The artificial intelligence (AI) boom is often narrated as a story of technological inevitability. Recurrent unease in technology markets tells a more complicated story. Nvidia is not an unprofitable company sustained only by speculation: its latest quarterly revenue was US$ 81.6 billion, 85% higher than a year earlier. Yet, present profits do not dissolve the wager.The scale of investment and market valuation also depends upon expectations that AI will become much cheaper, more capable and more widely used. Extraordinary capital has already been committed to chips, data centres and electricity infrastructure on that basis. There is real technology and real revenue, but also a collective wager on how quickly its future will arrive.We have always speculated on new technologies. What is distinctive about the present moment is the degree to which financial valuation has become tied to improvements that must continue almost in real time. Chips must improve; models must become more efficient and reliable; data centres must secure electricity and grid connections; and firms must find applications for which people are willing to pay. Inference cost – the expense of running a trained model – is not the only variable, but it is among the most consequential: mass adoption requires computation to become cheaper even as its capabilities expand.The efficiencies on which this wager rests are already arriving. According to Stanford’s 2025 AI Index, the cost of querying a model performing at roughly the level of GPT-3.5 fell from US$ 20 per million tokens in November 2022 to US$ 0.07 by October 2024, a decline of more than 280 times. Hardware costs have also fallen while energy efficiency has improved. The conventional inference is that continuing gains will widen adoption and vindicate the investment. Ecological economics, however, asks a different question: what happens to total energy and material use when each unit of computation becomes cheaper?The history of industrial efficiency gives little reason for complacency. A more efficient technology does not necessarily reduce aggregate resource use. By lowering the cost of a service, it can encourage that service to be used across many more activities. This rebound effect, associated with the Jevons paradox, is especially important for AI. A cheaper token may not mean that society undertakes the same amount of computation with less electricity. It may mean that AI is inserted into every search, email, classroom, office, hospital, financial transaction and bureaucratic process. Energy use falls at the level of a particular operation, while the number and complexity of operations expand much faster.This is no longer a hypothetical concern. What about the energy demand?The International Energy Agency (IEA) estimates that electricity consumption by data centres rose to around 485 terawatt-hours in 2025 and could reach about 950 TWh by 2030. Electricity demand from AI-focused data centres may triple over the same period. The IEA also notes the apparent contradiction directly: power consumption per AI task is declining rapidly, yet wider adoption and more energy-intensive applications continue to push total electricity demand upwards.AI is, therefore, not an immaterial technology residing somewhere in the cloud. Beneath every prompt lies a physical system of semiconductor fabrication, mines, power plants, transmission lines, cooling systems and data centres. It requires land, water, minerals, specialised equipment and a continuous supply of electricity. Technological progress can shift particular constraints, but it cannot remove computation from this material metabolism. The relevant question is not whether further efficiencies are possible. It is whether they can produce absolute reductions in resource use while the scale of computation continues to expand.This is where the financial and ecological dimensions of the gamble meet. According to the IEA, capital expenditure by five of the largest technology companies exceeded US$ 400 billion in 2025, with a further sharp increase anticipated in 2026. Financial markets require AI to grow rapidly enough to justify this infrastructure. Yet, that requires an equally rapid enlargement of material systems. When transformers are unavailable, grids cannot supply connections, advanced chips face constraints, or communities resist the diversion of land and water, a technological problem becomes a material and political one. Ecological limits return to a financial calculation from which they had largely been excluded.The outcome is a double bind. If capability, efficiency, reliable applications and paying demand do not advance quickly enough, expectations may be revised and losses may spread well beyond a few start-ups. If they do advance, cheaper computation may validate the financial wager while intensifying the ecological one. Commercial success and ecological success are not the same thing.Human costEfficiency also tells us nothing about justice. Markets ask whether cheaper computation can generate profitable products; they do not automatically ask which applications meet a social need, whose resources sustain them, or how benefits and burdens are distributed. The burdens of expansion are dispersed, while ownership and decision-making remain concentrated. Communities may provide land, water and electricity; workers may annotate data and moderate content; public institutions may subsidise infrastructure. Yet, a small group of corporations retains disproportionate authority over which applications are developed and who benefits from them. Nor can a more efficient chip decide whether scarce electricity should support household cooling during a heatwave, a rural health centre, or another data centre producing automated advertising and synthetic entertainment. That remains a social and political choice, even when presented as the inevitable consequence of innovation.The question of work exposes the same problem. The International Labour Organization (ILO) estimates that one in four workers is in an occupation with some exposure to generative AI, and transformation is more likely than complete replacement. This is not necessarily reassuring. A job need not disappear for a worker to lose autonomy, income, security or bargaining power. For India’s engineers, accountants, writers and back-office professionals, occupations long associated with mobility and middle-class dignity may be reorganised into smaller tasks, more closely monitored and performed by fewer people. The productivity gains, meanwhile, can remain concentrated among those who own the models, chips, platforms and data infrastructure.India needs some domestic computational capability to avoid complete dependence on foreign platforms. But amid persistent inequalities in access to reliable electricity and water, this cannot become an unquestioned race to build the largest possible computer infrastructure. Public support should be selective and conditional upon demonstrable social value, transparent accounting of energy and water use, additional clean-energy provision, safeguards for affected communities, and protection for workers whose occupations are being transformed. We need the capacity to decide which uses of AI justify their material cost, not merely the capacity to produce more tokens.The danger, then, is not simply that the AI gamble may fail. Under the prevailing model, either outcome carries serious risks. If the promised returns do not materialise, societies that never meaningfully consented to the wager may nevertheless bear its losses. If the gamble succeeds commercially, ubiquitous computation may intensify ecological overshoot, labour insecurity and corporate concentration. The transition we need cannot depend only upon making computation cheap enough for markets to respond. It requires deciding which forms of intelligence and production deserve to expand within a world of hard ecological limits. Otherwise, we may make computation extraordinarily cheap while making energy, water, meaningful work and a liveable planet increasingly scarce.Soumyajit Bhar teaches ecological economics and sustainability at the School of Liberal Studies, BML Munjal University, and directs its Community Transitions Research Centre.