Putting Down the Device: AI's Demand-Side Reckoning

Power and Concrete Are Solvable; Indifference Is Not

Two constraints dominate every conversation about artificial intelligence, and both are physical. The first is electricity: data centers now draw power faster than utilities can string new lines, and the grid has become the industry’s most honest bottleneck. The second is the build-out itself — the steel, the cooling, the silicon. By Bloomberg’s reckoning, the largest US technology firms may pour as much as seven hundred and twenty-five billion dollars into capital spending this year, the bulk of it AI data-center equipment. By some forecasts, free cash flow across the group turns negative for the first time in decades, the gap covered by record debt issuance.

These are enormous problems, but they are the kind money was invented to solve. A transformer shortage is a scheduling problem; a substation is a permitting problem; a chip is a manufacturing problem. Given enough capital and enough quarters, each one yields. There is a third constraint, though, that no balance sheet can underwrite directly, because it sits on no one’s balance sheet at all: whether the person holding the phone keeps wanting to open the app — and, more quietly, whether they will ever pay for the privilege. The grid can be expanded. Attention cannot be requisitioned.

The Whole Edifice Is Collateralized Against Future Attention

That word — attention — is the quiet collateral beneath the entire structure, and the spending only makes sense against a demand curve that has not fully arrived. The logic underneath is more circular than the headline figures suggest. Rising valuations justify heavier capital expenditure; heavier expenditure reads as a signal of explosive future demand; that signal feeds back into the valuations. The loop holds only until the revenue curve fails to steepen in time — at which point it breaks.

The arithmetic of the gap is stark. Even the largest AI revenue figures anyone cites — a few billion dollars a month, at most, for the category leader — remain a rounding error beside infrastructure commitments approaching three-quarters of a trillion a year. The bet is not on today’s usage. It is on a future in which hundreds of millions of people fold these tools so deeply, and so profitably, into daily life that the spending looks, in hindsight, conservative. The right question, then, is not whether people will open the app. It is whether their usage becomes the kind a fixed-cost empire can live on.

Breadth Is Not Depth: The Four Tests Demand Must Pass

By the raw count, the future looks secured. OpenAI says ChatGPT reached roughly nine hundred million weekly active users in early 2026, with the app reportedly crossing a billion monthly users not long after — a scale almost no consumer product in history has touched. But reach answers none of the questions that decide whether a build-out pays for itself. Four do: how often people come back, whether they stay, whether they pay, and how easily they leave.

On frequency, the texture thins fast — the average web session runs under thirteen minutes, and the typical weekly user sends fewer than three prompts a day: habitual, but light. On retention, enterprise adoption is broad and shallow at once, with most firms running AI somewhere in the business yet fewer than four in ten pushing it past a pilot. On switching cost, the picture is worse still: SimilarWeb-based reports put ChatGPT’s share of generative-AI web traffic down from roughly three-quarters to the high-fifties in about a year, with Gemini the main beneficiary, because the friction in moving from one chatbot to another is a single tap — and boycott calls after a contested defense deal only underscored how loosely the loyalty is held. A billion casual encounters can coexist with weak pricing power when the product is easy to substitute, bundled by incumbents, or used mostly for low-value tasks.

The Habit May Be Captured by Whoever Owns the Surface

That last point is the one the device metaphor obscures. Putting down the phone was always the easy version of the threat; the subtler version is that people never stop using AI at all — they simply stop using it as a thing they open, or pay for, on its own. Gemini’s traffic gains are not only evidence that rivals exist. They are evidence of distribution: Google already owns Search, Android, Chrome, Gmail, and Docs, and AI poured into surfaces a user visits anyway needs no separate habit and earns no separate fee. The risk is not that users reject AI. It is that the habit gets captured by whoever already owns the surface.

From there the economic failure mode is commoditization, not boredom. As models converge and incumbents fold AI into subscriptions people already buy, the standalone assistant slides from business to feature, and revenue per user erodes even as usage climbs. The optimist’s reply is the Jevons effect — make a resource cheaper and people consume far more of it. But Jevons helps only when falling cost unlocks latent, valuable demand; it guarantees nothing when the unlocked demand is low-value, price-sensitive, or captured by distributors. The more AI becomes expected, the harder it becomes to charge separately for it.

AI May Win the Interface and Still Lose the Economics

The largest industrial wager of the decade rests, finally, on a question of demand quality rather than demand volume. AI can plainly attract users; the open question is whether those users become durable, frequent, paying, switching-resistant demand — or whether AI becomes another ambient layer of the internet: everywhere, expected, expensive to provide, and difficult to price. Power and concrete are the visible walls of the maze; the real exit is not a dark phone but a billion people using AI all day while no one can bill them for it. The discipline this demands is unglamorous: measure frequency, retention, willingness to pay, and switching cost as ruthlessly as gigawatts and gross margin, because a build-out is only ever as sound as the economics of the habit it assumes will last.