Conceptual mechanical clock representing the changing pace of AI technology and economic cycles
MAi Edge Capital · PerspectiveOctober 9, 2026 · Research hypothesis

What If AI Is Changing Its Own Economic Clock?

Moore's Law, mainstream adoption, and Jevons paradox raise a question about the timing of the AI boom that historical economic cycles alone may not answer.

A forecast can identify the right danger and still get the calendar wrong.

In August, Arthur Hayes argued that the growth rate of announced AI infrastructure spending would begin to slow in mid-to-late 2027, becoming unmistakable by 2028. His warning deserves attention. Borrowed money can finance assets that age faster than expected, and technological enthusiasm does not repeal the requirement to earn an economic return.

But his timetable raises a question that deserves at least as much scrutiny as his warning: What happens when the technology underlying an economic cycle is changing fast enough to alter both the supply of a product and the demand for it?

This is not an argument that AI cannot be overbuilt. Nor is it a prediction that Hayes will be wrong. It is the beginning of an investigation into whether the assumptions used to date the next turning point are keeping pace with the technology itself.

The computer industry no longer moves to one clock

Gordon Moore's famous observation became the industry's shorthand for sustained improvements in chip density and cost. Intel still defends the trajectory, including through new manufacturing processes, packaging, and architecture. Yet the price of performing useful AI work is no longer determined by transistor counts alone.

Specialized accelerators, high-bandwidth memory, networking, model design, software optimization, and increasingly sophisticated data-center systems now contribute to what a unit of electricity and a dollar of spending can accomplish. Some advances make existing equipment more productive. Others make yesterday's hardware less competitive. Both effects can arrive within the same investment cycle.

This matters economically. A historical analogy assumes that certain relationships—between capacity, price, demand, and financing—remain reasonably stable long enough to forecast a turn. In AI, several of those relationships may be shifting at once.

Stanford's 2025 AI Index illustrates the pace. The price of querying models at a particular GPT-3.5-level benchmark fell from roughly $20 per million tokens in November 2022 to $0.07 in October 2024. That is a benchmark-specific comparison, not a universal price for intelligence. But the direction is unmistakable: capabilities that were expensive can become dramatically more accessible.

The more difficult question is what people—and machines—do when they become accessible.

Crossing the chasm changes more than the number of users

In our August essay, AI Has Crossed the Chasm. The Infrastructure Buildout Is Just Beginning, MAi Edge Capital argued that generative AI appeared to be moving from enthusiastic early adopters toward a broader mainstream market. That was an adoption hypothesis, not a proven universal threshold. We projected a substantial expansion in accelerated computing capacity if both adoption and usage intensity kept rising; that projection remains open to falsification.

One observation from that article matters particularly here: eventually, humans may stop being the correct denominator for AI compute demand.

A human is counted once as an adopter. But a business process might call thousands of models; a programmer might direct several coding agents; a laboratory might run autonomous cycles of simulation, criticism, and revision. Increasingly useful systems could generate workloads that do not map neatly onto the number of people who have tried a chatbot.

There is a necessary caution. Broad adoption is not the same as intensive, paid usage. Stanford's 2026 AI Index reports AI use at 88% of surveyed organizations, while agent deployment remains in single digits across nearly all business functions. That gap is consequential. It suggests enormous potential for deeper integration, but does not prove that the potential will turn into profitable demand.

Crossing the chasm may enlarge the market. It does not tell us how many compute-intensive tasks that market will ultimately purchase.

The paradox that could change the arithmetic

Jevons paradox begins with a counterintuitive possibility: when a resource becomes more efficient to use, its effective price falls, and the resulting increase in consumption can outweigh the savings per unit. Applied to AI, cheaper inference may make applications viable that were previously too expensive or too cumbersome to operate.

But a rebound is not guaranteed. Some efficiency gains simply reduce spending. Other gains expand use without expanding electricity consumption. The relevant question is not whether AI becomes more efficient; almost everyone expects that. It is whether economically valuable demand grows faster or slower than the computing resources needed for each unit of work.

The early evidence is suggestive rather than decisive. The International Energy Agency's 2026 assessment reports rapidly declining energy consumption for individual AI tasks while electricity use at AI-focused data centers increased about 50% in 2025. That contrast is consistent with growing adoption and increasingly intensive workloads. It is not, by itself, proof of Jevons paradox: the mix of tasks, facility construction, and hardware utilization are changing too.

What it does establish is that more efficient AI and greater total resource consumption can coexist. The economics cannot be inferred from either curve in isolation.

Hayes sees the paradox. Does his timetable fully account for it?

An honest challenge must begin by stating Hayes's argument correctly. In his August 4 essay, Situationship, Hayes explicitly discusses Jevons paradox. He accepts that AI-token consumption can soar even while the physical construction of data centers becomes excessive. His point is that demand might not expand quickly enough to protect older, debt-financed assets as chips improve and new capacity arrives.

That is a serious possibility. It is also possible for AI to flourish while infrastructure investors suffer large losses. More demand for intelligence does not guarantee higher rent for an older GPU cluster, or a return on a facility whose financing assumed more favorable economics.

The unresolved issue is the relative speed of the forces Hayes invokes. How fast will useful AI demand expand? How quickly will efficiency reduce compute requirements for existing tasks? How much new work will autonomous systems create? How long will physical capacity take to become productive? How quickly will investment costs, prices, and financing terms adjust?

Hayes's published essay specifies a mid-to-late-2027 turn in announced spending growth. We have not found in that essay a calibrated model demonstrating why those changing forces converge on that date. His forecast may prove correct; the stated calendar is not an independently established law of this market.

There are at least three coherent outcomes. Demand creation might outrun efficiency gains, sustaining rapid growth in utilized compute. Efficiency and overbuilding might outrun profitable new uses, causing capacity and pricing to disappoint. Or both could happen at once: aggregate AI consumption rises while particular hardware vintages, facilities, and capital structures lose value.

The third outcome is why neither a bullish adoption curve nor a bearish credit analogy, on its own, resolves the question.

The question we are putting on the record

MAi Edge Capital's working hypothesis is that the interaction of technological improvement, mainstream adoption, and induced AI usage may materially alter the timing and distribution of an infrastructure correction relative to familiar historical cycles. The direction of that alteration is not predetermined. Faster innovation might defer a demand shortfall—or bring forward the obsolescence of expensive assets.

Our planned independent investigation will test this rather than argue for it. We will compare like-for-like changes in AI task cost and computing intensity with actual workload growth; separate human adoption from agent-generated activity; examine paid usage rather than token totals alone; and track whether delivered capacity earns adequate returns after power, construction, financing, and replacement costs. We will revisit the forecast in our own Crossing the Chasm essay under the same standards applied to Hayes.

We will also identify what would prove the hypothesis wrong: evidence that the changing technology and adoption dynamics do not materially affect the timing or mechanism of investment cycles, or that they reinforce Hayes's proposed sequence rather than complicate it. The eventual research report will distinguish measurements from inference, preserve serious counterevidence, and explain any change in the conclusion. No completion date has been set.

Research status: pending independent validation

MAi Conclusion — Research Pending

Moore's Law is not a law of finance. Crossing the chasm does not guarantee an expanding market. Jevons paradox does not prevent infrastructure losses. These forces interact, and their relative speeds have not yet been independently established.

The final MAi Conclusion remains undetermined until the independent MAi meta-analysis has been completed, reconciled, and validated. Arthur Hayes's proposed 2027 inflection remains a forecast to test, not a conclusion we have rejected or accepted.

Open the forthcoming MAi research investigation →

Sources & research boundary

This editorial draws on Arthur Hayes's Situationship (August 4, 2026), Intel's explanation of Moore's Law (April 2025), the Stanford AI Index (2025) and Stanford AI Index (2026), and the IEA's Key Questions on Energy and AI (April 2026). It also revisits MAi Edge Capital's August 2026 Crossing the Chasm editorial, supplied from the publisher's archive.

Editorial status: A published hypothesis would not constitute an independent research finding, a validated economic forecast, or individualized investment advice. The planned investigation may confirm, revise, or reject it.