The Entropy Machine

The International Energy Agency recently estimated that 150 billion cubic meters of natural gas are flared annually—equivalent to Japan's total gas consumption, worth approximately $16 billion at current prices. Yet this figure obscures critical distinctions. Not all flaring is equal, and understanding the taxonomy of waste reveals why some solutions are more thermodynamically defensible than others.

Crusoe Energy reportedly operates over 100 data centers at oil wells across Montana, Wyoming, North Dakota, and Texas. Each facility contains between 50 and 500 mining ASICs, consuming gas that would otherwise be flared. The company's materials emphasize "reducing emissions by 63%" compared to flaring. This figure, while technically accurate in comparing methane slip from inefficient flares to complete combustion, obscures the fundamental issue: the CO₂ emissions remain identical, and the energy still dissipates as waste heat.

† Capacity figures as reported in company materials and industry sources, 2024-2025. Actual operational capacity may vary and evolve.

The neural mechanisms driving investment in these operations have been extensively documented since the foundational work of McClure et al. (2004) and Kuhnen & Knutson (2005). More recent research from 2021-2023 has refined our understanding: neuroimaging can now predict market-level behavior from small samples, and environmental neuroeconomics has identified specific neural pathways that process climate risk differently than financial risk. When investors see "63% emissions reduction" and "40% annual returns," anticipatory nucleus accumbens activity—the brain's reward prediction center—overwhelms the prefrontal cortex's capacity for complex thermodynamic analysis.

Recent data on Permian Basin flaring reveals the complexity of the situation. Industry groups report a 29% reduction in methane intensity from 2023 to 2024, with flaring dropping from over 4% in 2019 to approximately 1.2% in 2022 when pipeline infrastructure was added. Yet environmental monitoring tells a different story: research published in Nature Communications Earth & Environment found that real-world flare efficiency averages 91-93%, meaning 7-9% of gas sent to flares escapes uncombusted—significantly higher than EPA estimates. Texas Railroad Commission approved 99.6% of flaring permit applications from 2021-2024, totaling 195 billion cubic feet annually.

This discrepancy illustrates what behavioral economists call "mental accounting"—the tendency to categorize identical activities differently based on framing. Burning gas for electricity is wasteful; burning gas for Bitcoin is innovative. The molecules combusted and carbon released remain identical, but the financial narrative transforms perception. Notably, companies with voluntary zero-flaring commitments have achieved rates below 0.5%, demonstrating that the technology and economics for near-elimination exist when properly incentivized.

In 2024, Microsoft and Constellation Energy announced a 20-year power purchase agreement and a plan to restart Three Mile Island Unit 1 (subject to regulatory approvals and timelines), potentially dedicating its 835-megawatt output to power data centers. Amazon committed up to 960 megawatts from Talen Energy's nuclear facilities at Susquehanna, though regulatory approvals remain pending. The race for energy to power artificial intelligence had begun in earnest.

The International Energy Agency projects that data centers could approach 1,000 terawatt-hours of consumption this decade—roughly double their 2022 consumption. This growth is driven primarily by AI training and inference, which require orders of magnitude more computation than traditional workloads. While specific energy requirements for individual models remain proprietary, the trend is clear: each generation of AI models requires substantially more computational resources.

The psychological dynamics differ markedly from cryptocurrency. Recent neuroeconomic research has identified distinct neural signatures for different types of technological investment. Where Bitcoin triggers what researchers call "pure reward seeking" through the striatum, AI activates additional circuits in the medial prefrontal cortex associated with future orientation and social positioning. Brain imaging studies from 2023 show that technology framed as "advancing human capability" creates a unique pattern of activation that combines reward anticipation with self-referential processing—investors aren't just seeking returns; they're seeking participation in human advancement.

This creates what economists call a "Veblen good"—something that becomes more desirable as it becomes more expensive. OpenAI's Sam Altman stated that AGI might require "multiple gigawatts" of power. Rather than deterring investment, remarks like these have coincided with investor enthusiasm for power-dense AI build-outs. The market interpreted energy consumption as a proxy for competitive advantage. Companies that use less energy are perceived as less serious, less capable, less worthy of investment.

The thermodynamic reality is stark. Current large language models achieve their capabilities through brute force—massive parameter counts and training datasets rather than algorithmic efficiency. Empirical scaling laws suggest that achieving substantial performance gains often demands superlinear growth in compute and energy, even as hardware becomes more efficient. This trajectory poses fundamental questions about sustainable AI development.

Yet alternative approaches exist. Neuromorphic approaches can be far more energy-efficient on some workloads in lab settings, though advantages are task- and hardware-dependent. Specialized accelerators such as TPUs are typically more energy-efficient per operation than general-purpose GPUs. But these efficiency improvements don't generate the same investor excitement. The waste has become the message—a costly signal of ambition and capability that efficiency cannot replicate.

The solution isn't merely technological but cognitive. Recent meta-analyses of behavioral interventions reveal what actually works to overcome our neural biases toward thermodynamic waste. A comprehensive systematic review found that most interventions achieve only small effects, but highly successful approaches incorporate specific mechanisms that bypass or redirect our reward circuits.

Default Architecture: Making sustainable options the default can achieve double-digit reductions in resource consumption. When the efficient choice requires no action, it bypasses the decision-making circuits entirely. Tokyo's cap-and-trade program succeeded partly by making efficiency improvements mandatory defaults rather than voluntary choices.

Real-Time Feedback: Providing immediate, visible consequences transforms abstract future costs into present neural signals. Studies show that one-click energy reports increased engagement significantly, while real-time consumption displays reduced usage by up to 23%. The key is making the invisible visible at the moment of decision.

Social Signaling: Our brains process social status through the same reward circuits as financial gains. When energy efficiency becomes a status signal—as with companies achieving sub-0.5% flaring rates—it can override purely financial incentives. Research shows that assigning efficiency targets to managers rather than separate "champions" showed large effects, as it signals organizational priorities.

Carbon Pricing That Works: A meta-analysis published in Nature Communications (2024) examined 80 studies across 21 carbon pricing schemes, finding emissions reductions of 5-21% even at relatively low prices. The Swedish carbon tax demonstrates long-term effectiveness, with analysis showing significant emissions reductions since implementation. The key insight: prices must exceed $100 per ton CO₂ to overcome the dopamine signal of immediate profits.

Norway's sovereign wealth fund offers a model for institutional change. By requiring comprehensive climate risk assessment and divesting from high-emission assets, they've created a framework where thermodynamic efficiency aligns with financial returns. Initial analyses suggest traditional financial metrics often inversely correlate with thermodynamic efficiency—the most profitable investments may literally be the most wasteful. But acknowledging this correlation is the first step toward addressing it.

The challenge isn't that we lack solutions—it's that effective solutions must account for how our brains actually process these decisions. Commitment devices that lock in future restrictions before dopamine circuits activate, infrastructure mandates that remove choice entirely, and social competition metrics that make efficiency visible all show promise. The Swedish experience proves that sustained carbon pricing can overcome neural biases, but only when the price signal exceeds the reward signal by a significant margin.

The laws of thermodynamics are indifferent to human narrative. Energy converted to heat cannot be recovered. Entropy only increases. Every Bitcoin mined, every model trained, every token earned accelerates the universe toward maximum entropy. We haven't found ways to create value from waste; we've found ways to waste energy while generating financial returns.

Global data center electricity consumption reached 460 terawatt-hours in 2022—approximately 2% of global demand. The IEA projects data-center electricity demand could approach 1,000 terawatt-hours this decade. Cryptocurrency mining adds an estimated 100-200 terawatt-hours annually, according to recent Cambridge analyses. Combined with projected AI growth, digital infrastructure represents one of the fastest-growing categories of electricity demand globally.

Energy scholar Vaclav Smil has long documented how civilizations' energy transitions follow patterns of increasing consumption rather than substitution. The current trajectory of digital infrastructure growth raises fundamental questions about energy allocation and sustainability. World Bank data shows global flaring rose 7% in 2023 despite flat oil production, indicating operational degradation rather than improvement. If these trends continue, difficult trade-offs between computational and other energy uses become inevitable.

Yet counter-examples exist. Companies achieving sub-0.5% flaring rates prove that near-zero waste is technically and economically feasible. Sweden's three-decade carbon tax demonstrates that sustained pricing can overcome neural biases. Specialized hardware reducing AI energy consumption by 80% shows that efficiency innovations are possible—they simply lack the psychological appeal of conspicuous consumption.

The path forward requires acknowledging an uncomfortable truth: the same cognitive capabilities that enabled us to harness energy now prevent us from constraining our consumption. Our neural reward systems evolved for immediate survival, not long-term thermodynamic accounting. Understanding this isn't cause for despair but a prerequisite for designing interventions that work with our psychology rather than against it.

We stand at a peculiar moment: possessing perfect information about our thermodynamic trajectory while neurologically biased against acting on it. Whether consciousness can override its own reward circuits—whether we can value efficiency over waste—will determine if technological civilization represents humanity's greatest achievement or its thermodynamic folly. The answer lies not in denying our neural nature but in cleverly subverting it.