS Range | Empirical Basis | Observed Conditions | |————-|—————-|———————| | >15:1 | US 2015–16 (16–17) | Low energy cost pressure, comfortable surplus | | 12–15:1 | US most years, UK/DE non-crisis | Functional economy, normal operations | | 10–12:1 | US 2022 (12.72) | 9% inflation, cost-of-living pressure, but no structural breakdown | | 8–10:1 | UK 2022 (8.43) | Pension fund crisis, £40B energy subsidies, 11% inflation | | 7–8:1 | Germany 2022 (7.29) | Active deindustrialization: BASF closures | | <7:1 | France 2022 (6.79) | Fiscal stress, social unrest, industrial strain |

These are not theoretical thresholds: they are observations of what happened to advanced industrial economies at specific LEROIS levels on our scale during a real crisis.

---

3. Discussion

3.1 The Structural Nature of the J-Curve

The J-Curve is not a modeling artifact—it is a structural consequence of the physics of energy transitions. Building new energy infrastructure requires energy. A nuclear plant consumes energy during construction (steel, concrete, transport, assembly) before it generates any. An enhanced geothermal well requires drilling energy before it produces heat. Even a solar farm requires embodied energy in silicon purification, panel manufacturing, and grid interconnection before its first megawatt-hour.

During a transition, society is simultaneously:

• Consuming energy from existing sources at declining EROI (as conventional reserves deplete)
• Investing energy in new sources that have not yet begun producing
• Paying higher prices for marginal supply as demand outstrips available generation

The Lambert formula captures all three effects: construction energy increases E_T (the denominator), scarcity raises prices (increasing the weighted price-per-MJ term), and energy diverted to construction reduces the surplus available for GDP-generating activity (depressing the numerator).

The J-Curve resolves only when new capacity delivers enough energy, at low enough cost, to reverse all three trends simultaneously. This is why the speed of the buildout matters more than its ultimate scale: a slow transition means a longer time in the trough, where cascading failures become increasingly likely.

3.2 Why AI Efficiency Alone Is Insufficient

The AI-only scenario is instructive. There are credible, quantified pathways by which AI improves every component of the Lambert formula:

GDP productivity: Goldman Sachs (2023) estimates generative AI could raise global GDP by 7%. McKinsey (2023) projects $2.6–4.4 trillion in annual value. Acemoglu (2024) is far more conservative: at most a ~0.66% TFP gain over 10 years, with his preferred estimate below 0.53%. These gains enter the LEROIS numerator directly.

Energy efficiency: Google DeepMind's data center cooling AI reduced cooling energy by 40%. Scaled to the national building stock, similar systems could reduce total commercial building energy by 10–15%. DOE-documented voltage/VAR optimization yields ~2–4% distribution energy savings in field deployments. US transmission and distribution losses run roughly 5% of generation (~190–205 TWh/yr); AI-optimized dispatch could recover a portion.

Accelerated deployment: AI-driven permitting (automated environmental review, optimized site selection) could reduce project timelines by 1–2 years. AI materials science (Google DeepMind's GNoME: 2.2 million new stable crystal structures discovered in 2023) accelerates next-generation battery, solar cell, and superconductor development.

Fuel mix optimization: AI can shift industrial loads to off-peak hours, optimize renewable dispatch, and reduce curtailment. Curtailment in major US ISOs has reached roughly 5–9% of available wind and solar generation in recent years; AI-optimized dispatch could recover 10–15 TWh/yr.

Yet the AI-only scenario still declines to 9.11:1 by 2035—falling through the emerging-stress zone and into industrial strain (8–10:1), the band where the UK required emergency fiscal intervention in 2022. The efficiency gains operate on the margin while the demand shock operates on the base. A 3% reduction in building energy consumption saves ~10 TWh/yr. Data center demand growth is adding 100–200 TWh/yr. The efficiency gains are real but outmatched by an order of magnitude.

Policy implication: AI cannot solve the energy problem it creates through efficiency alone. Supply-side intervention—new generation capacity at massive scale—is necessary.

3.3 The Manhattan Scenario: What War-Speed Looks Like

The Manhattan scenario is the pathway that produces the strongest recovery—dramatic, reaching 20.10 by 2035, well beyond any historical value. (The Mobilization Case also holds LEROIS within the functional band, but without the explosive upside; see Section 2.2.) The full Manhattan portfolio requires:

• Nuclear restarts and new builds: Bringing shuttered plants back online (Three Mile Island Unit 1, Palisades) and accelerating new construction, targeting ~30 GW of baseload capacity by 2035
• Enhanced geothermal at scale: The US Department of Energy has assessed US enhanced-geothermal technical potential at over 5 TW (5,157 GW)—orders of magnitude beyond current deployment. Achieving even 5–10% of this potential (200–400 GW) within a decade would transform the supply picture. Enhanced Geothermal Systems (EGS) with 0.90 capacity factor and firm baseload characteristics could provide 1,000+ TWh/yr—more than the entire current data center demand projection.
• Fusion pilot plants (2033–2035): Under Manhattan-scale investment and permitting acceleration, fusion pilot plants begin contributing 20–300 TWh/yr from 2033, adding high-EROI baseload that compounds the recovery
• Streamlined permitting: Reducing the average permitting timeline from 4–7 years to 1–2 years for energy infrastructure
• Massive grid expansion: Doubling transmission capacity over the decade to interconnect new generation to demand centers
• AI-driven efficiency and productivity: Grid optimization (5%), industrial efficiency (8%), building efficiency (4%), transport efficiency (3%), and a 3% GDP productivity boost

This is not a normal infrastructure program. The scale is comparable to the US mobilization for World War II, when the country built ~297,000 aircraft and ~86,000 tanks and converted its entire industrial base in under four years. The energy equivalent: deploying roughly 2,000 TWh of new generation capacity by 2035—including nuclear, geothermal, and fusion—approximately tripling the current rate of annual capacity additions.

Historical precedent suggests this is physically possible but requires a command-economy approach to permitting, financing, and resource allocation that has not been demonstrated in the US since 1945. The interstate highway system, often cited as a peacetime analogue, took 35 years. The current permitting regime produces average timelines of 4.5 years for major transmission projects. The Manhattan scenario requires cutting this by two-thirds.

It is worth noting that this observation fits the LEROIS data itself: liberal democracy as we know it does not historically occur at the lower LEROIS values anyway. Lambert et al.'s cross-national data shows that below ~12:1 on our scale, governance quality, civil liberties, and democratic participation metrics all decline sharply. The difference between accidentally arriving at low-LEROIS authoritarianism through energy neglect versus deliberately accepting temporary command-economy mobilization with a defined goal and exit ramp is the difference between collapse and controlled investment. We would be steering into it with a target in mind—the 100 GW threshold—rather than sliding into it by default.

3.4 The Time Dimension: Trough Depth vs. Duration

A critical finding is that the Manhattan and Mobilization Case scenarios achieve defined troughs (2029 and 2033 respectively) followed by recovery or stabilization, while the baseline, AI-only, and disruption scenarios are still declining at the end of the projection window. For those three, the actual troughs lie beyond 2035—our model understates their ultimate severity.

This matters for policy. Manhattan's trough in 2029 means ~4 years of declining LEROIS before recovery. A society can sustain 4–5 years of economic stress—the US sustained the Great Depression for roughly that duration before New Deal infrastructure spending produced recovery. But baseline/AI-only/disruption scenarios mean 10+ years of sustained decline, long enough for institutional decay, capital flight, and political instability to become self-reinforcing.

In follow-up analyses, we will model specific energy modalities (nuclear, geothermal, advanced solar) and their individual impact on trough timing and depth. The core insight of this paper is simpler: if we do a concerted effort, we'll hit a local minimum and recover. If we don't, there is no local minimum—just continued descent.

3.5 The Manhattan Approach May Be the Only Way to Avoid Spiralling Chaos

The argument for a Manhattan-scale approach is not merely that it produces the best outcome—it may be the only approach that avoids a self-reinforcing downward spiral.

The mechanism is straightforward: declining LEROIS raises energy costs, which reduces economic growth, which reduces tax revenue, which reduces the state's capacity to fund infrastructure, which further reduces energy investment, which further lowers LEROIS. This is the "energy trap" identified by Tom Murphy (2011)—a situation where the energy needed to build new infrastructure exceeds what a declining-EROI economy can spare.

The bill for the AI transition is going to be due. It is either going to be paid by the masses—through rising electricity costs, declining public services, infrastructure decay, and reduced quality of life—or through a concentrated effort of Infrastructure-Led Development for the creation of national assets and jobs. The activists who oppose energy infrastructure development are, paradoxically, protesting toward the outcome in which ordinary people bear the cost most acutely. A Manhattan-scale buildout creates construction jobs, produces assets with 40–60 year lifespans, and generates power that reduces costs for everyone. Inaction produces the same demand growth (AI companies will simply bid electricity away from residential consumers) with none of the supply-side benefit.

3.6 The Security Asymmetry of a Capability Gap

Section 3.5 describes the economic downside of a deep, prolonged trough: the energy trap, in which a low-LEROIS economy can no longer afford the energy required to rebuild. But a sustained AI capability gap is not only an economic disadvantage. It is a standing security vulnerability—and, unlike an energy shortfall, one an adversary can actively exploit. Falling behind on superintelligence is not the passive condition of being slower. It hands a more capable rival the tools to ensure the gap never closes.

Asymmetric autonomous force. State authority rests on a monopoly on organized violence, and that monopoly has always depended on the state being able to out-resource any challenger—to field more soldiers, vehicles, and firepower than a rival could afford. AI-piloted autonomous weapons collapse that cost asymmetry. In Ukraine, first-person-view drones assembled for $300–$500 each have become the dominant weapon of the war—a US Army assessment attributes 60–80% of all 2025 combat casualties to unmanned systems, displacing artillery and armor and inverting a cost logic of mechanized warfare that had held since 1940. In June 2025, Ukraine's Operation Spider Web used 117 drones smuggled deep into Russia to strike airbases across the country's interior, damaging or destroying an estimated forty-plus aircraft for a claimed $7 billion in losses—strategic bombers, several no longer in production, eliminated by drones a state could buy by the thousand. Those drones were still human-piloted. The next step is autonomy: AI terminal guidance that defeats electronic jamming, and swarm coordination that removes the one-operator-one-drone limit entirely, so that fielded force scales with available compute rather than with manpower (CSIS, 2025). A society that cannot field comparable counter-autonomy faces a world in which a small, well-resourced actor—state or non-state—can project organized violence that once required an army. The local monopoly on force, on which every other institution depends, becomes contestable.

Precision sabotage of the technical base. The second mechanism is quieter, and for the J-Curve thesis more dangerous. In 2026, security researchers disclosed FAST16, a sabotage framework compiled in 2005—five years before Stuxnet. FAST16 did not steal data or crash systems. It subtly corrupted the numerical output of engineering and simulation software—crash-test, structural, and hydrodynamic modeling tools—by scaling values in the programs' internal arrays just enough to make the results quietly wrong. It went undetected for twenty-one years. Now project that capability forward to a superintelligence-scale adversary: sabotage that is adaptive, scalable, and aimed not at a single system but at the innovation base itself—chip-design verification, materials simulation, drug-discovery models, the numerical tools on which technical progress depends. The result would be a society in which nothing quite works—prototypes that fail for reasons no one can find, research that does not replicate, projects that run slightly and inexplicably wrong—with the cause never identified, because you cannot debug what you cannot detect. The West could be held in a technological stone age by an adversary it could not see.

These two mechanisms reveal why the framing in Section 2.3—that if no democratic nation reaches the threshold, the technology's governance "defaults to authoritarian regimes"—understates the stakes. The danger is not only who governs superintelligence. A decisive and durable AI advantage is itself a standing instrument for holding rivals down: militarily, by making their monopoly on force contestable; and technologically, by making their progress silently fail to compound. The second of these strikes the J-Curve exactly where recovery lives. LEROIS recovery depends on the GDP/E_T term—economic intensity, the value extracted from each unit of energy—which is precisely what technical progress raises and precisely what precision sabotage degrades. A capability gap does not merely lose the race. It can hand an adversary the means to ensure the race is never re-run. Seen this way, the 100 GW threshold is not only an opportunity to lead; failing it is an enduring, exploitable vulnerability.

---

4. Conclusion

The AI energy transition presents the United States with a civilizational J-Curve in LEROIS. This is not a prediction of doom—it is a description of the physics of energy transitions, quantified through an established methodology and validated against observed macroeconomic outcomes.

First, the descent is unavoidable. Under every scenario modeled, US LEROIS declines from its current level of 14.81:1 over the next 3–5 years. The AI energy demand shock is already in motion, and new generation capacity cannot be built fast enough to prevent an initial decline. The question is not whether LEROIS falls, but how far and for how long.

Second, recovery is possible but requires supply-side intervention. Two pathways hold LEROIS within the 12:1 functional economy band through 2035: the full Manhattan Project-scale portfolio, whose recovery is dramatic—reaching 20.10 by 2035, well beyond any historical value—and the more focused Mobilization Case, which holds LEROIS near 12.9 through an enhanced-geothermal buildout. AI efficiency gains alone are insufficient—the AI-only scenario declines to 9.11, into the industrial-strain zone. Doing nothing produces steady decline to 8.69, the same industrial strain zone where the UK required emergency fiscal intervention in 2022. The margin is stark: the difference between Manhattan (recovery to 20.10) and baseline (decline to 8.69) is approximately 2,000 TWh of new supply including fusion pilot plants, plus AI-driven efficiency gains, delivered within a decade.

Third, the 100 GW superintelligence threshold creates a geopolitical imperative. Achieving ASI-scale compute requires energy infrastructure that only a mobilization-scale approach can deliver within the relevant timeframe. If no democratic nation achieves this threshold, the technology's governance defaults to authoritarian states.

Fourth, the downside risk is already baked into the baseline—and a shock makes it far worse. Even without compounding negative shocks, the baseline trajectory reaches 8.69—the zone where the UK experienced pension fund near-collapse and required emergency fiscal intervention in 2022. The 2022 European crisis provides our empirical benchmarks: Germany at 7.29 began industrial exit; the UK at 8.43 required emergency fiscal intervention; France at 6.79 experienced social unrest. The baseline is the crisis scenario. The Disruption scenario quantifies what "any additional disruption" means: a compound geopolitical shock beginning in 2026—a Taiwan/Hormuz crisis, Chinese component export controls, a grid cyberattack—drives LEROIS to 6.30 by 2035, below the deindustrialization threshold entirely and into the severe-distress territory France occupied in 2022.

The policy implication is unambiguous: the speed of the energy buildout is the single most important variable determining whether the AI transition strengthens or weakens American civilization. Every year of delay deepens the trough, extends the recovery timeline, and increases the probability of cascading failures. The choice between a temporary valley and a permanent collapse is being made now, in permitting offices, capital allocation decisions, and legislative chambers. The physics is fixed. The policy is not.

---

5. Methods

5.1 The Lambert Financial-Proxy Formula

We use the Lambert et al. (2014) financial-proxy formula for societal EROI in its algebraically simplified form:

$$\text{LEROIS} = \frac{\text{GDP}}{E_T} \times \frac{1}{\sum_i \eta_i \times \text{price\_per\_MJ}_i}$$

Where:

• GDP: Gross Domestic Product in current USD
• E_T: Total Primary Energy Supply in MJ
• η_i: Share of fuel i in TPES (dimensionless; sum to 1.0)
• price_per_MJ_i: End-user delivered price of fuel i in USD/MJ

The formula decomposes LEROIS into two independent terms:

1. Economic intensity (GDP / E_T): dollars of GDP produced per megajoule of primary energy consumed
2. Energy price index (1 / Σ(η_i × price_per_MJ_i)): the inverse of the weighted average cost of energy in the mix

Five fuel categories follow Lambert's original decomposition: coal, oil, natural gas, primary electricity (nuclear + hydro + renewables), and combustible renewables/biofuels.

No calibration factors are applied. Our raw pipeline output for the US in 2009 is 13.81:1, compared to Lambert's published 11.0:1. The divergence is non-uniform across countries (US: 1.25×, China: 2.59×, UK: 1.92×), ruling out a simple unit-conversion explanation. Force-fitting a scalar calibration factor would be methodologically unsound. Instead, we use our direct values—which are internally consistent, fully transparent, and reproducible—and derive empirical thresholds from observed macroeconomic outcomes (see below).

Empirical thresholds. Rather than adopt Lambert's original thresholds verbatim (which were derived on a different numerical scale), we derive our own from the 2022 European energy crisis—a natural experiment in which countries at measured LEROIS levels experienced quantifiable economic outcomes:

| Threshold | LEROIS | Empirical Basis | |———–|——–|—————–| | Comfortable surplus | >15:1 | US 2015–16: low cost pressure, strong growth | | Functional economy | 12–15:1 | US most years: normal operations | | Emerging stress | 10–12:1 | US 2022 at 12.72: 9% inflation, no breakdown | | Industrial strain | 8–10:1 | UK 2022 at 8.43: pension crisis, £40B subsidies | | Deindustrialization | 7–8:1 | Germany 2022 at 7.29: BASF closures | | Severe distress | <7:1 | France 2022 at 6.79: fiscal stress, social unrest |

5.2 Historical Series (2009–2024)

The historical LEROIS series is computed year-by-year from:

• GDP: World Bank national accounts (current USD)
• Energy mix: Our World in Data / Energy Institute Statistical Review (TWh by source, converted to MJ)
• Prices: End-user delivered prices from EIA (US), Eurostat (EU), IEA (international)

No calibration factors are applied. Values are the direct output of the Lambert formula using contemporary data sources.

End-user delivered prices—not commodity spot prices—are used throughout, consistent with Lambert's methodology. The delivered price includes transport, distribution, storage, and retail margins, reflecting the actual cost the economy pays for energy.

5.3 Supply-Demand Model

The forward-looking model ingests a 23-sheet Excel workbook containing:

• Project Ledger: 3,283 individual power generation projects with MW capacity, commercial operation date, development status, and modality classification
• DC Demand Ledger: 1,500 individual data center facilities with MW capacity, operational status, and expected online date
• 15 modality sheets: Detailed per-modality projections (geothermal, wind, solar+LDS, hydro, offshore, coal, diesel, reciprocating gas, aero gas, heavy gas, biomass, nuclear available, nuclear restart, nuclear new, SMR)

Each project's contribution to supply is computed as:

$$\text{TWh} = \text{MW} \times \text{CF} \times \text{FE} \times 0.00876 \times P(\text{realization})$$

Where P(realization) ranges from 1.00 (operational) to 0.05 (MOU/LOI):

| Development Status | P(realization) | |——————-|—————-| | Operational | 1.00 | | Under construction (>50%) | 0.90 | | Under construction | 0.85 | | Permitted/financed/contracted | 0.70 | | Regulatory approved | 0.50 | | Announced with site | 0.30 | | Announced, no site | 0.10 | | MOU/LOI | 0.05 |

Firm-equivalence factors account for the intermittency of renewable sources:

| Modality | Capacity Factor | Firm-Equivalence | Notes | |———-|—————-|——————|——-| | CCGT (heavy gas) | 0.544 | 1.00 | Dispatchable baseline | | Nuclear | 0.920 | 1.00 | Baseload | | Coal | 0.395 | 1.00 | Dispatchable, declining | | Geothermal (EGS) | 0.900 | 1.00 | Firm, baseload | | Solar + LDS | 0.201 | 0.30 | Contested (NREL: 0.40–0.50; LBNL: 0.15–0.25) | | Onshore Wind | 0.336 | 0.20 | Limited peak correlation | | Offshore Wind | 0.420 | 0.25 | Minimal deployment through 2030 | | Hydro | 0.400 | 0.85 | Dispatchable, water-year dependent |

The workbook provides supply and demand figures for the full 2024–2035 window directly; values for 2031–2035 are curve-fit extrapolations of the 2026–2030 bottom-up pipeline, computed within the workbook and ingested as-is.

5.4 Price Model

Energy prices are projected using a supply-demand elasticity model:

1. Compute net energy balance for each year (supply growth + scenario additions - DC demand)
2. Convert supply-demand gap to price premium using historical short-run elasticity (-0.3)
3. Apply GDP feedback: higher energy prices reduce GDP growth (sensitivity: -3% GDP per 100% energy price increase)
4. Bound prices by historical crisis peaks (Germany 2022: $350/MWh electricity, $45/MMBtu gas)

5.5 LEROIS Projection

For each projected year (2025–2035) and each scenario:

1. Start with 2024 baseline energy mix (coal: 8.3%, oil: 37.8%, gas: 34.2%, primary electricity: 17.7%, biofuel: 2.0%)
2. Evolve mix forward: coal declines ~0.5 pp/yr, primary electricity increases ~0.9 pp/yr
3. Apply scenario-specific prices from the price model
4. Use GDP projections (IMF WEO April 2025 baseline, ~2.5% nominal growth, with energy-price feedback)
5. Use TPES projections (EIA AEO 2025 baseline, ~0.8% annual growth)
6. Compute LEROIS using the Lambert formula directly (no calibration factor)

5.6 Scenarios

| Scenario | Description | Key Assumptions | |———-|————-|—————–| | Manhattan Project-scale | All levers at maximum: nuclear + geothermal + fusion + grid + permitting reform + AI gains | Nuclear restarts + new builds + geothermal + fusion pilots (2033–2035) + 50% permitting acceleration + AI efficiency/GDP gains. Aggressive and conservative variants bracket an uncertainty band (the conservative variant assumes no fusion and a slower manufacturing ramp). | | Mobilization Case | Single-modality enhanced-geothermal buildout, composed with the AI overlay | Enhanced Geothermal Systems (EGS) scaled to 1,202 TWh/yr of firm baseload by 2035 (CF 0.90, firm-equivalence 1.0) plus AI efficiency/GDP gains. Composed at load time from the geothermal and AI-only scenario definitions. | | AI-only | Baseline + AI efficiency/productivity gains | 1–3% energy efficiency, 1–15% GDP boost, lower generation costs. No new supply mobilization. | | Baseline | Announced projects only, probability-weighted | No intervention beyond current pipeline | | Disruption | Baseline subjected to a compound 2026 geopolitical/infrastructure shock | From 2026 onward: -25% supply growth, -10% DC demand, oil +40% / gas +30% / electricity +20% / coal +15%, -4% GDP. Shock persists through 2035 with no recovery modeled. Years before 2026 track the baseline exactly. |

5.7 Software and Reproducibility

The computational pipeline is implemented in Python as the jcurve module. It reads directly from the workbook (xlsx format via openpyxl/pandas), applies zero hardcoded fallbacks, and produces all figures programmatically. Every constant is defined in a central config.py file. The pipeline is designed to be audit-ready.

python -m jcurve                     # All scenarios, all figures
python -m jcurve --scenario baseline # Single scenario
python -m jcurve --validate          # Validate workbook parsing
python -m jcurve --output-json out.json  # Export all results

Data and code are available on request.

---

Appendix A: LEROIS Decomposition Analysis

What drives the J-Curve: price effect, mix effect, and GDP residual.

The Lambert formula decomposes into three independent drivers of LEROIS change: (1) the price effect—changes in delivered energy prices; (2) the mix effect—shifts in fuel mix composition; and (3) the GDP/energy residual—changes in economic intensity.

The decomposition reveals that:

1. Price effects dominate the descent—then drive the recovery: Rising energy prices account for the largest share of LEROIS decline in every scenario during the trough years. In Manhattan, the price effect reaches -2.8 points by 2029 before reversing: by 2035, the price effect is positive (+2.2 points) as massive new supply drives electricity prices to ~$95/MWh, well below the baseline of ~$302/MWh. In the baseline, the price effect deepens continuously to -6.3 points by 2035 with no reversal; in the disruption scenario it reaches -8.4 points.

2. Mix effects contribute secondarily: The shift from cheaper fuels (coal, oil) toward more expensive primary electricity accounts for roughly -0.8 to -1.5 LEROIS points across all scenarios, growing over time as electrification increases the electricity share of TPES.

3. GDP growth amplifies the recovery: The GDP/energy residual—reflecting steady ~2.5% nominal growth plus AI productivity gains—provides a contribution growing to +4.6 by 2035 in the Manhattan scenario. The combination of falling prices and rising GDP produces the recovery: Manhattan's LEROIS gains +7.2 points from its 2029 trough to 2035, driven roughly equally by price relief and GDP/productivity growth.

---

Appendix B: UK LEROIS — A Skeptical Analysis

Our pipeline produces a UK LEROIS of 15.39 for 2024—which seems implausibly high given observable UK conditions (NHS in crisis, infrastructure decay, real wages stagnant since 2008, energy price cap required since 2022). Note that even in the crisis year of 2022, our pipeline produces UK LEROIS of 8.43—which does align with the observable industrial strain (see Appendix C). This appendix unpacks why the non-crisis values may overstate UK conditions.

B.1 What Produces the High Number

The Lambert formula for the UK in 2024:

• GDP: £2.72 trillion (~$3.4 trillion USD)
• TPES: ~7.0 EJ
• Mix: Gas-heavy (38%), oil (35%), primary electricity (20%), coal (3%), biofuel (4%)
• Prices: End-user delivered prices relatively moderate due to the energy price cap (suppressing the gas price signal)

The UK's relatively high GDP per unit of energy consumed (due to its service-heavy, low-manufacturing economy) produces a favorable first term (GDP/E_T). The UK simply does not use much energy per dollar of GDP because it does not manufacture much of anything. The second term (energy price index) benefits from the regulated price cap which holds consumer energy costs below market-clearing levels—but this is an artificial suppression, funded by taxpayer subsidy rather than genuine market abundance.

B.2 Why We Are Skeptical

1. The price cap is a fiscal transfer, not energy abundance: The UK's Energy Price Guarantee costs taxpayers approximately £40 billion. The Lambert formula sees low consumer prices and interprets them as favorable EROI. In reality, the energy is expensive but the cost is hidden in government debt. This is a measurement artifact: the formula measures delivered prices, not the true cost to society including the fiscal burden.

2. Service economy illusion: A country that imports its manufactured goods (steel from China, chemicals from Germany, electronics from Asia) does not account for the embodied energy of those imports in TPES. The UK's low energy intensity reflects outsourced manufacturing, not genuine efficiency. Adjusting for embodied energy in trade would substantially increase E_T.

3. Observable conditions contradict 15.39: On our scale, 15+ corresponds to "comfortable surplus" (anchored to US 2015–16). The UK's current reality—150,000+ NHS wait-list deaths above pre-pandemic baseline, 4% of bridges at "substandard" condition nationally (with hotspots above 15% in some councils), real wages flat since 2008, declining life expectancy in deprived areas—is inconsistent with a LEROIS in the comfortable surplus zone.

4. UK steel production has collapsed: From ~17.8 million tonnes in 1990 to roughly 4 million by 2024—the lowest output since the Great Depression—the UK has lost industrial capacity that the LEROIS metric would expect at higher values.

B.3 Assessment

We retain the UK LEROIS of 15.39 in our dataset as the mathematical output of the Lambert formula, but flag it as likely overstating the UK's true energy-socioeconomic position due to:

• Artificial price suppression via fiscal transfer (price cap)
• Embodied energy trade deficit not captured in TPES
• Service economy distortion of GDP/E_T ratio

A corrected estimate accounting for fiscal cost of the price cap (~£40bn added back to delivered prices) and estimated embodied energy of net imports (~2.5 EJ) would place UK LEROIS closer to 10–12:1, which is more consistent with observable conditions. Notably, the UK's crisis-year value (8.43 in 2022) does align well with observed outcomes—suggesting the Lambert formula works best when prices reflect true market costs rather than being distorted by fiscal transfers. Future work should develop a trade-adjusted LEROIS methodology.

---

Appendix C: Lambert Divergence and Threshold Derivation

C.1 Our Numbers vs. Lambert's Original

Our replicated LEROIS values for 2009 (the overlap year with Lambert et al. 2014) differ from the original:

| Country | Lambert (2009) | Our Raw (2009) | Ratio | |———|—————|—————-|——-| | USA | 11.0 | 13.81 | 1.26× | | China | 3.2 | 8.29 | 2.59× | | Germany | 8.0 | 12.80 | 1.60× | | France | 10.0 | 11.85 | 1.19× | | UK | 9.0 | 17.30 | 1.92× |

The divergence stems from:

• Price data sources: Lambert used IEA/OECD delivered price data (circa 2012 publication); we use EIA, Eurostat, and national statistical office data compiled in 2025. Energy accounting conventions have changed.
• GDP data vintage: Lambert used GDP data available at time of writing; we use revised World Bank series.
• Primary electricity accounting: The "primary equivalent" convention for nuclear and renewables varies between IEA and national accounts.

Critically, the divergence is non-uniform across countries (ranging from 1.19× to 2.59×). This rules out a simple unit-conversion or systematic bias explanation and means that applying a country-specific calibration scalar to match Lambert's values would be fitting five free parameters to five data points—a calibration with no predictive power and no physical justification.

C.2 Why We Derive Our Own Thresholds

Rather than calibrate our values to Lambert's scale and adopt his thresholds, we derive empirical thresholds directly from observed macroeconomic outcomes at measured LEROIS levels on our scale. The 2022 European energy crisis provides the natural experiment:

• US at 12.72: 9.1% CPI inflation, cost-of-living crisis, but no structural economic breakdown. The economy continued to function.
• UK at 8.43: Pension fund near-collapse requiring Bank of England emergency intervention. Energy Price Guarantee costing £40 billion in taxpayer subsidies. 11.1% inflation. NHS crisis deepening.
• Germany at 7.29: Active deindustrialization begins. BASF announced closure of European chemical operations. Energy-intensive manufacturing relocating permanently.
• France at 6.79: Fiscal stress, social unrest, industrial strain.

These outcomes validate Lambert et al.'s core insight—that LEROIS thresholds correspond to civilizational function—while establishing those thresholds on our own numerical scale rather than through force-fitting.

---

References

Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487.

Amodei, D. (2024). Machines of Loving Grace. darioamodei.com/essay/machines-of-loving-grace.

Aschenbrenner, L. (2024). Situational Awareness: The Decade Ahead. situational-awareness.ai.

Brockway, P. E., Owen, A., Brand-Correa, L. I., & Hardt, L. (2019). Estimation of global final-stage energy-return-on-investment for fossil fuels with comparison to renewable energy sources. Nature Energy, 4, 612–621.

Cleveland, C. J. (1992). Energy quality and energy surplus in the extraction of fossil fuels in the U.S. Ecological Economics, 6(2), 139–162.

EPRI (2024). Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption. Electric Power Research Institute.

Goldman Sachs (2023). Generative AI Could Raise Global GDP by 7 Percent. Goldman Sachs Research.

Google DeepMind (2023). Scaling Deep Learning for Materials Discovery. Nature, 624, 80–85.

Hall, C. A. S., Lambert, J. G., & Balogh, S. B. (2014). EROI of different fuels and the implications for society. Energy Policy, 64, 141–152.

Lambert, J. G., Hall, C. A. S., Balogh, S., Gupta, A., & Arnold, M. (2014). Energy, EROI and quality of life. Energy Policy, 64, 153–167.

McKinsey Global Institute (2023). The Economic Potential of Generative AI: The Next Productivity Frontier. McKinsey & Company.

Murphy, D. J., & Hall, C. A. S. (2010). Year in review—EROI or energy return on (energy) invested. Annals of the New York Academy of Sciences, 1185, 102–118.

Murphy, T. W. (2011). The Energy Trap. Do the Math (UCSD), 18 October 2011. dothemath.ucsd.edu/2011/10/the-energy-trap/.

Smil, V. (2017). Energy and Civilization: A History. MIT Press.