AI Costs Plummet 280x: Will Global Growth Get an Extra 0.5% Boost?

Author Karmendra Pandey has spent a long career in IT budgeting. Drawing on frontline data, he makes a judgment far grander than "AI cuts costs": AI will push global GDP growth beyond the past two decades.

IT Cost Reduction Is Already a Fact, But That's Not the Whole Story

AI's reduction of IT costs is backed by solid empirical data. In customer support scenarios, agents using AI assistants resolved 13.8% more issues per hour, with the least experienced employees seeing the largest gains—about 35%. In a controlled writing experiment, professionals completed tasks 40% faster while quality improved by 18%. A field experiment on large development teams showed developers using AI coding assistants completed about 26% more tasks.

A more fundamental change is happening in AI inference costs. The cost to achieve a fixed level of AI performance fell from $20 per million tokens in late 2022 to about $0.07 roughly eighteen months later. A price drop of more than 280x means tasks that were once too costly to automate—documenting legacy systems, summarizing every incident, testing every line of old code—are quietly becoming the default option.

But the author stresses that IT cost reduction is only the beginning of the story. The real foreshadowing lies in the fact that when the marginal cost of intelligence approaches zero, the way the entire economy is organized will be forced to redesign itself.

Growth Forecasts Diverge Widely, But the Consensus Is More Solid Than It Seems

Forecasts for AI's pull on global GDP vary enormously across institutions. The IMF estimates AI will contribute about 0.5 percentage points to global growth annually over the next decade; the OECD believes AI could bring 0.4 to 1.3 percentage points of additional productivity growth per year in the most exposed advanced economies; Goldman Sachs estimates AI will raise global GDP levels by 7% over about a decade. Meanwhile, leading skeptic Daron Acemoglu calculates far more conservatively, estimating total US GDP gains of only 1% to 1.6% over ten years.

The author reads a key signal from this divergence: the serious debate is not about whether AI works at the task level, but about how quickly firms can reorganize around AI, how many tasks AI can profitably complete, and whether it creates new work or merely automates old work. These are management questions, not model questions.

He ultimately settles on the median forecast. An extra half percentage point of growth per year sounds modest, but advanced economies have been hovering in the 1% to 1.5% productivity growth range for two decades. Compounded, half a percentage point is enough to reverse most of the slowdown. We don't need heroic forecasts to come true—we only need the "boring" one to hold, and that alone would constitute a historic turning point.

The productivity J-curve is the key concept the author uses to explain "why current data doesn't show AI's effects." It borrows the metaphor of the "J-curve effect" from macroeconomics: an investment or shock initially worsens measurable indicators temporarily, before improvements that exceed trend emerge, tracing a path shaped like the letter J. In the AI context, firms must first invest in "intangible complements"—data governance, workflow redesign, employee training, compliance and security—costs that hit current expenses immediately but only translate into output once the organization has readapted. Therefore, a "stagnation" or even "regression" in national or industry-level productivity statistics during the early diffusion of a technology does not necessarily mean the technology is ineffective. Understanding the J-curve helps distinguish between two narratives: one is "AI has no productivity effect," the other is "AI's productivity effect has not yet crossed the threshold of complementary investment." The author clearly holds the latter view, using it to explain the huge gap between current macro data and micro experiments.

History Is an Honest Judge: Every Technological Turning Point Was Underestimated

The author uses industrial history to remind us to be patient. James Watt patented the steam engine in the 1770s, but steam's greatest contribution to British productivity growth came about a century later, in the 1850s to 1870s. Electricity demanded even more patience: power stations appeared in the 1880s, but by 1900, fewer than 5% of factory machinery was driven by electric motors. The productivity boom didn't arrive until the 1920s, roughly forty years later, and only after factories were redesigned—a single large steam engine plus a line shaft was replaced by one motor per machine and single-story production lines built around workflow rather than power source.

Computers replicated this pattern almost perfectly. Robert Solow wrote his famous line in 1987: "You can see the computer age everywhere but in the productivity statistics." The statistics didn't really move until 1995 to 2004, by which time firms had already redesigned processes around the technology.

The internet and smartphones offer the reverse lesson: adoption can happen almost instantly—smartphone penetration in the US went from 5% to 40% in about four years—yet the productivity boom can still be absent, because consumer delight is not the same as the restructuring of production organization.

This is precisely why the author isn't worried about the current stagnation in productivity data. AI is following what researchers call the "productivity J-curve": firms first invest in intangible complements such as data, workflow redesign, training and governance, investments that drag on measurable output before the benefits appear. Electricity needed new factories; AI needs new workflows. Technology is the easy part to buy; redesign is the real work.

The Solow Productivity Paradox is an important historical reference for understanding the AI growth debate. In 1987, economist Robert Solow remarked: "You can see the computer age everywhere but in the productivity statistics." At the time, computer investment was growing rapidly, but official productivity statistics showed no clear acceleration. The paradox wasn't broken until 1995–2004, which scholars generally attribute to firms finally redesigning processes, organizations and supply chains around computers, rather than merely digitizing old processes. The core lesson of the Solow paradox is that the productivity dividend of a general purpose technology (GPT) depends on the speed of complementary organizational change, not on the technology's performance itself. Today's debate over whether AI drives GDP is almost a replay of the Solow paradox—micro experiments show significant task gains, while macro statistics remain flat. The author cites this history to show that the current data vacuum is more likely a measurement and reorganization lag than evidence that AI is ineffective.

This Time Is Different: Four Real Differences

The author does not believe this is merely a "repackaged electricity revolution." He points to four essential differences.

Speed. ChatGPT reached 100 million users in about two months. Electricity took decades to reach half of all factories. AI builds on devices, networks and cloud platforms that already exist—it doesn't need to wait for the world to be wired.

Replication is nearly free. Steam engines had to be built, transported and installed one by one. Once an AI capability is created, it can be replicated to a million companies overnight at near-zero marginal cost. Intelligence has never behaved like software—now it does.

It automates thinking, not muscle. Steam and electricity amplified what hands could do; AI amplifies what minds can do—drafting, summarizing, coding, analyzing, supporting customers—including the slow, expensive service sectors that resisted every previous wave of automation.

It can help invent its own successor. AI systems have already predicted the structures of 200 million proteins, work recognized by a Nobel Prize. A technology that accelerates research itself is not just a tool within the economy—it is leverage on the rate at which the economy acquires new tools.

General Purpose Technology (GPT) is the economic concept that runs through the entire analytical framework. It refers to technologies that can penetrate almost all industries and continuously spawn complementary innovations—steam engines, electricity, computers and the internet are all considered classic GPTs. The economic effects of a GPT typically don't manifest as the invention itself, but as the secondary innovation dividend released after production organization is rebuilt around it, which explains why its productivity contribution often lags by decades. The author places AI within this lineage, while emphasizing four differences—diffusion speed, marginal replication cost, the object of automation (cognitive rather than physical tasks), and acceleration of its own R&D—effectively arguing that AI may be a GPT with a "compressed lag phase." This framework helps readers understand that debating the magnitude of AI's pull on GDP is essentially debating the speed of GPT diffusion and reorganization, not model capability itself.

An Honest Warning: Task Gains Leak

The author includes two warnings from inside IT organizations.

The first is that task gains leak. In a rigorous trial, experienced developers using AI tools on their own mature codebases were 19% slower, yet they believed they were 20% faster. AI is fastest on bounded, pattern-rich work; on deep, context-dense work it can be a tax. Across the economy, the time saved by the average worker amounts to only about 1.4% to 2.2% of total hours, because most work hours are not yet AI-assisted. Multiplying lab gains by all of GDP is marketing, not analysis.

The second warning concerns the IT industry itself. AI reduces the cost per unit of IT work far faster than it reduces total IT budgets, because demand expands to eat the savings—firms themselves estimate about 29% of cloud spending is wasted, AI creates new waste while eliminating old waste, and data center power consumption will roughly double by 2030. For IT services firms, the business model itself is being repriced: work sold by the labor hour is priced against an input that AI is cheapening. The future belongs to outcome-based pricing, redesigned workflows, governance and modernization—selling complements, not hours. The author sees this as a healthy correction, but we shouldn't pretend it's painless, especially for the junior roles on which the old pyramid was built.

My Prediction: The Winners Are Those Who Reorganize Work Around AI

The author offers his most direct judgment: within this decade, global growth will be faster than the sluggish decade before AI. Not because the models are magical, but because the cost of intelligence is collapsing, technology replicates for free, and redesigned work—though slow—has already begun in support desks, software teams and back offices.

The winners will not be the countries and companies that bought AI first. Electricity rewarded the factories redesigned around it, roughly forty years after power stations appeared. AI will reward organizations redesigned around it—and this time it won't take forty years.

The steam age belonged to those who built engines; the electrical age belonged to those who rebuilt factories. This age will belong to those who rebuild how work is organized—around a mind that can be replicated. This is a turning point at least as large as previous ones, and for the first time in history, it is arriving faster than institutions can prepare. That gap, not the technology, is the real risk. Closing it is the real opportunity.

One-line Summary

AI inference costs have plummeted 280x; the real growth dividend depends on how fast organizations reorganize around AI.

Paragraph Summary

Drawing on frontline IT budget data, the article notes that AI has already delivered 13.8% to 40% efficiency gains in tasks like customer support, writing and coding, with inference costs falling more than 280x in eighteen months. But the author stresses that IT cost reduction is only the prologue—the real growth story lies in the economic reorganization that follows as the marginal cost of intelligence approaches zero. Forecasts for AI's pull on global GDP range from Acemoglu's conservative estimate to the IMF's roughly 0.5 percentage points per year; the author argues that if the median forecast holds, compounding would be enough to reverse two decades of productivity slowdown in advanced economies. History shows steam, electricity and computers all endured decades-long lags and productivity J-curves, but AI differs fundamentally in speed, replication cost, automating cognitive tasks and accelerating its own research. The author also warns that task gains leak and that the IT services model of pricing by labor hours is being repriced, concluding that the winners will be organizations that reorganize how work is done around AI.