The Age of AI Is Here. What Happens Next?

In June 1998, Paul Krugman — eleven years before his Nobel — wrote a short, deliberately provocative essay predicting that by 2005 the internet's effect on the economy would prove "no greater than the fax machine's" [10]. He wasn't being reckless; he was reasoning from the fax machine's own history, a genuinely useful device that never once showed up in a productivity chart. A year after he published that line, markets were pricing internet stocks as though the medium had no ceiling at all. History does not often hand you two opposite errors made twelve months apart about the same technology, but it did here, and both of them are worth sitting with before anyone says a confident word about AI.

This Convergence doesn't ask you to believe the Age of AI has arrived. It asks what would count as evidence either way, and then goes looking for it. Here is some of what turns up: by August 2024, less than two years after the first mass-market chatbot shipped, 39 percent of American adults under 65 were already using generative AI — a faster climb than the personal computer or the internet managed at the same age, and one that shows up across a genuinely wide range of occupations and tasks rather than clustering in one industry [1]. Globally, adoption reached roughly half the population inside three years [2]. The four largest cloud companies have guided to about $700 billion in combined capital spending for 2026 alone, most of it AI infrastructure, up from roughly $410 billion the year before [3]. And the International Energy Agency now expects data-center electricity use to roughly double by 2030 [9] — a projection large enough to move national forecasts.

None of that settles what kind of thing is happening. It only rules out the option of pretending nothing is.

Economists have a name for the small set of technologies that don't just sell a product but rearrange the economy around themselves: general-purpose technologies. Timothy Bresnahan and Manuel Trajtenberg coined the modern definition — pervasive across sectors, improving over time, and spawning further rounds of innovation — using the steam engine, electricity, and the semiconductor as their working examples [4]. Researchers are now applying that label to generative AI without much hedging, and the peer-reviewed adoption data is the reason why: the technology is genuinely in use across occupations and task types, the exact pattern the definition requires [1].

The label comes with two hundred years of fine print, though. Paul David's classic study of factory electrification found that it took roughly forty years for electric motors to show up in the productivity statistics — not because the technology was weak, but because factories, supply chains, and entire cities had to be rebuilt around it before the gains materialized [5]. Carlota Perez, looking across five technological revolutions, found a second regularity: each one runs on a rhythm of frenzied capital, a crash, and only then the long, unglamorous deployment phase that actually changes how people live [6].

AI carries a third piece of history nobody else's general-purpose technology can claim: it has already cried wolf twice. The field collapsed in the mid-1970s, when the Lighthill Report helped freeze British research funding after early promises went unmet, and again in the late 1980s, when the commercial expert-systems boom went bust [7]. Anyone arguing that this time is structurally different is taking on the burden of proof, not receiving the benefit of the doubt — which is exactly why the next section sticks to what has actually been measured.

Start with what several independent instruments agree on. The 39 percent adoption figure comes from a peer-reviewed national survey; the roughly 53 percent global adoption figure, and the finding that 70 percent of organizations now use generative AI in at least one business function, come from a separate research institute [1][2]. This is the concrete difference between the present moment and the failed springs of 1970 and 1985: those were funding booms confined to laboratories [7]. This one is unfolding across the population. The capital tells the same story from a different angle — US private AI investment reached $285.9 billion in 2025 [2], and the hyperscalers' disclosed 2026 capital guidance, roughly $700 billion combined, is the largest corporate capital program on record [3].

Then the number splits. Consumers appear to be capturing an estimated $172 billion a year in value from generative AI tools in the United States alone [2]. But the most-cited study of enterprise deployments — MIT's Project NANDA — reports that roughly 95 percent of corporate generative-AI pilots show no measurable effect on profit or loss Competing hypothesis [8]. That finding is contested; the study isn't peer-reviewed and its definition of "failure" has been challenged, and it's labeled accordingly here. But the underlying tension shows up in more than one source: value that's easy to find in an individual's daily use is hard to find in a company's books. A slow-arriving transformation and an overbuilt disappointment would both produce exactly this pattern at this early stage [5][6].

That tension is exactly where qualified researchers disagree. The first dispute is over speed: one camp reads the adoption curves as proof this general-purpose technology will transform the economy faster than its predecessors did — software doesn't require anyone to rebuild a factory. The other camp points straight back at David: the binding constraint was never getting the technology into people's hands, it was reorganizing work around it, and the enterprise ROI gap is early evidence that the old forty-year lag is still very much alive [5][8]. Competing hypothesis Both sides include serious economists; the data cannot yet adjudicate between them. The second dispute is over what $700 billion actually is — the railroad grid of the next economy, or the frenzy phase of Perez's cycle arriving on schedule [3][6]. History's most direct precedent suggests it can be both: the dot-com crash wiped out enormous amounts of capital while leaving behind the fiber-optic network the next twenty years of the internet ran on. Whether today's chips, which depreciate far faster than fiber or rail ever did, leave a comparable residue is one of the sharper open questions in this entire investigation. Which reading is actually happening is the open question the rest of this piece works through.

This isn't one story, it's three ledgers moving at once and being read as if they were one. The adoption ledger (population-scale usage) [1][2], the capital ledger (the largest corporate spending program on record) [3], and the physical ledger (electricity demand large enough to move national forecasts) [9] don't have to agree with each other, and right now they don't: usage is broad, capital is concentrated, and the enterprise P&L is quiet [8]. That mismatch is also a labor-market signal — a companion investigation on this site tracks what happens when AI-attributed job cuts run ahead of the same kind of measured productivity gap documented here (Are AI Layoffs a Strategic Mistake?), and a general-purpose technology's capital and energy buildout doesn't resolve independently of how the workforce underneath it gets reorganized.

None of what follows is a forecast. It's what the historical parallels above imply is possible, laid out so the next few years of data can actually confirm or rule one out.

Put the measured pieces side by side — the fastest consumer-technology adoption on record [1][2], the largest corporate capital program on record [3], a physical buildout large enough to bend national electricity forecasts [9] — and researchers are using the term general-purpose technology without much hesitation [1][4]. What isn't established is what it becomes, and both of Krugman's misses from 1998 argue for humility about claiming otherwise in either direction [10]. These are the public, trackable indicators worth watching.

• Corporate capital guidance — Whether hyperscalers revise the ~$700B 2026 capex guidance up, down, or not at all [3].
• Adoption & productivity statistics — Whether enterprise adoption in Census Bureau business surveys keeps climbing or plateaus; whether national productivity statistics show any AI signal at all, or repeat the electrification-era silence [5]; whether independent researchers can replicate the 95-percent enterprise-ROI finding using better methodology [8].
• Consumer-vs-enterprise value gap — Whether the consumer-value and firm-level P&L estimates start converging or keep diverging [2][8] — arguably the cleanest single test between the electrification path and the bounded-tool path.
• Physical infrastructure — Measured data-center consumption against the IEA's trajectory [9]. By 2031, either the aggregate productivity numbers have moved, or the electrification-era lag [5] is repeating on schedule — either result answers this investigation's question.

Krugman changed his mind in public, years later, about the fax-machine line. So here's the question this whole investigation rests on: what single piece of evidence would change your mind about AI — in either direction? Write to us and tell us what it would take. We answer what we can, and the best replies shape what we investigate next.