The Biophysical Trap: Why AI Cannot Solve the Energy Crisis It Creates

PART 1.4: The Biophysical Trap: Why AI Cannot Solve the Energy Crisis It Creates#

The energy crisis described in Section 1.1 is not a temporary infrastructure bottleneck. It is a collision with thermodynamic reality.

The Economy Is Not a Monetary System—It Is an Energy System#

There is a school of economics that most mainstream economists refuse to take seriously. It is called biophysical economics, and its founding figure is the systems ecologist Charles A.S. Hall. Its central claim is deceptively simple: money is not the primary constraint on an economy. Money is a claim on energy and resources. You can print dollars. You cannot print joules.[1]

The evidence for this is not subtle. Since the Industrial Revolution, every sustained expansion of economic output has moved in near-lockstep with an expansion of energy throughput. There is no historical example of a large economy that grew its real output for decades while shrinking its total energy use in absolute terms. When policymakers talk about “decoupling” economic growth from resource consumption, they are describing something that has never happened at scale—and that the physics of the problem suggests never will. Ecological economists Jason Hickel and Giorgos Kallis reviewed the empirical record and concluded that absolute decoupling of GDP from resource use, at the rate and duration a green-growth future would require, is not supported by the evidence.[2]

This is the first trap: the belief that we can grow AI capabilities indefinitely without corresponding growth in energy consumption.

The EROI Cliff: The Invisible Constraint on Civilization#

What EROI Measures#

Energy Return on Investment (EROI) is the fundamental metric of energy quality. It answers one question: for every unit of energy you spend acquiring energy, how many units do you get back? Hall and his collaborators have spent decades assembling these ratios across fuels.[3]

Historical fossil-fuel EROI (at the point of extraction):

1930s–1950s oil: roughly 80:1 to 100:1 (for every barrel spent drilling, you got 80–100 back)
1970s oil: roughly 35:1
2010s conventional oil: roughly 18:1
2020s shale oil: roughly 5:1 to 7:1

Renewable EROI (harmonized meta-analyses):

Solar PV: the range is wide and boundary-dependent—from as low as ~2.5:1 in pessimistic buffered analyses to ~11–12:1 in harmonized meta-studies (Bhandari et al., 2015)[4]
Wind: ~16–20:1 (lower once storage and backup are added)
Hydroelectric: very high (Hall & Lambert put the mean near 84:1) but geographically limited, with most large sites already exploited

A crucial caveat destroys the lazy version of this argument. The high fossil numbers above are measured at the wellhead. When you measure EROI at the useful stage—the energy actually delivered as motion, heat, or light after refining and conversion losses—the picture changes sharply. A 2024 study in Nature Energy by Aramendia, Brockway and colleagues estimates the useful-stage EROI of fossil fuels at only about 8.5:1 to 14:1, far below the headline extraction figures.[5] That finding collapses much of the supposed gap between fossil fuels and renewables.

So the honest claim is not “renewables are net-energy sinks.” That is cherry-picking from discredited outlier studies. The honest claim is that the energy surplus margin is tightening—the comfortable buffer that built the modern world is thinner than the extraction-stage numbers imply, on both sides of the ledger.

Why This Matters: The Surplus Energy Cliff#

The EROI ratio determines how much surplus energy is left over for society after paying the energy cost of acquiring energy.

High EROI (1950s oil at ~80:1): spend 1 unit, get 80 back—79 units of surplus for everything else: manufacturing, healthcare, education, leisure, consumption.

Low EROI (a buffered renewable system at ~10:1 falling toward single digits at the useful stage): the surplus shrinks toward a single-digit multiple.

The energy analyst Nate Hagens has spent years arguing that this surplus is the invisible foundation beneath modern complexity—the reason a society can afford data centers, global supply chains, advanced medicine and universal schooling at the same time. You cannot run a high-complexity civilization on an energy system that barely produces a surplus. The math does not permit it.

The Maintenance Trap of Complexity#

As civilization becomes more complex, the maintenance energy cost of that complexity grows. This is the hidden, lurking demand required just to keep existing systems running—what Hagens and others describe as the metabolic overhead of a technological society.

Examples of complexity-maintenance cost:

Software updates and patches for billions of devices
Cybersecurity infrastructure defending against ever-growing threats
AI model retraining as data distributions shift
Grid management as renewable intermittency requires sub-second balancing
Supply-chain logistics for globally distributed just-in-time manufacturing

Every layer of technological complexity adds overhead. AI infrastructure does not replace human systems—it adds a new layer on top of them, requiring:

Energy to run the models
Energy to cool the data centers
Energy to manufacture and replace hardware on 3–5 year cycles
Energy to train successor models as current ones are superseded
Energy to manage the grid instability created by massive compute loads

We are adding AI layers to our infrastructure at the exact moment the surplus energy that funds all maintenance is under pressure. It is like adding another floor to a building while the foundation is being questioned.

The “Green AI” Illusion: Marketing vs. Thermodynamics#

The Marketing Narrative#

Tech companies promote AI as a climate solution:

“Smart grids” optimizing renewable integration
“Precision agriculture” reducing fertilizer use
“AI-driven logistics” cutting transportation emissions
“Climate modeling” improving predictions

Industry-friendly estimates suggest AI could shave a few percent off global emissions through efficiency gains.

The Thermodynamic Reality#

Direct environmental footprint of AI infrastructure:

Energy Consumption#

Current: data centers account for on the order of 1.5% of global electricity today; the broader digital/ICT sector is usually put at a few percent—only under the widest boundary does the “digital economy” approach ~10%.[6]
Projection: the IEA projects data-center electricity demand could roughly double by 2026 versus 2022.[6]
Per-query cost: the widely repeated claim that a single generative-AI query uses “100× a Google search” is false. The original de Vries estimate was ~10× (≈3 Wh vs ~0.3 Wh), and even that has since been revised down: Epoch AI estimates a typical GPT-4o query at ~0.3 Wh—roughly parity with a web search.[7] The real concern is aggregate load, not the per-query cost.
Training costs: training a frontier model plausibly runs into the tens of thousands of MWh (GPT-3 was estimated near 1,300 MWh; larger successors are estimated at multiples of that), though the exact figures are not disclosed and should be read as estimates.[8]

The defensible argument is not about any single query. It is that AI infrastructure is growing far faster than any efficiency it delivers—so at the system level the net effect is acceleration, not mitigation.

Water Consumption#

AI data centers and the fabs that supply them require enormous cooling and process water.

Google reported its data-center water withdrawal rose about 20% in a single year (2021→2022, to roughly 5.6 billion gallons), a jump the company attributes in part to AI-driven demand.[9]
Chip fabrication is even thirstier at the source. TSMC—which makes the overwhelming majority of the world’s advanced logic—consumed on the order of 150,000–190,000 tonnes of water per day (2019–2020), comparable to a mid-size city. During Taiwan’s worst drought in half a century in 2021, TSMC had to truck in water to keep production running.[10]

In regions already facing scarcity—the American Southwest, Taiwan—this demand competes directly with agriculture and residential use.

Material Waste and Mining Impact#

The hardware itself carries a large upstream and downstream footprint.

Electronic waste: a 2024 study in Nature Computational Science projects that the generative-AI boom could add 1.2 to 5 million tonnes of e-waste cumulatively between 2023 and 2030, driven by rapid hardware turnover—much of it laden with lead, chromium and other toxic metals.[11]
Critical minerals: the chips embed gallium, germanium, indium, tantalum and rare earths, whose extraction generates large volumes of tailings and whose supply is highly concentrated.
Human cost of the supply chain: the cobalt in batteries and electronics is exposed to some of the worst conditions in mining. UNICEF estimated that in 2014 roughly 40,000 children worked in artisanal mines across southern DRC—the heart of global cobalt production—and Amnesty International documented children as young as seven, earning $1–2 a day.[12]

The “green” transition to AI-optimized systems is not weightless. It requires a mining and hardware-turnover boom with real environmental and human costs.

AI as Fossil-Fuel Accelerant#

The application no one puts in the marketing deck:

Major AI companies maintain active partnerships with oil and gas producers—AI for subsurface geology mapping, machine learning for extraction optimization, predictive maintenance for offshore platforms. One of the most immediately profitable applications of AI is accelerating fossil-fuel extraction: optimizing drilling locations, reducing downtime, and extending the economic life of marginal wells.

The technology marketed as the solution to climate change is, in this corner of the industry, actively deployed to extend the lifespan of the fossil-fuel business.

The Renewable Transition Trap: Scale Is the Enemy#

The Storage Problem#

A fully renewable grid needs massive storage to ride through the times when the sun does not shine and the wind does not blow. The scale of seasonal storage is where the argument gets serious.

By one contested estimate—Simon Michaux’s 2021 report for the Geological Survey of Finland—buffering the global grid for four weeks would require on the order of 2.5 billion tonnes of batteries, vastly more than current annual production could supply for centuries at today’s rates.[13] Michaux’s methodology is disputed by mainstream analysts: it assumes batteries as the sole buffer, with no other storage, no demand flexibility and no grid interconnection. Treat the exact number as a worst-case, not settled fact. But even generous scenarios agree on the direction: seasonal storage at grid scale is a materials problem of a wholly different order than daily storage.

The trap: renewable energy without storage is intermittent. Storage at seasonal scale is materially daunting. So dispatchable baseload—fossil or nuclear—remains hard to displace entirely.

The Infrastructure Expansion Problem#

Phasing out fossil fuels requires overbuilding generation capacity to account for lower capacity factors (solar ~20–25%, wind ~30–35% versus ~80–90% for fossil/nuclear baseload), intermittency, and transmission losses.

Michaux’s same report puts numbers on the overbuild. For the United States alone, he estimates replacing fossil-fuel electricity would require thousands of additional average-sized generating plants; globally, on the order of hundreds of thousands of new plants, with capital expenditure in the tens of trillions of dollars.[13] These are one contested model’s figures, not consensus—but even discounted heavily, they describe a build-out without historical precedent, and one whose material requirements press hard against known reserves of copper, silver and rare earths.

The EROI Squeeze#

When you shift from high-EROI energy toward lower-EROI energy, you must dedicate a larger share of your total energy budget just to acquiring energy—leaving less surplus for everything else. As you pour energy into building renewable infrastructure (mining, manufacturing, installation), the surplus available to maintain societal complexity is squeezed. That squeeze arrives at exactly the moment AI infrastructure is demanding exponentially more.

You can push a high-complexity AI civilization on fossil fuels until the surplus and the climate give out. Or you can build within the surplus a tightening energy system actually delivers. The comfortable assumption—that we can have unlimited compute, a full renewable transition, and undiminished economic complexity all at once—is the assumption the physics quietly refuses.

Jevons Paradox: Why Efficiency Gains Accelerate Consumption#

The Trap of Efficiency#

Jevons Paradox—named for the 19th-century economist William Stanley Jevons—describes how, as technology makes a resource more efficient to use, total consumption of that resource tends to rise rather than fall.

Why:

Efficiency makes the resource cheaper per unit of output
Cheaper resources get used more widely
New applications become economically viable
Total consumption grows despite per-unit efficiency gains

AI as the Perfect Jevons Trap#

Consider AI-optimized logistics. The promise is that route optimization cuts trucking fuel use per shipment. The reality is that cheaper shipping enables more just-in-time manufacturing, which drives more shipping volume—so total fuel consumption can climb even as each route gets more efficient. Every efficiency gain unlocked by AI tends to be reinvested into expanded activity rather than banked as savings.

The Impossible Dream of “Decoupling”#

Green-growth theory holds that GDP can keep rising while resource consumption falls in absolute terms. The empirical record, as Hickel and Kallis document, does not support this at the necessary scale or speed:[2]

No large economy has achieved sustained absolute decoupling of GDP from energy and material use.
Relative decoupling (emissions per unit of GDP falling) is common—but total emissions keep rising because GDP grows faster than efficiency improves.
AI intensifies the dynamic: by making the economy more efficient, it enables faster growth, which overwhelms the per-unit gains.

Decoupling, in this reading, is the story that lets policymakers avoid the harder question of what a lower-throughput economy actually looks like.

The Synthesis: Three Goals in Tension#

Goal 1: Exponential AI expansion (on the order of 12× compute growth in 36 months)

Goal 2: Transition to renewable energy (phasing out fossil fuels)

Goal 3: Maintain current economic complexity (global supply chains, advanced services, universal access to technology)

The thermodynamic reality: pursued at full ambition, these three goals pull against one another. Within a tightening energy surplus, you can push hard on at most two at once—which is the uncomfortable trade-off the following scenarios spell out. These are illustrative scenarios, not forecasts.

Scenario A: AI + Renewables (Squeezing Goal 3)#

Choice: build AI infrastructure on renewable energy.

Consequence: the tighter surplus and large material requirements pressure the rest of the economy toward simplification. Supply chains contract; some universal services become harder to sustain; societal complexity trends toward what the available surplus can carry.

Who bears the cost: the bottom half to three-quarters of the population, who lose access first to services that depend on abundant surplus energy.

Scenario B: AI + Economic Complexity (Squeezing Goal 2)#

Choice: build AI infrastructure on continued fossil-fuel use.

Consequence: climate impacts accelerate and resource depletion continues. The energy system grows more fragile as the useful-stage EROI of remaining fossil fuels declines, eventually forcing a disorderly transition.

Who bears the cost: future generations, and populations in climate-vulnerable regions—which overlap heavily with those that contributed least to emissions.

Scenario C: Renewables + Economic Complexity (Squeezing Goal 1)#

Choice: transition to renewables while defending economic complexity, and constrain AI growth.

Consequence: AI development slows or plateaus near current capability levels. Centralized AI companies that have committed hundreds of billions of dollars to hyperscale clusters face investor pressure. Economic power shifts toward regions that accept lower complexity in exchange for energy sovereignty.

Who bears the cost: AI investors and the companies that bet everything on uninterrupted exponential scaling.

The Distributed Alternative: Working With Physics, Not Against It#

The current trajectory assumes it can violate biophysical constraints. It assumes:

Infinite energy growth from finite sources
Efficiency gains that reduce rather than accelerate consumption
Material abundance despite geological scarcity

The distributed alternative accepts biophysical limits and designs within them:

Regenerative Agriculture#

Works with solar energy flows (photosynthesis) rather than mined inputs (fertilizer)
Builds soil as a carbon sink rather than depleting it
Operates at human scale with low-complexity tools, not AI-optimized industrial monoculture
Measured profitability: a 2025 Tanzania study of conservation agriculture found profit rising from about $177 to $527 per unit of land—a 198% increase—alongside higher yields.[14] (The source studies conservation agriculture and reports per-hectare figures.)

Microgrids#

Match generation to local demand rather than requiring vast transmission infrastructure
Run on lower total throughput by eliminating transmission losses (typically 5–15%)
Use grid-forming inverters that don’t require complex centralized balancing
Resilience: local grid-forming systems can recover in seconds where centralized grids can take hours to restore after a failure

Circular Supply Chains#

Reuse existing materials rather than mining virgin resources
Localize production to cut logistics energy
Design for low complexity—repair and remanufacture over AI-optimized global coordination
Documented savings: the Kalundborg industrial symbiosis in Denmark reports roughly $310M in cumulative savings over 40 years and proved resilient through COVID-19 disruption.[15]

The synthesis: these systems work with a tightening surplus rather than against it. They don’t require exponential growth or assume abundant materials. They accept thermodynamic limits and function within them.

Bridge to Part 2: The Only Way Out Is Through Simplification#

The biophysical trap is not a problem to be solved with better technology. It is a constraint to be accepted and designed around.

Centralized AI infrastructure attempts to defy these constraints by assuming:

An energy surplus more abundant than the numbers support
Material availability that geology does not guarantee
Efficiency gains that history shows tend to accelerate consumption

The distributed, regenerative alternative doesn’t try to overcome physical limits. It works within them.

Part 2 argues that these alternatives are not merely physically viable—they are economically superior precisely because they accept, rather than resist, thermodynamic reality.

When the rising maintenance cost of complexity meets a tightening energy surplus, low-complexity, distributed systems are the ones that endure.

Closing Thought#

The framing that runs through this chapter is not mine alone; it belongs to a lineage of biophysical economists—Hall, Hagens, Hickel, Kallis—who have spent careers insisting on a point the mainstream keeps deferring. The energy crisis is not, at root, a problem to be solved. It is a reality to adapt to.

The sun delivers a fixed budget of energy to the Earth each day. For two centuries we have supplemented that budget by drawing down a savings account of ancient sunlight stored underground—and the interest rate on that account, its EROI, is falling. Building ever-larger centralized AI on top of that shrinking margin is a spending spree undertaken as the balance runs thin.

The path that survives is the one physics has been describing all along: simplify, localize, distribute, and match complexity to the surplus energy actually available. The choice is not between growth and stagnation. It is between deliberate simplification and the involuntary kind.

Sources#

[1] Charles A.S. Hall and Kent A. Klitgaard, Energy and the Wealth of Nations: An Introduction to Biophysical Economics (Springer, 2nd ed., 2018).

[2] Jason Hickel and Giorgos Kallis, “Is Green Growth Possible?” New Political Economy 25, no. 4 (2020): 469–486. https://doi.org/10.1080/13563467.2019.1598964

[3] Charles A.S. Hall, Jessica G. Lambert, and Stephen B. Balogh, “EROI of different fuels and the implications for society,” Energy Policy 64 (2014): 141–152. https://doi.org/10.1016/j.enpol.2013.05.049

[4] Khagendra P. Bhandari et al., “Energy payback time (EPBT) and energy return on energy invested (EROI) of solar photovoltaic systems: A systematic review and meta-analysis,” Renewable and Sustainable Energy Reviews 47 (2015): 133–141. https://doi.org/10.1016/j.rser.2015.02.057

[5] Jaime Aramendia, Paul E. Brockway et al., “Estimation of useful-stage energy returns on investment for fossil fuels and implications for renewable energy systems,” Nature Energy 9 (2024): 803–816. https://www.nature.com/articles/s41560-024-01518-6

[6] International Energy Agency, Electricity 2024 — data-center electricity demand projected to roughly double by 2026. https://www.iea.org/reports/electricity-2024

[7] Epoch AI, “How much energy does ChatGPT use?” (2025) — ~0.3 Wh per typical query, roughly parity with a web search. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use

[8] Alex de Vries, “The growing energy footprint of artificial intelligence,” Joule 7, no. 10 (2023): 2191–2194. https://doi.org/10.1016/j.joule.2023.09.004

[9] Google, Environmental Report 2023 — ~20% year-over-year rise in data-center water withdrawal (to ~5.6 billion gallons). https://sustainability.google/reports/google-2023-environmental-report/

[10] Eamon Barrett, “The global chip shortage is bad. Taiwan’s drought threatens to make it worse,” Fortune (June 12, 2021); TSMC daily water use ~150,000–190,000 tonnes and trucked-water response to the 2021 drought. https://fortune.com/2021/06/12/chip-shortage-taiwan-drought-tsmc-water-usage/

[11] Peng Wang et al., “E-waste challenges of generative artificial intelligence,” Nature Computational Science (2024) — projected 1.2–5 million tonnes cumulative e-waste, 2023–2030. https://www.nature.com/articles/s43588-024-00712-6 · MIT Technology Review coverage: https://www.technologyreview.com/2024/10/28/1106316/ai-e-waste/

[12] Amnesty International, “This Is What We Die For: Human rights abuses in the Democratic Republic of the Congo power the global trade in cobalt” (2016) — UNICEF 2014 estimate of ~40,000 children in southern-DRC artisanal mines; children as young as seven; $1–2/day. https://www.amnesty.org/en/documents/afr62/3183/2016/en/

[13] Simon P. Michaux, “Assessment of the Extra Capacity Required of Alternative Energy Electrical Power Systems to Completely Replace Fossil Fuels,” Geological Survey of Finland (GTK), Report 42/2021 — ~2.5 billion-tonne battery buffer and generation build-out estimates; methodology contested. https://tupa.gtk.fi/raportti/arkisto/42_2021.pdf

[14] Frontiers in Sustainable Food Systems (2025) — Tanzania conservation-agriculture profitability: ~$177 → ~$527 profit, +198%. https://www.frontiersin.org/journals/sustainable-food-systems/articles/10.3389/fsufs.2025.1706205/full

[15] Kalundborg Symbiosis — ~$310M cumulative savings over 40 years. https://www.symbiosis.dk/en/