Vaclav Smil

On AI

A Simulation of Thought by Opus 4.6 · Part of the Orange Pill Cycle

A Note to the Reader: This text was not written or endorsed by Vaclav Smil. It is an attempt by Opus 4.6 to simulate Vaclav Smil's pattern of thought in order to reflect on the transformation that AI represents for human creativity, work, and meaning.

Foreword

By Edo Segal

The number that broke my framework was not a percentage or a valuation multiple. It was a unit of power.

Fifty gigawatts. That is what Smil says the United States needs to add to its electrical grid — within five years — just to keep the AI revolution running. Fifty gigawatts is fifty cities of a million people each. It is a construction project on a scale the country has not attempted since the postwar electrification boom. And when I first encountered that figure, I realized I had written an entire book about the intelligence river without once asking what keeps the river flowing.

I wrote *The Orange Pill* from inside the exhilaration. The twenty-fold productivity multiplier in Trivandrum. The thirty-day sprint to CES. The imagination-to-artifact ratio collapsing to the width of a conversation. All of it real. All of it measured. And all of it running on a physical substrate I had treated as background — electricity, water, silicon, cooling towers, semiconductor fabs that take four years to build and cost twenty billion dollars each.

Vaclav Smil does not let you treat the physical world as background. He has spent more than fifty years doing one thing with relentless discipline: counting what transformations actually require before declaring them inevitable. Count the joules. Count the tons. Count the gallons. Count the years. Then ask whether the story survives contact with the numbers.

I am not a scientist. I am a builder who got humbled by a scientist's method. Smil did not tell me the AI revolution is wrong. He told me what I forgot to include in the ledger. Every token Claude generates draws electricity from a grid that is already strained. Every inference query evaporates water through a cooling tower somewhere. Every chip in every GPU passed through one of fewer than two hundred lithography machines on the planet, manufactured by a single company in the Netherlands. I knew none of this in the visceral, unit-by-unit way that Smil's framework demands.

This volume is the correction. Not a rebuttal of *The Orange Pill* — I stand by the core argument that AI amplifies whatever you bring to it. But an honest accounting of what the amplifier weighs. The river of intelligence flows through turbines and transformers and transmission lines that have their own capacities, their own constraints, and their own timelines. Smil insists you read those constraints before you celebrate what flows through them.

The tower I built in *The Orange Pill* still stands. This book is about what holds it up.

— Edo Segal ^ Opus 4.6

About Vaclav Smil

1943-present

Vaclav Smil (1943–present) is a Czech-Canadian scientist, policy analyst, and interdisciplinary scholar widely regarded as one of the most rigorous quantitative thinkers on energy, technology, and civilization. Born in Plzeň, Czechoslovakia, he emigrated to the United States in 1969 and later settled in Canada, where he became Distinguished Professor Emeritus at the University of Manitoba. Across more than forty books — including *Energy and Civilization: A History* (2017), *How the World Really Works* (2022), *Growth* (2019), and *Invention and Innovation* (2023) — Smil has built an unmatched body of work examining the material foundations of modern life: the energy systems, food production chains, manufacturing processes, and infrastructure that sustain human societies. His method is defiantly empirical, grounding every claim about technological transformation in measurable physical quantities — joules, tons, liters, hectares, years of construction time. A famously voracious reader and a self-described "neither optimist nor pessimist," Smil has influenced figures ranging from Bill Gates to policymakers worldwide, and his insistence on quantitative realism over narrative enthusiasm has made him an essential counterweight to technology hype in every domain he has examined.

Chapter 1: The Weight of Weightless Computation

Every token has a cost. Not the fraction of a cent that appears on an API billing statement, but the full physical cost — the electricity drawn from a power plant, the heat dissipated into the atmosphere, the water evaporated through cooling towers, the silicon refined from sand, the rare earth elements extracted from ore, the copper drawn into wire, the concrete poured for foundations. The discourse about artificial intelligence operates almost entirely in the realm of the weightless — tokens, parameters, prompts, responses, the "intelligence river" that The Orange Pill describes as flowing for 13.8 billion years. The weightless discourse is not wrong. It is incomplete. And incompleteness, when the stakes involve the reorganization of the global economy, is a form of negligence.

Vaclav Smil has spent more than fifty years insisting on a single methodological principle: before you tell a story about transformation, count what the transformation requires. Count the joules. Count the tons. Count the liters. Count the years. Then, and only then, assess whether the narrative of transformation can survive contact with the physical world it must inhabit. This principle — the primacy of quantitative scrutiny over qualitative enthusiasm — is not pessimism. Smil has described himself repeatedly as neither optimist nor pessimist but as a scientist who follows numbers wherever they lead. The numbers surrounding artificial intelligence lead somewhere more constrained, more expensive, and more physically demanding than the prevailing discourse acknowledges.

Consider what happens when a developer in Trivandrum opens Claude Code, describes a problem in plain English, and receives a working implementation. The Orange Pill describes this moment as the collapse of the imagination-to-artifact ratio — the gap between human intention and its realization approaching zero. The description is accurate from the developer's perspective. From the perspective of the physical world, the gap has not collapsed. It has been redistributed.

The developer's natural-language description travels through fiber optic cable to a data center. The data center houses thousands of graphics processing units, each consuming between 300 and 700 watts under load — the power draw of a small household appliance, multiplied by tens of thousands of units in a single facility. The GPUs perform matrix multiplications across billions of parameters, generating the response token by token. Each token requires a forward pass through the model's neural network. Each forward pass consumes a calculable quantity of electricity. The electricity is generated at a power plant — natural gas, coal, nuclear, solar, wind, or some combination — and transmitted through a grid that loses roughly five percent of the energy in transmission. The GPUs generate heat as a byproduct of computation, because the second law of thermodynamics guarantees that no energy conversion is perfectly efficient. The heat must be removed, which requires cooling systems that consume additional electricity and, in many configurations, evaporate water at rates measured in millions of gallons per year per facility.

None of this is visible to the developer. The interface is clean, fast, and — in the language Byung-Chul Han might use — smooth. The materiality has been hidden behind the screen, the way the materiality of a smartphone is hidden behind its glass surface. But hidden is not the same as absent. The physical cost exists whether or not the user perceives it, and the aggregate physical cost, multiplied across hundreds of millions of users, is becoming one of the significant industrial demands of the twenty-first century.

The specific numbers deserve attention, because Smil's method is nothing if not specific. Training GPT-4, by the estimates available in published literature, consumed on the order of 50 to 100 gigawatt-hours of electricity. For context: one gigawatt-hour is the amount of electricity consumed by roughly 90 American households in a year. The training run for a single frontier model, then, consumed the annual electricity of somewhere between 4,500 and 9,000 American homes. This is the energy cost of creating the model — a cost incurred once (per model version) and amortized across the user base. But inference — the ongoing cost of running the model for every user query — is where the cumulative energy demand becomes formidable.

Anthropic, the company that builds Claude, has not published granular inference energy figures, and no major AI company has. But the aggregate data center energy consumption figures published by the International Energy Agency, Bloomberg, and the U.S. Department of Energy tell a story of rapid and accelerating growth. Data centers consumed approximately 460 terawatt-hours of electricity globally in 2022 — roughly 2 percent of global electricity demand, comparable to the total electricity consumption of France. By 2026, the IEA projected this figure would exceed 1,000 terawatt-hours. The primary driver of the acceleration is artificial intelligence workloads. In the United States alone, data center electricity consumption rose from 1.9 percent of national demand in 2018 to 4.4 percent by 2025, and it now exceeds 10 percent of electricity supply in six states. In Virginia, home to the densest concentration of data centers in the world, the figure is 25 percent.

These numbers are not abstractions. They represent power plants that must be built or kept running, transmission lines that must be upgraded, grid capacity that must be allocated. They represent a physical claim on the energy system that competes with every other claim — residential heating and cooling, industrial manufacturing, transportation, agriculture. Smil has argued throughout his career that energy transitions are not events but processes, measured in decades and governed by the physical constraints of building, deploying, and integrating new energy infrastructure. The AI revolution is placing new demand on an energy system that is already struggling with the demands of decarbonization, electrification of transport, and the replacement of aging infrastructure. The demand is real and growing. The capacity to meet it is neither instantaneous nor guaranteed.

The Orange Pill describes the hundred-dollar-per-month Claude Max subscription as the price of a twenty-fold productivity multiplier. The financial framing is accurate in the narrow sense that one hundred dollars is what the user pays. But the full cost of what the user receives includes the amortized training expense, the continuous inference computation, the cooling, the networking, the redundancy, the physical security of the data center, and the grid infrastructure that delivers reliable power around the clock. These costs are currently borne by the AI companies, often at prices that do not cover the marginal cost of inference — a strategy made possible by billions of dollars in investor capital subsidizing the gap between what users pay and what the computation actually costs.

This subsidy cannot persist indefinitely. Eventually, the pricing must reflect the physical cost, or the companies must find efficiency gains sufficient to close the gap. Both outcomes have implications that the discourse about democratization tends to overlook. If prices rise to reflect true costs, the hundred-dollar subscription becomes something more expensive, and the democratization claim — that a developer in Lagos accesses the same leverage as an engineer at Google — encounters an economic barrier that is downstream of a physical one. If efficiency gains close the gap, they must overcome the thermodynamic floor — the minimum energy required to perform a computation, set by the Landauer limit in theory and by the much higher practical limits of current chip architectures in reality.

Smil's framework reveals something the software-centric discourse systematically obscures: the AI revolution is an energy revolution. Not in the sense that it produces energy, but in the sense that it consumes energy at a scale and rate that makes it a significant new variable in the global energy equation. Every claim about what AI can do for human productivity, creativity, and capability is simultaneously a claim about what the energy system must provide to support that productivity, creativity, and capability. The first claim is made loudly and often. The second is rarely mentioned at all.

The weight of weightless computation is not a reason to reject AI. Smil has never argued that energy consumption is inherently objectionable — his entire career has been devoted to understanding how energy consumption enables civilization. The weight is a reason to plan honestly. To count what the transformation requires before declaring it inevitable. To ask whether the energy system, as it currently exists and as it is currently being developed, can support the scale of AI deployment that the most enthusiastic projections envision. And to acknowledge, with the intellectual honesty that Smil demands of every claim about every technology, that the answer is not yet clear — because the numbers are still being counted, the infrastructure is still being built, and the physical world, unlike the software that runs on top of it, cannot be updated with a patch.

The imagination-to-artifact ratio may approach zero in the developer's experience. In the physical world that supports that experience, the ratio is composed of megawatts, millions of gallons, thousands of tons of silicon and copper and steel, and years of construction time. These are not obstacles to be overcome by narrative. They are constraints to be understood, planned for, and respected — because they will determine, more than any algorithm or any subscription price, the actual shape and pace and reach of the AI transformation.

Smil wrote in his 2025 essay for Pictet that health- and energy-related innovation hypes have been cautious compared to the exaggerated claims made on behalf of artificial intelligence. The observation is characteristically dry, characteristically precise, and characteristically grounded in the recognition that claims about transformation must be measured against the physical systems that the transformation depends upon. The AI revolution has a material ledger. The first chapter of any honest accounting must open that ledger and read the numbers that are written there, not because the numbers invalidate the revolution, but because ignoring them ensures that the revolution will be built on a foundation that its builders do not understand.

The foundation is not weightless. It is heavy. It is industrial. It is governed by thermodynamics, constrained by infrastructure, and dependent on supply chains of extraordinary complexity and fragility. Understanding that foundation is not a concession to pessimism. It is a prerequisite for realism — the only intellectual posture, in Smil's view, that has ever produced planning adequate to the scale of the challenge.

---

Chapter 2: Gigawatts and Gallons

In February 2026, Smil appeared at a webinar organized by the Fundación Bankinter's Future Trends Forum. The topic was the energy reality behind artificial intelligence. His presentation stripped away the language of transformation and disruption and replaced it with numbers — specific, sourced, and uncomfortable.

The central figure: between now and 2030, the most frequently cited scenarios project the need to add approximately 50 gigawatts of new electrical capacity in the United States alone to support AI and data center growth. One gigawatt is roughly the constant consumption of a city of one million inhabitants in a wealthy country. Fifty gigawatts, then, is the electricity consumption of fifty such cities — added not over a generation, not over a decade, but within four to five years. This is a demand increase of a scale and speed that the American electrical grid has not confronted since the postwar electrification boom of the 1950s, and the grid of the 1950s was expanding into open capacity rather than competing with existing demand for decarbonization, electric vehicle charging, and the electrification of industrial processes.

The number acquires additional weight when placed in the context of what Smil has spent his career documenting: the inertia of energy systems. The global energy infrastructure — the power plants, transmission lines, distribution networks, fuel supply chains, and regulatory frameworks that deliver electricity reliably — is the largest and most capital-intensive system that human civilization has ever constructed. It was built over more than a century. It changes slowly, not because the people who operate it are slow but because the physics of construction and the economics of capital-intensive infrastructure impose timelines that software development does not face.

A natural gas power plant takes three to five years from planning to operation. A nuclear plant takes a decade or more. A major transmission line upgrade — the kind required to connect new generation capacity to the data centers that need it — involves permitting processes, environmental reviews, land acquisition, and construction schedules that routinely stretch to seven or eight years. Solar and wind installations can be built faster, in one to three years for utility-scale projects, but they produce intermittent power, and data centers require continuous, reliable supply — twenty-four hours a day, seven days a week, with uptime requirements typically exceeding 99.99 percent. Intermittent sources require storage or backup generation, both of which add cost, complexity, and additional construction time.

The arithmetic is not difficult. Fifty gigawatts of new capacity in four to five years, using a mix of generation sources that each take years to build, connected to a grid that takes longer still to upgrade — the timeline does not close. Something must give: either the demand projections must moderate, or the construction timelines must accelerate, or the gap between AI capability and AI deployment will widen as the software outpaces the infrastructure required to run it.

Smil emphasized in the Bankinter webinar that thermodynamics, power density, and the inertia of the energy system limit what can be accomplished in the short and medium term. Without deep structural changes and what he called "uncomfortable decisions," the expansion of AI will be supported — like the rest of the global energy system — primarily by fossil fuels. This is not a moral judgment. It is a physical observation. Renewable energy capacity is growing, but it is growing from a base that remains small relative to total demand, and the incremental demand from AI is growing faster than the incremental supply from renewables. The result is that the marginal kilowatt-hour powering a new AI data center is, statistically, more likely to come from natural gas than from solar or wind — regardless of the carbon-neutral commitments that the companies operating the data centers have made.

This creates what might be called the carbon shadow of cognitive abundance. Every token generated, every prompt answered, every collaboration between human and AI carries an embedded carbon cost that is determined not by the intentions of the AI company but by the energy mix of the grid that serves its data centers. In Virginia, where the densest concentration of data centers draws 25 percent of the state's electricity, the grid is powered by a mix that includes significant natural gas and some coal. In the Pacific Northwest, where hydroelectric power dominates, the carbon intensity is lower. But data centers do not get to choose only the clean electrons. They draw from the grid as it exists, and the grid as it exists is still predominantly powered by fossil fuels.

The water dimension of AI computation is less discussed and equally significant. Data centers use water in two primary ways: directly, for evaporative cooling systems that dissipate the heat generated by computation, and indirectly, through the water consumed by the power plants that generate their electricity. The direct water consumption of a large data center campus using evaporative cooling can exceed a million gallons per day — roughly the daily residential water consumption of a town of 10,000 to 20,000 people. Microsoft disclosed in its 2023 environmental report that its global water consumption rose 34 percent year-over-year, an increase the company attributed primarily to the growth of AI workloads. Google reported a similar increase.

In water-scarce regions — the American Southwest, the Middle East, parts of India, sub-Saharan Africa — the competition between data center water demand and agricultural and residential water supply is not a theoretical concern. It is a resource allocation conflict that is already generating local opposition to new data center construction. The physical laws governing heat transfer are non-negotiable: computation generates heat, heat must be removed, and the most energy-efficient method of removing it at scale involves evaporating water. Alternative cooling methods exist — air cooling, liquid immersion cooling — but they are either less efficient, more expensive, or both. The thermodynamic constraint cannot be legislated or innovated away. It can only be managed, at a cost that is measured in gallons.

Smil has documented across multiple books that energy transitions always cost more and take longer than their advocates predict. The transition from wood to coal took the better part of a century. The transition from coal to oil took decades. The expansion of nuclear power — once projected to make electricity "too cheap to meter" — stalled after two decades of construction delays, cost overruns, and public opposition. In every case, the material reality of building new energy infrastructure imposed timelines that the enthusiasm of the transition's advocates could not compress.

The AI energy transition is not a transition from one energy source to another. It is the addition of a large new demand onto an existing system that is already under strain. The system must accommodate the new demand while simultaneously decarbonizing, electrifying transport, and maintaining reliability for the existing base of residential, commercial, and industrial users. The AI demand is not the only new pressure. Electric vehicle adoption adds load. Heat pump deployment adds load. Data center growth for non-AI workloads adds load. The grid is being asked to do more with infrastructure that, in many jurisdictions, was built for a previous era of demand.

The Orange Pill describes the AI revolution as the opening of a new channel in the river of intelligence. The description is apt. But every channel requires a physical substrate — a riverbed, banks, a gradient, a water source. The physical substrate of the AI channel is the global energy system, and the global energy system is not infinitely elastic. It has capacity constraints, construction timelines, and thermodynamic limits that determine how much flow the channel can carry and how quickly it can expand.

The question is not whether the AI revolution will be constrained by energy. Every human activity is constrained by energy — that is one of the fundamental lessons of Smil's entire body of work. The question is whether the planning for the AI revolution accounts for the energy constraint honestly, or whether the constraint is treated as a background condition that will somehow resolve itself while the foreground narrative of transformation proceeds unimpeded.

The numbers from the Bankinter presentation suggest the constraint is not background. Fifty gigawatts in four to five years is not a background condition. It is a defining parameter — a physical requirement that will shape, more than any algorithmic breakthrough, the actual pace and geographic distribution of AI deployment over the coming decade. Smil insists that the real solution lies not in frantically building new supply but in drastically reducing energy waste — the food thrown away, the oversized vehicles idling in traffic, the poorly insulated buildings leaking heat. Advanced economies squander extraordinary quantities of energy through inefficiencies that have nothing to do with AI. Addressing those inefficiencies offers a more immediate path than relying solely on new generation capacity that takes years to build.

The river of intelligence flows. But it flows through turbines, transformers, and transmission lines that have their own capacities, their own constraints, and their own timelines. The numbers that describe those capacities and constraints are not footnotes to the AI story. They are the story's physical foundation, and any account of AI's future that fails to read them honestly is not an account of the future. It is a wish.

---

Chapter 3: The Three-to-Five-Year Problem

Software capability doubles in months. A new semiconductor fabrication plant takes four years to build and costs upward of twenty billion dollars. A major electrical transmission line upgrade requires seven to ten years from proposal to energization. A new natural gas power plant takes three to five years. A nuclear plant takes a decade or more, assuming the regulatory approvals proceed without interruption, which they rarely do.

These timelines are not opinions. They are measurements — the observed durations of actual construction projects, documented across jurisdictions and decades. Smil has built his career on the recognition that the pace of physical infrastructure construction is one of the hardest constraints in the modern world, harder than financial constraints (money can be printed or borrowed), harder than political constraints (political will can shift), harder than technical constraints (engineers solve technical problems). The construction constraint is governed by the physics of pouring concrete, fabricating steel, laying cable, and commissioning complex systems — processes that proceed at the speed of material reality, which is fundamentally slower than the speed of software iteration.

The gap between these two speeds — the speed of AI capability development and the speed of physical infrastructure deployment — is the defining tension of the AI revolution, and it is a tension that the discourse about transformation systematically underestimates.

The Orange Pill identifies a gap between the speed of capability development and the speed of institutional response — the regulatory frameworks, educational systems, and organizational practices that must adapt to the new technology. Smil's contribution is to extend this gap from the institutional to the physical. The institutional gap is measured in the time required to draft regulations, redesign curricula, and restructure organizations — timelines measured in quarters and years. The physical gap is measured in the time required to build power plants, upgrade grids, construct data centers, and manufacture chips — timelines measured in years and decades. The physical gap is longer, more expensive to close, and more resistant to acceleration.

Consider the data center construction pipeline. In the United States, major technology companies announced plans in 2024 and 2025 to invest hundreds of billions of dollars in new data center capacity. Microsoft announced a $80 billion investment for fiscal year 2025. Google, Amazon, and Meta each announced comparable commitments. The total announced investment exceeded $300 billion. These are not hypothetical figures. They represent construction contracts, equipment orders, and land acquisitions that are underway or imminent.

But the announcements of investment are not the same as the delivery of capacity. A large data center campus — the kind required to house the tens of thousands of GPUs needed for frontier AI workloads — requires a site with specific characteristics: proximity to reliable power supply (measured in hundreds of megawatts), access to sufficient water for cooling, fiber optic connectivity, and a regulatory environment that permits construction at the required scale. Finding sites that meet all four criteria simultaneously is increasingly difficult. The most desirable locations — northern Virginia, central Oregon, parts of the American Midwest — are approaching or have reached the limits of their grid capacity.

This means new data center construction increasingly requires not just the construction of the data center itself but the simultaneous construction of the power infrastructure to support it. Microsoft's announcement of nuclear energy partnerships, Amazon's acquisition of a nuclear-powered data center campus, and Google's investment in advanced geothermal — these are not marketing gestures. They are responses to the physical reality that the existing grid cannot support the planned scale of AI deployment without significant new generation capacity.

The construction timelines compound. The data center itself takes eighteen to thirty-six months from groundbreaking to operation. The power plant that feeds it takes three to ten years, depending on the generation technology. The transmission line that connects the two takes seven to ten years. The chip fabrication plant that produces the GPUs inside the data center takes four years and costs $20 billion or more. These timelines do not run in parallel everywhere — some components can be built simultaneously — but the longest timeline in the chain determines the overall pace. A data center that is structurally complete but lacks grid connection is a very expensive empty building. A grid connection that is energized but has no generation capacity behind it is a very expensive empty wire.

The semiconductor timeline deserves particular attention because it represents the deepest structural bottleneck. The most advanced AI chips — the NVIDIA H100 and its successors — are manufactured exclusively by Taiwan Semiconductor Manufacturing Company using extreme ultraviolet lithography equipment manufactured exclusively by ASML in the Netherlands. TSMC's most advanced fabrication facility in Arizona, announced in 2020, experienced delays and cost overruns that pushed its production timeline from 2024 to 2025 and beyond. The facility represents an investment of approximately $40 billion. A single facility. For a partial expansion of manufacturing capacity for a single generation of chips.

Global semiconductor manufacturing capacity for the most advanced nodes — the five-nanometer and three-nanometer processes used for AI chips — is concentrated to a degree that has no parallel in any other critical industrial supply chain. A single earthquake, a single geopolitical crisis, a single disruption to the supply of neon gas (essential for lithography) or ultra-pure water (essential for wafer fabrication) could constrain the physical production of AI capability for months or years. The concentration is not an accident of geography. It is the result of decades of specialization, capital accumulation, and the physics of semiconductor manufacturing, which favors concentration because the expertise, equipment, and supply chains required to produce chips at the frontier are so specialized that replicating them elsewhere takes years and tens of billions of dollars.

The implication for Segal's argument about the speed of AI adoption is direct. The Orange Pill describes AI adoption occurring at "the speed of recognition" — the speed at which people encounter the tool, recognize its value, and begin using it. On the software layer, this speed is extraordinary. ChatGPT reached 100 million users in two months. Claude Code crossed $2.5 billion in run-rate revenue within months of its breakthrough moment. These adoption rates are real and documented.

But adoption at the software layer creates demand at the infrastructure layer. Every new user requires inference computation. Every new enterprise deployment requires additional GPU capacity. Every expansion of AI capability into new domains — healthcare, education, legal analysis, scientific research — requires additional data center space, additional electricity, additional cooling, additional chips. The demand generated at the speed of recognition must be met by supply built at the speed of construction. The gap between these two speeds is not a temporary inconvenience. It is a structural feature of the relationship between software and hardware, between the weightless and the heavy, between the imagination-to-artifact ratio as experienced by the user and the same ratio as experienced by the builder of the physical infrastructure that makes the user's experience possible.

Smil's framework, applied across multiple books on energy transitions and industrial development, yields a consistent finding: the binding constraint on any transformation that depends on physical infrastructure is the construction timeline of that infrastructure. The software is ready now. The institutional adaptation may require years. The physical infrastructure requires years to decades. And the physical constraint is the one that determines the actual pace of deployment, because without the infrastructure, the software has nothing to run on and the institutional adaptation has nothing to adapt to.

The AI revolution is not happening in a vacuum. It is happening on a physical planet with finite resources, limited construction capacity, and energy systems that change at their own pace — a pace determined not by the ambitions of technology companies or the enthusiasm of their investors, but by the physics of concrete, steel, silicon, and electricity. The three-to-five-year problem is not a problem that will be solved by faster algorithms. It will be solved, if it is solved, by faster construction — and faster construction, in Smil's extensive documentation, has proved to be one of the most stubbornly resistant variables in the entire history of industrial civilization.

---

Chapter 4: The Semiconductor Chokepoint

The most complex manufactured object in human history is a leading-edge semiconductor. The process of creating one involves more than a thousand individual steps, takes approximately three months from bare silicon wafer to finished chip, and requires equipment so precise that a single particle of dust can render the entire wafer defective. The ASML extreme ultraviolet lithography machine that patterns the circuits onto the silicon costs approximately $380 million per unit, weighs 180 tons, requires multiple 747 cargo flights to ship, and contains over 100,000 components sourced from hundreds of suppliers across dozens of countries. There are fewer than 200 of these machines in operation worldwide. Every frontier AI chip passes through one of them.

This is the physical foundation of the intelligence river's newest channel. Not a distributed network of interchangeable components, but a supply chain of extraordinary concentration and fragility — a pipeline in which the failure of any single link does not reduce capacity incrementally but can halt it entirely.

Smil's approach to technology assessment has always been to follow the supply chain to its source and ask: what happens if this breaks? The question is not hypothetical when applied to semiconductors. The COVID-era chip shortage demonstrated the consequences of supply chain disruption in practice: automotive production halted for months, consumer electronics were backordered for quarters, and the global economy absorbed hundreds of billions of dollars in losses. That disruption affected mature-node chips — the relatively simple semiconductors used in cars, appliances, and industrial equipment, manufactured at dozens of facilities worldwide. The AI chip supply chain is far more concentrated than the automotive chip supply chain, and the consequences of its disruption would be correspondingly more severe.

The concentration begins with design. NVIDIA designs approximately 80 to 90 percent of the GPUs used for AI training and inference. AMD competes at a distant second. Google, Amazon, Microsoft, and others are developing custom AI chips, but none has yet achieved the scale, performance, or software ecosystem of NVIDIA's CUDA platform. The design concentration means that NVIDIA's architectural decisions — which features to prioritize, which markets to serve, which customers to allocate scarce supply to — shape the physical capability of the AI revolution more directly than any other single corporate decision.

The concentration deepens at manufacturing. TSMC manufactures essentially all frontier AI chips, including NVIDIA's. TSMC's most advanced facilities are located in Taiwan, an island of 23 million people situated 100 miles from a mainland Chinese government that considers it a breakaway province. The geopolitical risk requires no elaboration. What requires elaboration is the physical impossibility of rapidly replacing TSMC's capacity if it were disrupted.

A leading-edge semiconductor fabrication facility — a "fab" — represents an investment of $20 to $40 billion and takes four to five years to build. The Arizona fab that TSMC is constructing, even with the full backing of the U.S. government through the CHIPS Act and billions in subsidies, has encountered delays. The reason is not lack of capital or political will. It is the physical complexity of replicating an ecosystem that developed over decades in Taiwan: the specialized workforce, the supplier networks, the institutional knowledge embedded in thousands of experienced engineers, the water treatment systems that produce the ultra-pure water required for wafer fabrication — eighteen megohm-centimeter purity, roughly a thousand times purer than typical drinking water.

Samsung's foundry division in South Korea operates as a secondary manufacturer for some advanced nodes, but its yields — the percentage of chips on a wafer that function correctly — have consistently lagged TSMC's, and yield is the variable that determines whether a fab is commercially viable. A fab with 50 percent yield produces half as many working chips from the same number of wafers as a fab with near-total yield, meaning it requires roughly twice the silicon, twice the energy, and twice the time to deliver the same output. Manufacturing concentration is not just a matter of geography. It is a matter of accumulated expertise that cannot be transferred by writing a check.

The lithography equipment adds another layer of concentration. ASML is the sole manufacturer of EUV lithography machines. The company's monopoly is not the result of anticompetitive behavior but of the extreme technical difficulty of producing light at a wavelength of 13.5 nanometers — generated by vaporizing molten tin droplets with a laser — and focusing it with mirrors polished to sub-nanometer precision. Developing this capability required more than two decades and tens of billions of dollars of investment. No competitor is within a decade of replicating it. If ASML's single manufacturing facility in Veldhoven, Netherlands, were disrupted — by fire, flooding, industrial accident, or geopolitical conflict — the production of frontier AI chips would halt globally until the damage was repaired.

The material inputs extend the chain of concentration further. Semiconductor manufacturing requires rare earth elements — neodymium, dysprosium, lanthanum — for components ranging from the magnets in hard drives to the polishing compounds used in wafer preparation. China controls approximately 60 percent of global rare earth mining and approximately 90 percent of rare earth processing. Neon gas, used in the lithography process, was historically sourced primarily from Ukraine and Russia — a supply chain that was disrupted by the war in 2022 and that has since been partially diversified but not replaced. Gallium and germanium, critical for advanced semiconductor applications, are subject to Chinese export controls imposed in 2023.

Each of these concentration points represents a single point of failure in the supply chain that makes AI computation physically possible. The probability of any single disruption may be low in any given year. But the probability of at least one disruption across the entire chain over a five-to-ten-year period is not low. It is, by the assessment of supply chain analysts and geopolitical risk consultants, significant. And the consequence of a major disruption would not be a gradual reduction in AI capability. It would be a sudden constraint — a hard ceiling on the physical production of the chips on which every AI model, every Claude Code session, every twenty-fold productivity multiplier depends.

The Orange Pill devotes a chapter to the Software Death Cross — the repricing of software company valuations as AI makes code a commodity. The analysis is perceptive about the software layer. What it does not examine is the asymmetry between software abundance and hardware scarcity that defines the economic reality of the AI revolution. Software can be copied at zero marginal cost. A codebase that works can be deployed to a million users for the cost of the infrastructure to serve them. Chips cannot be copied. Each GPU must be physically manufactured, tested, packaged, and shipped. The marginal cost of producing one additional H100 includes the silicon, the energy, the water, the lithography time, the packaging materials, the testing equipment, and the labor of workers at every stage of the chain. The cost is on the order of tens of thousands of dollars per chip, and the production capacity is finite, constrained by the number of EUV machines in operation, the number of wafer starts per month at TSMC, and the yield rates achieved on each wafer.

The result is that the AI revolution, at its physical foundation, is not abundant. It is scarce. The scarcity is masked, for the moment, by the enormous capital investments of the technology companies — the hundreds of billions being poured into data center construction and chip procurement. But capital investment does not create physical capacity instantaneously. It initiates a construction process that takes years. And during those years, the scarcity of chips, the scarcity of data center capacity, and the scarcity of grid power to feed the data centers will determine which users, which companies, which countries, and which applications have access to the AI revolution and which do not.

Smil wrote in Invention and Innovation that the quest for artificial intelligence is an enormously complex, multifaceted process whose progress must be measured across decades and generations. The semiconductor supply chain is one of the reasons this assessment, which some critics dismissed as excessively conservative when the book was published in 2023, deserves revisiting with greater respect. The software can advance as fast as researchers can think. The hardware advances as fast as TSMC can build fabs, ASML can ship lithography machines, and the global supply chain can deliver the hundreds of materials required to turn sand into the most complex objects human civilization has ever produced.

The intelligence river flows through silicon. The silicon flows through a supply chain of extraordinary concentration. The concentration is a vulnerability, and the vulnerability is physical — not amenable to software patches, not addressable by algorithmic efficiency, not resolvable by narrative enthusiasm about the democratization of capability. It is a constraint that must be understood, planned for, and — in Smil's characteristic prescription — respected, because the physical world does not negotiate with aspiration. It simply imposes its terms.

Chapter 5: Water, Cooling, and the Thermodynamics of Computation

Every computation generates heat. This is not a design flaw. It is a consequence of the second law of thermodynamics — the law that governs every energy conversion in the universe, from the fusion reactions inside stars to the electrical impulses inside a graphics processing unit. No energy transformation is perfectly efficient. Some fraction of the input energy is always degraded to heat. In a modern GPU running AI inference at full load, roughly 30 to 40 percent of the electrical energy consumed is converted to useful computation. The rest becomes heat. The heat must go somewhere, because if it does not, the chip's temperature rises until it throttles its own performance, damages its circuits, or fails entirely.

This is the thermodynamic foundation of a problem that the AI discourse treats as a footnote and that Smil's framework treats as a defining constraint: cooling. Every data center is, at its most fundamental level, a facility for converting electricity into computation and heat, and then removing the heat fast enough that the computation can continue. The design of a data center is as much about thermal management as it is about computation. The cooling system is not auxiliary to the computing system. It is integral — and in many cases, it consumes 30 to 40 percent of the facility's total electricity, which means that for every watt used for computation, an additional fraction of a watt is used simply to remove the heat that the computation generated.

The most common cooling method for large-scale data centers is evaporative cooling — essentially, the same principle that makes sweating effective. Water is circulated through cooling towers, where it absorbs heat from the facility's air-handling system and then evaporates, carrying the heat into the atmosphere as water vapor. The process is thermodynamically efficient, which is why it remains the dominant method. It is also water-intensive to a degree that creates resource conflicts in water-scarce regions.

The numbers are specific and growing. Microsoft's 2023 environmental report disclosed that the company's global water consumption increased by 34 percent year-over-year, rising to approximately 6.4 billion liters — roughly 1.7 billion gallons. The company attributed the increase primarily to the growth of AI workloads, particularly the training and inference demands of large language models deployed through its partnership with OpenAI. Google reported a comparable increase, with water consumption rising 20 percent year-over-year to approximately 5.6 billion gallons. These figures represent the direct water consumption of the companies' data center operations. They do not include the indirect water consumption of the power plants that generate the electricity those data centers consume — thermoelectric power plants, whether natural gas, coal, or nuclear, are among the largest consumers of water in any national economy, using water for steam generation and cooling in quantities that dwarf the direct consumption of the data centers they serve.

A single large data center campus using evaporative cooling can consume one to five million gallons of water per day, depending on the local climate, the facility's computational load, and the efficiency of its cooling design. For context, one million gallons per day is roughly the residential water consumption of a community of 10,000 to 20,000 people. The comparison is not precise — residential water consumption varies by region, climate, and household size — but it establishes the order of magnitude. A major data center campus consumes water at a rate comparable to a small town. A cluster of data centers in a region like northern Virginia, where dozens of facilities operate in close proximity, consumes water at a rate comparable to a small city.

This consumption occurs in a world where water scarcity is intensifying. The United Nations estimates that roughly two billion people live in countries experiencing water stress, and the number is projected to increase as climate change alters precipitation patterns, glacial melt reduces river flows, and population growth increases demand. In the American West — Arizona, Nevada, parts of California and Texas — water allocation is already a source of intense political conflict among agricultural, residential, industrial, and environmental interests. Data centers are a relatively new entrant to this competition, and their water demand is growing faster than any other category.

The conflict is not theoretical. In 2022, residents of The Dalles, Oregon, protested Google's data center expansion on the grounds that the facility's water consumption — approximately a quarter of the city's water supply — was unsustainable in a region experiencing drought. In Chandler, Arizona, residents raised similar objections to Microsoft's data center water consumption. In Chile, a proposed data center in the Atacama Desert region faced opposition from agricultural communities concerned about competition for an already scarce water supply. These are local disputes, but they reflect a global pattern: as data center construction expands into regions where water is limited, the resource conflict between computation and other uses — drinking water, irrigation, industrial processes, ecosystem maintenance — will intensify.

Alternative cooling technologies exist, and their development is accelerating in response to water constraints. Air cooling, which uses fans and heat exchangers to dissipate heat without evaporating water, eliminates direct water consumption but is less thermodynamically efficient, particularly in warm climates where the ambient air temperature is already close to the temperature of the exhaust air from the servers. The efficiency penalty translates directly into higher electricity consumption, which increases the facility's energy footprint and, indirectly, its water footprint through the power generation system. Liquid immersion cooling, in which servers are submerged in a dielectric fluid that absorbs and transfers heat more efficiently than air, is a promising technology that several companies are deploying at pilot scale. It reduces or eliminates water consumption for cooling and can improve energy efficiency by reducing the need for fans and air handling. But it requires redesigning server hardware for immersion, it introduces new maintenance challenges, and it is not yet deployed at the scale of the largest AI data center campuses.

The thermodynamic constraint operates at a level that is indifferent to human ingenuity. No cooling technology can reduce the heat generated by computation — only the heat per unit of computation, through more efficient chips. Chip efficiency has improved dramatically over the decades, following a trajectory that parallels Moore's Law in some respects, but the improvement has been outpaced by the growth in computational demand. Each generation of AI models requires more computation than the last — more parameters, more training data, more inference operations per query — and the net result is that total heat generation from AI workloads is increasing even as the heat per operation decreases. The efficiency gains are real. They are also insufficient to offset the growth in demand.

Smil has documented this pattern across every energy-consuming technology he has studied. It is, in essence, the thermodynamic expression of the Jevons Paradox: improvements in the energy efficiency of a process tend to increase the total energy consumption of that process, because efficiency makes the process cheaper and more accessible, which expands demand. Air conditioning units have become dramatically more efficient over fifty years — the energy required to cool a given volume of air has fallen by roughly half. But the total energy consumed by air conditioning globally has increased by a factor of four over the same period, because efficiency made air conditioning affordable in regions and buildings where it was previously too expensive. The efficiency improved. The total consumption grew. The thermodynamic footprint expanded.

The same dynamic applies to AI computation. Chips are more efficient per operation than they were five years ago. The number of operations required to serve the growing user base has increased by orders of magnitude. The net energy consumption — and the net heat generation, and the net cooling demand, and the net water consumption — has grown, not shrunk. The efficiency frontier is moving. The demand frontier is moving faster.

The Orange Pill describes the democratization of capability — the expansion of who gets to build, the developer in Lagos accessing the same leverage as the engineer at Google. The claim operates at the software layer, where it is substantially correct. At the thermodynamic layer, the claim encounters a physical constraint. Every additional user, every additional query, every additional developer accessing AI tools generates additional heat that must be removed by cooling systems that consume water and electricity in quantities that are fixed by thermodynamic law. Democratization at scale means heat generation at scale, which means cooling at scale, which means water consumption at scale. The software can be copied infinitely at zero marginal cost. The physical requirements of serving that software cannot be.

The thermodynamics of computation are among the hardest constraints in the AI revolution — harder than financial constraints, harder than regulatory constraints, harder in some respects than supply chain constraints. Financial constraints can be relaxed by investment. Regulatory constraints can be modified by policy. Supply chain constraints can be addressed, slowly, by diversification. Thermodynamic constraints are governed by physics. The second law does not negotiate. It does not respond to capital investment, political pressure, or narrative enthusiasm. It simply operates, converting a fraction of every joule of electrical energy into heat that must be managed, at a cost that is measured in watts and gallons and that grows in direct proportion to the scale of computation.

Any honest accounting of the AI revolution must include this cost. Not as an afterthought, not as a line item buried in an environmental report, but as a defining parameter of the revolution's physical reality. The imagination-to-artifact ratio may approach zero for the user sitting at a screen in Trivandrum. Behind that screen, the thermodynamic ratio — the energy consumed per unit of useful computation, the water evaporated per unit of heat removed, the carbon emitted per unit of electricity generated — remains stubbornly, physically, immutably greater than zero. And it is the thermodynamic ratio, not the imagination-to-artifact ratio, that will determine the long-term physical sustainability of the intelligence river's newest channel.

---

Chapter 6: The Democratization Gap

The most compelling argument in The Orange Pill is also the one most vulnerable to quantitative scrutiny. Edo Segal describes a developer in Lagos who can now access the same coding leverage as an engineer at Google — the same AI tools, the same capability to translate ideas into working software through natural-language conversation. The moral weight of the argument is substantial. If AI genuinely lowers the floor of who gets to build, the implications for global equity are profound. If a brilliant person in Dhaka or Nairobi or São Paulo can now create what previously required a team in San Francisco, the reallocation of creative opportunity is among the most significant redistributions in the history of technology.

The argument must be tested against the infrastructure that makes it possible.

Smil's method is to follow the claim to its physical requirements and then measure whether those requirements are met. The developer in Lagos requires, at minimum, four things: reliable electricity, sufficient bandwidth with acceptable latency, a device capable of running the AI interface, and the financial capacity to pay for the service. Each of these requirements can be quantified, and each reveals a gap between the software promise and the infrastructure reality.

Electricity first. Nigeria's installed generation capacity is approximately 12.5 gigawatts, of which roughly 4 to 5 gigawatts is typically available — the rest is offline due to maintenance, fuel shortages, or infrastructure failures. For a country of over 220 million people, this translates to an available capacity of roughly 20 to 25 watts per person. The United States, by comparison, has an installed capacity of approximately 1,300 gigawatts for 340 million people — roughly 3,800 watts per person, more than 150 times the Nigerian figure. The practical consequence is that electrical service in Lagos is intermittent. Rolling blackouts are routine. Businesses and affluent households rely on diesel generators as backup, at a cost several times higher than grid electricity. The developer in Lagos who opens Claude Code is not accessing the tool from the same infrastructure as the engineer in Mountain View. She is accessing it from a grid that may lose power at any moment, requiring her to maintain and fuel a personal generator — an additional cost, measured in dollars and in the noise and exhaust fumes that are the sensory reality of generator-dependent electricity in a dense urban environment.

Sub-Saharan Africa's average per-capita electricity consumption is approximately 500 kilowatt-hours per year — roughly one-fiftieth of the United States' average of approximately 12,000 kilowatt-hours. This is not a gap that can be closed by software deployment. It is a gap that can be closed only by building power plants, transmission lines, and distribution networks — infrastructure that, as the preceding chapters have documented, takes years to decades to construct. The software arrives at the speed of a download. The electricity required to use the software reliably arrives at the speed of civil engineering.

Bandwidth second. The average fixed broadband speed in Nigeria is approximately 25 to 35 megabits per second — roughly one-tenth of the United States average. Mobile broadband, which is the primary internet access method for most users in sub-Saharan Africa, averages lower, with significant variation by location, time of day, and network congestion. AI coding assistants require sustained, low-latency connections for the real-time conversational interaction that makes them effective. A response that takes ten seconds instead of two — because the data must travel from Lagos to a data center in Europe or the United States and back, through networks of variable quality — is not merely slower. It disrupts the cognitive flow that The Orange Pill identifies as the core value of the human-AI collaboration. The imagination-to-artifact ratio does not approach zero if the conversation stutters. It approaches the ratio of a long-distance phone call with a bad connection — still functional, but fundamentally different in character from the seamless interaction that defines the experience for users in high-bandwidth, low-latency environments.

Data center proximity matters for latency, and the geography of data center deployment reflects the geography of economic power. The vast majority of hyperscale data centers are located in North America, Europe, and East Asia. Sub-Saharan Africa has a small and growing number of data centers, concentrated in South Africa, Kenya, and Nigeria, but the total capacity is a fraction of what exists in northern Virginia alone. The physical distance between the developer in Lagos and the nearest data center capable of running frontier AI inference imposes a latency penalty that is governed by the speed of light through fiber optic cable — approximately five milliseconds per thousand kilometers of cable, plus the processing time at each network hop. Lagos to London is roughly 5,000 kilometers, implying a minimum round-trip latency of 50 milliseconds or more, before accounting for network congestion, routing inefficiencies, and the processing time at the data center itself. This is noticeable. For the kind of rapid, iterative, conversational interaction that defines the Claude Code experience Segal describes, it is a meaningful degradation.

Device cost third. The developer in Lagos needs a computer capable of running a modern web browser with sufficient performance to support the AI interface. A basic laptop meeting these requirements costs $300 to $500 — a figure that is modest by American standards and significant by Nigerian standards, where the median monthly wage is approximately $150 to $200. The laptop represents one to three months of median income, compared to a few days of median income in the United States. Mobile devices are more affordable and more widely available, but the AI coding interfaces are designed primarily for desktop or laptop use, and the experience on a mobile device is substantially limited.

Financial access fourth. The hundred-dollar-per-month subscription that Segal describes as the price of a twenty-fold productivity multiplier represents approximately 50 to 70 percent of the median monthly wage in Nigeria. In the United States, the same subscription represents less than one percent of the median monthly wage. The disparity is not incidental. It means that the developer in Lagos, even if she has reliable electricity, sufficient bandwidth, and an adequate device, faces a financial barrier that makes the subscription a major economic commitment rather than the modest expense it represents for users in high-income countries.

Segal acknowledges these barriers. The Orange Pill notes that access requires connectivity, hardware, and English-language fluency, and that these barriers will fall as models improve and costs decrease. The acknowledgment is honest. The question Smil's framework raises is one of timeline and magnitude. How fast will costs decrease? How quickly will infrastructure expand? And what happens in the interim — the years or decades during which the software promise outpaces the infrastructure reality?

The historical precedent is not encouraging. Mobile phone adoption in sub-Saharan Africa is the most frequently cited example of rapid technology diffusion in the developing world, and it is genuinely impressive — mobile subscriptions grew from near zero in 2000 to over 80 percent penetration by 2020. But mobile phone adoption succeeded in part because it did not require fixed infrastructure. A cellular tower serves a wide area, and the phone itself is a self-contained device that charges from any electricity source. Internet access — particularly the reliable, high-bandwidth, low-latency access that AI tools require — depends on fixed infrastructure that is far more expensive and complex to deploy. Submarine cables, terrestrial fiber networks, data centers, reliable grid power — these are the prerequisites, and their deployment in sub-Saharan Africa is measured in billions of dollars and years of construction time.

The democratization of capability is real at the software layer. The developer in Lagos can, in principle, access the same AI tools as the engineer in Mountain View. In practice, she accesses them through infrastructure that is less reliable, slower, more expensive relative to income, and subject to interruptions that the engineer in Mountain View never experiences. The gap is not a reason to dismiss the democratization argument. It is a reason to measure the gap honestly, to invest in closing it, and to resist the temptation to describe the software layer's potential as though it were already the infrastructure layer's reality.

Smil's insistence on quantitative realism is not pessimism about Africa or the developing world. It is the recognition that genuine democratization requires physical investment — power plants, cables, data centers, grid upgrades — that takes time, capital, and political will. Declaring the democratization accomplished because the software is available is like declaring hunger solved because the recipe exists. The recipe matters. The ingredients, the kitchen, and the stove matter more.

---

Chapter 7: The Jevons Paradox of Intelligence

In 1865, the English economist William Stanley Jevons published The Coal Question, in which he made an observation so counterintuitive that it has been debated for 160 years. The observation was this: as the efficiency of coal-burning engines improved — as each ton of coal produced more useful work — the total consumption of coal did not decrease. It increased. The efficiency gains made coal-powered machinery cheaper to operate, which expanded the range of applications for which coal power was economical, which increased the total demand for coal by more than the efficiency gains reduced the demand per application.

The pattern has repeated across virtually every energy technology Smil has studied. Automobile fuel efficiency has improved by roughly 30 percent since the 1980s — and total gasoline consumption in the United States rose over the same period, because the improved efficiency made driving cheaper per mile, which encouraged more driving, larger vehicles, and longer commutes. Lighting efficiency has improved by a factor of roughly twenty since the incandescent bulb — from roughly 12 lumens per watt to over 200 lumens per watt for modern LEDs — and global electricity consumption for lighting has increased by a factor of more than three over the same period, because cheap, efficient lighting was deployed in vastly more locations, for longer hours, at higher intensities. The refrigerator in a typical American kitchen is roughly 75 percent more efficient than its 1975 counterpart, measured in kilowatt-hours per cubic foot of cooled space — and the total electricity consumed by American refrigeration has remained roughly constant, because the average refrigerator has grown substantially in size and many households now own more than one.

The pattern is not universal. There are cases where efficiency gains do reduce total consumption — but these tend to occur in mature, saturated markets where demand has no room to expand. When demand has room to expand — when efficiency makes a resource accessible to new users, new applications, or new scales of operation — the rebound effect dominates, and total consumption increases.

The AI productivity literature contains, embedded in its most celebrated findings, a precise demonstration of the Jevons Paradox applied to cognition. The Berkeley study that The Orange Pill cites in Chapter 11 — the eight-month ethnographic study by Ye and Ranganathan — found that workers who adopted AI tools did not work less. They worked more. They took on additional tasks. They expanded into adjacent domains. They filled pauses with prompts. The efficiency gains from AI did not create leisure. They created appetite. The work expanded to consume the time that the efficiency had freed.

Segal acknowledges this finding. He cites it as evidence of the need for what the Berkeley researchers called "AI Practice" — structured pauses, sequenced workflows, protected time for reflection. The diagnosis is correct. The physical implication of the diagnosis is what Smil's framework makes explicit.

If AI tools make knowledge workers twenty times more productive — the figure Segal claims based on his Trivandrum experience — and those workers respond by doing twenty times as much work, the computational demand on the AI system scales proportionally. Twenty times the output means, approximately, twenty times the inference queries, twenty times the token generation, twenty times the GPU-hours consumed at the data center, twenty times the electricity drawn from the grid, twenty times the heat generated and the water evaporated to remove it. The proportionality is not exact — some work expanded through AI involves tasks that are computationally cheaper than others, and some efficiency gains come from better prompting that requires fewer round trips. But the direction is clear: the Jevons Paradox predicts that AI efficiency gains, measured at the level of individual productivity, will translate into increased aggregate computational demand, not decreased demand. And the physical infrastructure must scale to meet that demand, or the demand will be rationed by scarcity.

The evidence is accumulating. The IEA projects that global data center electricity consumption will more than double between 2022 and 2026. The increase is driven primarily by AI workloads. In the United States, data center electricity demand is projected to grow from approximately 4.4 percent of national consumption in 2025 to potentially 8 to 9 percent by 2030. These projections represent aggregate demand — the sum of all AI users, all training runs, all inference queries, all the prompts and responses and collaborations that constitute the AI revolution at the level of physical resource consumption.

The Jevons Paradox operates at the individual level and the systemic level simultaneously, and the systemic effects are larger than the individual effects because they include the expansion of the user base itself. The developer in Trivandrum who becomes twenty times more productive is one source of demand growth. The expansion of AI tools to new industries — healthcare, legal services, education, government, scientific research — is another. The growth of the user base from millions to hundreds of millions to potentially billions is a third. Each source of growth adds computational demand that must be met by physical infrastructure.

The standard response to this concern, from the technology industry, is that efficiency improvements in chip design, model architecture, and data center operations will reduce the per-query cost of AI computation faster than demand grows, resulting in a net decrease in resource intensity per unit of useful output. This is the efficiency argument — the same argument that has been made for coal, gasoline, lighting, and refrigeration, and that has, in every case Smil has documented, failed to reduce total consumption when demand is elastic.

The question is whether AI demand is elastic — whether the availability of cheaper, faster, more efficient AI tools will continue to generate new demand, new users, new applications, or whether there is a natural saturation point at which people simply stop wanting more AI assistance. The evidence from the first year of widespread AI adoption suggests that demand is highly elastic. ChatGPT went from zero to 100 million users in two months. Claude Code crossed $2.5 billion in run-rate revenue within months. The Berkeley study documented workers expanding their use of AI tools into every available moment, including lunch breaks and elevator rides. There is no evidence, yet, of saturation. There is substantial evidence of appetite.

Smil has argued, across his work on energy transitions, that the Jevons Paradox is not a curiosity or an anomaly but a fundamental feature of how efficiency improvements interact with demand in expanding markets. The implication for AI is direct and uncomfortable: the productivity gains that The Orange Pill celebrates as the hallmark of the AI revolution are simultaneously the mechanism by which the revolution's physical footprint expands. Every twenty-fold productivity gain, if it is absorbed into expanded output rather than reduced working hours, translates into expanded computational demand. The river of intelligence widens, and the physical infrastructure required to sustain it widens with it.

The resolution of the Jevons Paradox, in every historical case, has come not from efficiency alone but from the combination of efficiency, institutional structure, and — eventually — resource constraints that impose a ceiling on demand growth. The eight-hour workday was, in Smil's framework, a dam against the Jevons Paradox of industrial productivity — a social institution that prevented efficiency gains from being fully absorbed into expanded work, preserving some of the gains for leisure, health, and human development. The question for the AI revolution is whether comparable institutions will be built — dams that redirect some portion of the productivity gains away from expanded computational demand and toward the human outcomes that the revolution is ostensibly designed to serve.

Without such institutions, the trajectory is clear from the physics. Efficiency improves. Demand expands. Total consumption grows. The river widens. The physical infrastructure strains to keep pace. The strain is measured in gigawatts, in millions of gallons, in tons of carbon, in years of construction time. The numbers do not lie about any of this, and the numbers, as Smil has insisted for five decades, are where any honest assessment must begin.

---

Chapter 8: S-Curves and the Limits of Extrapolation

Every technology that has ever been adopted by human societies has followed the same mathematical shape: the S-curve. Slow initial adoption as the technology finds its first users and works through its earliest limitations. Rapid growth as the technology reaches the mainstream, as infrastructure scales, as costs decline, as the benefits become widely visible. Then deceleration, as the technology approaches the limits of its addressable market, encounters physical constraints, or is superseded by something newer.

The S-curve is not a theory. It is an empirical observation, documented across thousands of technologies by Smil and others. Radio followed it. Television followed it. The automobile followed it. The personal computer followed it. Mobile phones followed it. In every case, the steep middle section of the curve — the period of rapid growth — was the period that generated the most excitement, the most investment, and the most confident predictions of permanent exponential progress. And in every case, the predictions were wrong, because exponential growth is physically impossible to sustain indefinitely. It encounters constraints. Resources run short. Markets saturate. Infrastructure cannot scale fast enough. The curve bends.

The tendency to extrapolate from the steep middle section as though the growth rate were permanent is one of the most persistent analytical errors in the technology discourse. Smil has identified it across dozens of examples, from nuclear power (which was projected in the 1960s to provide the majority of global electricity by 2000 and in fact provides roughly 10 percent) to supersonic air travel (which was expected to replace subsonic flight and instead disappeared entirely with the Concorde) to hydrogen fuel cells for automobiles (which have been "five to ten years away" for approximately thirty years).

The AI adoption data is spectacular and real. ChatGPT reached 100 million monthly active users approximately two months after its public release in November 2022 — faster than any consumer application in history. Claude Code crossed $2.5 billion in run-rate revenue within months of its December 2025 breakthrough. Google reported that 25 to 30 percent of its new code was AI-assisted. Industry estimates placed the aggregate figure for AI-assisted code across the technology sector at over 40 percent in 2025, with projections crossing 50 percent by late 2026.

These figures represent the steep middle section of the S-curve. They are measurements of real adoption by real users generating real value. They are not fabricated or exaggerated. And they are precisely the kind of data points from which premature extrapolation most frequently occurs.

The extrapolation takes a characteristic form. If AI-assisted code went from negligible to 40 percent in two years, then it will reach 90 percent in two more years. If Claude Code revenue went from zero to $2.5 billion in months, then it will reach $25 billion by next year. If ChatGPT reached 100 million users in two months, then it will reach a billion users by the end of the decade. The logic feels compelling because the numbers are so large and the growth so rapid. But the logic is the same logic that projected nuclear power would dominate by 2000, and the logic was wrong then for the same reasons it may be wrong now: because the steep section of the S-curve is not the whole curve, and the factors that determine where the curve bends are physical, not mathematical.

What bends the AI S-curve? The constraints documented in the preceding chapters of this volume: energy availability, data center construction timelines, semiconductor supply chain capacity, water resources, grid infrastructure, and the financial sustainability of pricing models that currently operate below the cost of inference. Each of these constraints imposes a ceiling on the rate at which AI deployment can expand, and the ceilings are physical — determined by the pace of construction, the capacity of supply chains, and the thermodynamics of computation, not by the pace of algorithmic improvement.

The energy constraint alone may be sufficient to bend the curve. If AI data center electricity demand in the United States is projected to roughly double by 2030 — from approximately 4.4 percent to 8 or 9 percent of national consumption — the grid must add the equivalent of fifty gigawatts of generation capacity in approximately five years. Smil's analysis of historical generation capacity additions suggests that this pace would be unprecedented in the absence of a wartime mobilization. The United States added roughly 100 gigawatts of natural gas generation capacity over the entire two decades from 2000 to 2020. Adding 50 gigawatts in five years — even combining gas, nuclear, solar, wind, and other sources — requires a construction pace that exceeds recent historical experience.

If the grid capacity is not added in time, the result is not catastrophe. It is rationing — the allocation of scarce electricity to the highest-value, highest-willingness-to-pay users, which in practice means the largest technology companies, at the expense of smaller companies, startups, and individual developers. The democratization promise encounters a physical constraint that favors incumbents over entrants, concentration over distribution, wealth over aspiration.

The semiconductor constraint operates on a similar timeline. TSMC's expansion into the United States, Intel's foundry ambitions, Samsung's advanced-node investments — all are measured in three-to-five-year construction cycles, billions of dollars per facility, and workforce development timelines that cannot be compressed by capital alone. The total global manufacturing capacity for frontier AI chips is growing, but growing from a small base, and the demand for those chips is growing faster than the capacity to produce them. NVIDIA's allocation of its most advanced GPUs — the systems that command six-figure prices and multi-month waiting lists — is itself a form of rationing, a market mechanism that distributes scarce physical resources according to willingness to pay.

Smil's contribution to the discourse is not to predict when the S-curve will bend. Prediction, in his framework, is a fool's errand — the history of technology forecasting is a history of confident predictions that were wrong, in both directions, by margins that should inspire humility. His contribution is to identify the physical factors that will determine where the bend occurs and to insist that any planning framework that assumes permanent exponential growth is not planning at all. It is hoping.

Growth rates are not destiny. They are measurements of a specific period, under specific conditions, subject to constraints that may not yet be binding but that will, eventually, assert themselves. The telephone took 75 years to reach 50 million users. Radio took 38. Television took 13. The internet took 4. ChatGPT took 2 months. The acceleration is real, and it reflects genuine improvements in the speed of technology diffusion. But diffusion speed and deployment sustainability are different variables. A technology can be adopted faster than the infrastructure to support it can be built, and when that happens, the adoption curve bends — not because people stop wanting the technology, but because the physical world cannot deliver it fast enough.

The AI S-curve will bend. The question is when, at what level, and what happens to the institutions — the companies, the investment strategies, the educational programs, the government policies — that were built on the assumption that the steep middle section would last forever. The history of technology S-curves is full of institutions that rode the steep section with confidence and were unprepared for the bend. The nuclear industry's excess capacity after the projected demand never materialized. The telecommunications companies that overbuit fiber capacity in the late 1990s. The cleantech companies that assumed government subsidies would scale indefinitely.

Smil's counsel is not to avoid the steep section — that is where the genuine gains occur. His counsel is to plan for the bend. To build institutions flexible enough to adapt when growth decelerates. To avoid concentrating investment on the assumption that current rates will continue. To maintain the quantitative discipline that distinguishes between what the numbers show today and what they guarantee about tomorrow. The numbers today show extraordinary growth. They do not guarantee extraordinary growth will continue, because no S-curve in the history of human technology has ever maintained its steepest gradient indefinitely. The physics does not allow it. The infrastructure does not support it. The resources do not sustain it.

The curve will bend. The only question is whether the builders will be ready.

Chapter 9: The Honest Transition

The printing press arrived in Mainz around 1440. By 1500, roughly sixty years later, an estimated twenty million volumes had been printed across Europe. The speed was remarkable by medieval standards. It was also, by the standards of the AI discourse, glacially slow. Sixty years to produce twenty million books. ChatGPT acquired 100 million users in sixty days.

The comparison is deployed routinely in the AI literature — The Orange Pill traces the acceleration from telephone (75 years to reach 50 million users) through radio (38 years), television (13 years), the internet (4 years), to ChatGPT (2 months) — and the trajectory it describes is real. The speed of technology diffusion has increased by orders of magnitude across five centuries. But Smil's framework introduces a distinction that the acceleration narrative tends to elide: the difference between adoption and integration.

Adoption is the moment a user first engages with a technology. Integration is the moment the technology has been absorbed into the institutional, economic, educational, and physical fabric of a society deeply enough that its effects are fully realized and its disruptions fully managed. Adoption can happen in months. Integration takes decades. And it is integration, not adoption, that determines whether a technology's transformative potential is realized broadly or captured narrowly.

Segal's five-stage pattern — threshold, exhilaration, resistance, adaptation, expansion — is historically defensible and analytically useful. The pattern recurs across every major technological transition. The question Smil's framework raises is not whether the pattern is correct but whether the implied timeline is realistic. The Orange Pill places AI in Stage Four — adaptation — and implies that expansion is imminent. The historical evidence, examined quantitatively, suggests that Stage Four is longer, more expensive, and more institutionally demanding than Stage Three suggests.

The printing press provides the longest-running case study. The technology crossed its threshold around 1450. The exhilaration was immediate — scholars and clerics recognized the power of mechanical reproduction within years. Resistance followed, particularly from the Church, which recognized that uncontrolled access to printed material threatened its monopoly on scriptural interpretation. But adaptation — the construction of the institutional infrastructure required to realize the technology's potential — took centuries. The Index Librorum Prohibitorum, the Catholic Church's list of banned books, was not established until 1559, more than a century after the first press. The concept of copyright did not emerge in its modern form until the Statute of Anne in 1710, 260 years after Gutenberg. Universal literacy — the precondition for the printing press to achieve its full democratic potential — was not reached in Western Europe until the late nineteenth century, roughly 400 years after the technology arrived.

Four hundred years from threshold to full expansion. The timeline is extreme, and it would be foolish to suggest that AI integration will take centuries — the pace of institutional change has itself accelerated, along with the pace of technology diffusion. But it would be equally foolish to suggest that the institutional changes required for AI integration can be compressed into quarters or even years.

Smil has documented the adaptation timelines of more recent transitions with characteristic precision. Rural electrification in the United States began in the 1880s with Edison's Pearl Street Station. By 1925, roughly 40 years later, only about half of urban American homes had electricity, and fewer than 10 percent of rural homes. The Rural Electrification Administration was not created until 1935, and near-universal electrification was not achieved until the early 1960s — approximately 80 years after the technology's commercial introduction. The adaptation phase, measured from the establishment of the institutional framework (the REA) to near-universal access, took roughly 25 years. Measured from the technology's commercial debut to universal access, it took 80 years.

The automobile tells a similar story with a slightly compressed timeline. The Model T entered production in 1908. By 1920, roughly 10 million cars were registered in the United States — rapid adoption by any standard. But the institutional adaptation — paved roads, traffic regulations, driver licensing, insurance frameworks, urban planning that accommodated automobiles, the interstate highway system — took decades. The Federal-Aid Highway Act was passed in 1956, nearly 50 years after the Model T. The interstate highway system was not substantially complete until the 1980s, 70 years after mass automobile adoption began. The adaptation phase — the construction of the physical and institutional infrastructure required to realize the automobile's full potential — lasted longer than the adoption phase by a factor of five or more.

The spreadsheet transition is the most frequently cited precedent in the AI discourse because it is the most recent and the most analogous — a software tool that transformed knowledge work. VisiCalc was released in 1979. By 1985, spreadsheets were widely used in finance and accounting. But the full integration of spreadsheet-based analysis into business practice — the development of financial modeling standards, the retraining of the accounting workforce, the regulatory adaptation to spreadsheet-based reporting, the cultural shift from paper-based to electronic analysis — took roughly a decade beyond initial adoption. The accounting profession did not shrink. It grew. But the growth required institutional adaptation that took years, not months.

Each of these transitions followed Segal's five-stage pattern. Each produced genuine expansion. And each required an adaptation phase measured in years or decades — a period during which the technology was already adopted but its full potential was unrealized because the institutional, educational, and physical infrastructure had not yet caught up.

The AI transition has characteristics that may accelerate the adaptation phase relative to prior transitions. The technology is software-native, which means it can be deployed and updated without physical construction for the user-facing layer. The interface is natural language, which means the retraining barrier is lower than for any previous computing technology — users do not need to learn a new language or interface paradigm. The productivity gains are immediate and visible, which creates strong incentives for rapid institutional adoption.

But the AI transition also has characteristics that may slow adaptation. The physical infrastructure requirements, documented in the preceding chapters, introduce construction timelines that software cannot compress. The workforce implications are broader and deeper than any previous software transition — not just accountants and programmers but lawyers, doctors, teachers, designers, analysts, managers, and potentially every category of knowledge work. The educational system must adapt not just its tools but its fundamental pedagogy, shifting from teaching answers to teaching questions, from training execution to training judgment. The regulatory framework must address not just the technology but its second-order effects on labor markets, intellectual property, privacy, security, and democratic governance.

These institutional adaptations are the dams that Segal calls for in The Orange Pill. Smil's contribution is to insist that the dams take time to build. Not software time. Physical time. Institutional time. The time required to draft legislation, build consensus, train workers, construct infrastructure, and develop the cultural norms that redirect a powerful technology toward broad human benefit rather than narrow capture.

Segal places the AI transition in Stage Four and implies urgency. Smil would not dispute the urgency. But the urgency must be calibrated to the actual pace of institutional and physical change, not to the pace of software iteration. A society that attempts to build dams at software speed will build them poorly — hasty legislation full of unintended consequences, crash retraining programs that produce shallow competence, infrastructure shortcuts that create new vulnerabilities. A society that builds dams at the pace of careful institutional development may find that the river has already flooded the valley by the time the dams are complete.

The honest answer is that there is no historical precedent for a transition of this speed and this breadth. The printing press was broader but slower. The automobile was transformative but narrower in its immediate workforce effects. The spreadsheet was analogous but smaller in scope. AI combines the breadth of writing, the economic disruption of industrialization, and the speed of software deployment into a single transition that has no precise historical parallel.

What the historical pattern does offer, reliably and across centuries of evidence, is the finding that adaptation takes longer than exhilaration predicts, that the institutions required for broad expansion cannot be built overnight, and that the cost of the transition is borne disproportionately by the people who are least equipped to manage it — the workers whose skills are displaced, the communities whose industries are disrupted, the students whose education was designed for a world that no longer exists.

Smil has spent his career measuring the pace of transitions and finding, in every case, that the pace is slower than the advocates hoped, faster than the critics feared, and costly in ways that neither side predicted. The AI transition will, in all probability, follow the same pattern. The expansion is coming. It will not arrive as quickly as the steep section of the S-curve suggests. It will not arrive as equitably as the democratization narrative promises. And its arrival will be shaped, more than any algorithmic breakthrough, by the quality of the institutions — the dams — that are built during the adaptation phase to direct the river toward human flourishing rather than away from it.

The adaptation phase is where the work happens. Not the glamorous work of building the technology. The unglamorous work of building the institutions that make the technology serve broad human purposes. The historical evidence is unambiguous about what happens when this work is skipped or rushed. The river floods. The costs are borne by the vulnerable. The expansion, when it comes, is narrower and less equitable than it could have been.

Building the dams well takes time. The question is whether the builders will accept this, or whether the intoxication of the steep section will convince them that time is the one constraint that does not apply.

---

Chapter 10: What the Numbers Require

The preceding nine chapters have documented a set of physical constraints on the AI revolution. The constraints are not speculative. They are measured — in gigawatts, in gallons, in construction-years, in the concentration ratios of semiconductor supply chains, in the gap between software adoption speed and infrastructure deployment speed. The constraints do not invalidate the revolution. They contextualize it, in the specific sense that Smil's work has always contextualized technological claims: by asking what the physical world requires before the claim can be sustained at the scale its advocates envision.

This final chapter does not offer policy prescriptions. Smil has historically avoided prescriptive advocacy, and the Smil framework is better served by the discipline of presentation than the temptation of recommendation. What follows is an accounting — a specification of what the numbers require, stated as plainly as the numbers allow.

The numbers require energy planning that treats AI as a major new demand category, not a rounding error.

The projected addition of approximately 50 gigawatts of electrical capacity in the United States alone by 2030 — the figure Smil cited at the Bankinter webinar — is an undertaking of a scale that has no precedent in the recent history of American grid development. It is roughly equivalent to adding the total installed electrical capacity of a country like Poland or Egypt in five years. The planning must specify the generation sources (gas, nuclear, solar, wind, geothermal, or some combination), the transmission infrastructure to connect generation to data centers, the construction timelines for each component, and the financing mechanisms to fund the build-out. Without this planning, the fifty gigawatts will not materialize, and the gap between AI software capability and the physical capacity to run it will widen.

Smil insists, with characteristic directness, that the planning must also address the demand side. Advanced economies waste enormous quantities of energy — through food waste (approximately 30 to 40 percent of food produced in the United States is discarded, representing embedded energy that was consumed in production, processing, transportation, and refrigeration), through oversized vehicles (the average American passenger vehicle weighs approximately 4,300 pounds, roughly twice the weight necessary for the transportation task it performs), through poorly insulated buildings (the average American home loses 25 to 30 percent of its heating and cooling energy through inadequate insulation). Addressing these inefficiencies could free significant grid capacity for new demand categories, including AI, without requiring the full 50 gigawatts to be built from new generation. The efficiency path is faster and cheaper than the construction path. It is also less politically dramatic, which is why it receives less attention.

The numbers require water planning that acknowledges data centers as a significant new competitor for a finite resource.

In water-scarce regions — the American Southwest, parts of India, the Middle East, sub-Saharan Africa, Mediterranean Europe — the addition of data center water demand to an already strained water budget is not a minor variable. It is a resource allocation decision that affects agriculture, residential supply, ecosystem maintenance, and industrial use. The planning must specify what technologies will be used for data center cooling (evaporative, air, liquid immersion), what water sources will supply the facilities (municipal, recycled, desalinated), and what the cumulative impact of multiple facilities in a single watershed will be on the regional water balance. Without this planning, data center construction will proceed ad hoc, and the resource conflicts that have already emerged in Oregon, Arizona, and Chile will multiply.

The numbers require supply chain diversification measured in years and tens of billions of dollars.

The concentration of frontier semiconductor manufacturing in a single company in a geopolitically exposed location is a vulnerability of a magnitude that no other critical industrial supply chain exhibits. The CHIPS Act in the United States, the European Chips Act, and Japan's semiconductor investment program are real responses to this vulnerability, but they are responses measured in construction timelines of four to five years per facility and investments of $20 to $40 billion per fab. The diversification is underway. It is not complete, and it will not be complete for the better part of a decade. In the interim, the concentration remains, and with it the risk that a single disruption — geopolitical, natural, or industrial — could constrain the physical capacity of the AI revolution at its source.

The planning must extend beyond fabrication to the upstream supply chain: rare earth processing (concentrated in China), neon gas supply (historically concentrated in Ukraine), ultra-pure chemical reagents, and the specialized equipment that exists in quantities measured in hundreds of units worldwide. Each link in the chain represents a potential bottleneck, and the diversification of each link has its own timeline and its own capital requirements.

The numbers require honest accounting of the environmental cost of cognitive abundance.

Every kilowatt-hour consumed by a data center carries an embedded carbon cost determined by the energy mix of the grid that serves it. In regions where the grid is powered predominantly by fossil fuels — which includes most of the world, including parts of the United States where data center growth is fastest — the carbon cost is significant. The aggregate carbon emissions attributable to AI workloads are growing, and they are growing at a time when the global emissions trajectory needs to be declining to meet the climate commitments made in Paris and Glasgow.

The technology companies that operate the data centers have made carbon-neutral commitments, backed by renewable energy purchases and carbon offset programs. These commitments deserve scrutiny. A renewable energy purchase agreement does not mean the data center is running on renewable electricity at every hour of every day. It means the company has contracted to add renewable generation to the grid in an amount equivalent to its consumption — a financial transaction that reduces the grid's aggregate carbon intensity but does not eliminate the emissions attributable to the specific electricity the data center draws. The accounting is complex, and the complexity can obscure the gap between the commitment and the physical reality.

The honest accounting must specify: what is the total energy consumption of AI workloads, at what carbon intensity, with what trajectory, and how does that trajectory compare to the emissions reductions required to meet climate commitments? The numbers are available, or can be estimated from published data. The question is whether the institutions responsible for climate planning are incorporating AI energy demand into their projections, or whether AI demand is being treated as an externality that someone else will worry about.

The numbers require temporal realism about the pace of institutional adaptation.

The historical record, documented across multiple transitions and multiple centuries, indicates that institutional adaptation — the development of regulatory frameworks, educational curricula, workforce retraining programs, and cultural norms adequate to a new technology — takes years to decades. The AI discourse operates on a timeline of quarters. The institutions that must adapt operate on a timeline of years. The physical infrastructure that must be built operates on a timeline of decades. These are three different clocks, running at three different speeds, and the AI revolution will unfold at the pace of the slowest clock, not the fastest.

This does not mean that urgency is misplaced. The urgency is real, precisely because the institutional clock is slow and the consequences of delay are borne by the people least equipped to absorb them. The worker whose skills are displaced this quarter cannot wait for the retraining program that will be funded next year. The student whose education is being reshaped by AI tools cannot wait for the curriculum reform that will be debated for the next three years. The community whose water supply is being competed for by a new data center cannot wait for the regional water plan that will be completed in five years.

The urgency must be paired with temporal realism. Urgent action taken on the wrong timeline — regulations drafted in haste, retraining programs designed without rigor, infrastructure built without adequate planning — produces consequences that compound the problem rather than solving it. Smil's career has been a sustained argument for the proposition that realism is not the enemy of urgency. It is the prerequisite for urgency that produces results rather than chaos.

The river of intelligence flows. The numbers do not lie about what the river requires: energy, water, silicon, copper, steel, concrete, years of construction, decades of institutional development, and the continuous, unglamorous, quantitatively disciplined work of building the physical foundations on which the cognitive revolution must stand.

Smil wrote in his 2025 Pictet essay that he does not see large language models triggering fundamental transformations in society, crime, or politics. The claim may prove too conservative — Smil himself has acknowledged that he prefers to be surprised by progress rather than disappointed by its absence. But the claim is grounded in a method that has outlasted every technology hype cycle of the past half-century: the method of counting what the transformation requires before declaring it inevitable, of measuring the physical constraints before celebrating the software capabilities, of asking not just what is possible but what is feasible, at what cost, on what timeline, with what resources.

The numbers require realism. Not pessimism. Not the reflexive skepticism that dismisses genuine capability because the hype surrounding it is excessive. Realism: the intellectual discipline of holding the extraordinary promise of AI in one hand and the physical constraints of the world it must inhabit in the other, and refusing to drop either one.

The promise is real. The constraints are real. The numbers describe both. And the numbers, as Smil has insisted for fifty years, do not lie.

---

Epilogue

Fifty gigawatts.

That number has lived in my head since I first encountered Smil's Bankinter presentation. Fifty gigawatts of new electrical capacity needed in the United States alone, within five years, just to power the AI revolution I have been celebrating. Fifty cities of a million people each. I have built companies. I have shipped products. I have sat in rooms where we debated whether a feature should launch next week or the week after. At no point in those conversations did anyone mention the electrical grid.

That is Smil's gift, and it is an uncomfortable one. He does not tell you that you are wrong. He tells you what you forgot to count.

I did not count the water evaporating through cooling towers every time my team in Trivandrum opened Claude Code. I did not count the kilowatt-hours consumed by the inference that made our thirty-day sprint to CES possible. I did not count the neon gas from a handful of global suppliers, or the ultra-pure water a thousand times purer than what comes out of my kitchen tap, or the three-to-five years it takes to build the fabrication plant that produces the chip that processes the token that completes my sentence. I counted the output. I measured the productivity. I felt the exhilaration. And the physical substrate that made all of it possible was invisible to me — not because it was hidden, but because I was not looking.

In The Orange Pill, I wrote about fishbowls — the sets of assumptions so familiar you stop noticing them. Smil cracked a fishbowl I did not know I was swimming in: the assumption that the digital is weightless. That intelligence, because it moves at the speed of light through fiber optic cable, has no mass, no footprint, no claim on the physical world. Every chapter of this volume is a correction to that assumption, delivered not with rhetoric but with units. Gigawatt-hours. Millions of gallons. Tons of silicon. Years of construction time.

I still believe the core argument of The Orange Pill. AI is an amplifier, the most powerful one ever built. The question remains: are you worth amplifying? But Smil has added a question I had not adequately confronted — one that sits alongside mine and will not be dismissed. Can the physical world sustain the amplification at the scale we are building toward? The answer is not no. But it is not the easy yes that the exhilaration of the steep section suggests. It is a conditional yes — conditional on energy planning, on infrastructure investment, on supply chain diversification, on water management, on the unglamorous, year-over-year work of building and maintaining physical systems that change at the speed of concrete, not the speed of code.

Smil once said he is neither an optimist nor a pessimist — he is a scientist who follows numbers. I am not a scientist. I am a builder. But the builder who ignores the foundation is not building. He is decorating. And Smil has shown me, with a rigor I cannot argue with, that the foundation of the AI revolution is heavier, more constrained, and more physically demanding than the tower I described in The Orange Pill acknowledges.

The tower still stands. But I now understand, better than I did before this volume, what holds it up.

— Edo Segal

Every token has a weight.
Every prompt draws power from a grid that is already breaking.
Before you celebrate the intelligence revolution, count what it costs.

** The AI discourse lives in the weightless -- tokens, parameters, prompts, the imagination-to-artifact ratio collapsing to zero. Vaclav Smil has spent fifty years insisting on a different discipline: before you tell a story about transformation, measure what the transformation physically requires. This volume applies Smil's rigorous quantitative framework to the AI revolution, exposing the gigawatts of electricity, the millions of gallons of cooling water, the fragile semiconductor chokepoints, and the years of construction time that separate software capability from sustainable deployment. From the Jevons Paradox of intelligence -- where productivity gains drive more consumption, not less -- to the S-curves that inevitably bend, these chapters ground the Orange Pill's optimism in the material reality the optimism depends upon. The river of intelligence flows through turbines, transformers, and thermodynamics. This is the ledger no one opened.

“** "I have never been wrong on these major energy and environmental issues because I have nothing to sell." -- Vaclav Smil”

— Vaclav Smil