The cheapest words ever written may be the most expensive to compute. That is the paradox at the core of the AI boom. The product feels weightless, but the bill is physical. It arrives as power draw, land, copper, cooling, water, and years of capex. Markets act as if inference is free and scale is inevitable. Physics does not.
There is a wedge between what users pay and what AI costs to deliver. That wedge is being bridged by venture funding, hyperscaler balance sheets, and public markets that will presently fund losses if the story has momentum. AI’s economics today ride an invisible subsidy made of cheap capital and cheap power. Bloomberg projects electricity demand from AI training and services to roughly quadruple within a decade. That is not a rounding error. It is a structural shock to grids built on decade-long planning cycles and slow permitting. Basic unit economics are being deferred by accounting tricks and growth-at-any-cost narratives. The spread between marginal inference cost and marginal revenue is being justified by tomorrow’s productivity miracle. That is a faith-based position, not a cash flow.
AI’s constraint is not algorithms. It is amperage. Large models consume serious electricity when trained and a nontrivial amount with every query. Industry insiders warn that routine AI usage for trivial tasks could push consumption sharply higher. Data centers are already one of the fastest-growing loads on power systems. In many regions, interconnection queues stretch years. Power prices reflect local generation mix, grid congestion, and weather. The math flips fast if electricity gets costlier or intermittent. Cooling requires water or additional energy for alternative systems. Transmission requires steel, transformers, and environmental approvals. None of that moves on startup timelines. When the unit bottleneck is megawatts, not MAUs, valuation fantasies meet transformer lead times.
The industry is behaving as if Moore’s Law will paper over the energy gap. The physics say otherwise. Power density per rack is rising, cooling is approaching practical limits, and memory bandwidth—not raw flops—now constrains performance. Gains in performance per watt are real, but they are not exponential. Jevons paradox also lurks: better efficiency increases total use as demand expands to fill it. If power does not get cheaper on a delivered basis, hardware refresh cycles will lengthen. The useful life of GPUs will stretch from two years to five or seven as boards are sweated for capex recovery. That is the opposite of current consensus, which assumes short cycles, rapid monetization, and permanent scarcity premia. Thermodynamics is an indifferent counterparty. It will not honor growth guidance.
Hyperscalers are trapped in a game-theory loop. If one cuts capex, another will claim the future and the stock premium that comes with it. So everyone runs. The result is a Red Queen race that converts equity enthusiasm into substations, generators, and chip backlogs. It is a classic prisoners dilemma: the cooperative outcome would be slower, disciplined spending. The dominant strategy is to defect and spend anyway. In a regime of rising funding costs, this arms race becomes fragile. Expected value logic breaks if model quality improvements yield diminishing returns while inference costs per user rise. The “growth” multiple the market awards is subsidized by externalities not on the income statement—grid strain, policy risk, and environmental costs that end up socialized.
The market assumes abundant content means abundant value. It does not. When content supply explodes, the marginal price trends to zero unless distribution monopolies or brand moats exist. AI has already flooded channels with low-grade outputs. If users must start paying closer to the real marginal cost of inference, a lot of engagement dies. Advertisers will not fund the world’s inference bill for junk impressions. Measurement noise rises as bots use bots to generate traffic, eroding ad yields and trust. Meanwhile, moderation, provenance, and legal costs increase. The paradox deepens: the cheaper it feels to generate content, the more expensive it becomes to filter and monetize it. In past manias, shovel sellers got rich. Here, even shovels require massive electricity and specialist labor to operate at scale. The bill keeps surfacing.
This script is not new. The railroad mania built too much track, then consolidated. The dot-com era overbuilt fiber and data centers, then value migrated to a few durable networks with pricing power. In each case, investors paid up for a growth curve that could not clear its cost of capital. Long, dull mean reversion followed the crash. Today’s AI cycle adds a twist: the binding constraint is not just capital, but energy. If power grows dearer while money grows scarcer, valuations compress violently. Even boosters admit the bust case is plausible, though they treat it as a tail. Others argue the boom is fed by central bank liquidity and fiscal push, making it an unsustainable speculative surge. Consensus concentrates in the upside scenario because it must; that is how arms races are justified. Probability, however, does not negotiate with optimism. The fat tail sits on the downside because physical capacity cannot scale as quickly as narratives.
Strip away the marketing. Three paths exist. In the first, financial discipline arrives. Low-value inference dries up as usage-based pricing meets real energy costs. Growth slows, infra remains core, multiples deflate. In the second, energy prices rise and capital tightens together. Training slows, inference prices lift, and many firms vanish. Hyperscalers retreat to core workloads, sometimes with policy help. In the third, AI more than pays its own bill by delivering durable productivity to enterprises. Valuations, in hindsight, prove fair. Investors love the third because it absolves them of hard trade-offs. Yet all three paths run through the same choke point: electrons. That is why the effective bet today is not just on models, but on the cost and availability of power over the next decade. History says plan on disappointment and budget for delays.
Markets are repeating a classic error: confusing subsidized usage for product-market fit. Free trials at population scale feel like destiny. They are not. They are a stress test of how much an ecosystem will consume when the meter is off. When pricing normalizes, behavior normalizes. That is when psychology flips from fear of missing out to fear of paying up. We saw it in streaming as content costs rose and password sharing ended. We saw it in ride-hailing when subsidies faded. The same curve applies here, only the fixed costs are higher and the assets heavier. Once investors accept that AI is a power business wrapped in software, they will demand power-business returns. That means longer paybacks, lumpy growth, and an end to stories untethered from watts and wires.
Systems become robust when they absorb volatility rather than depend on permanent subsidy. AI will be no different. Antifragility here looks like metered usage, co-location with cheap baseload energy, smarter scheduling of training to off-peak windows, and honest accounting for water and grid impacts. It looks like fewer vanity models and more narrow systems where the value per query exceeds the cost to compute it. It looks like acceptance that the cheapest content is not the best business, and that the scarcest input is power, not prompts. The only real question for investors is whether their thesis survives when electricity stops being background noise. If your model needs free watts to work, it is not a model. It is a hope.
