In April, 2025, the International Monetary Fund still treated artificial intelligence as a medium-term possibility: a candidate for lifting weak potential growth, a source of electricity demand, and an item on the list of upside risks. By the spring and summer of 2026, the Fund was describing something else. AI had become a present-tense force in investment, trade, and financial markets, even as the broad productivity payoff remained unproven. The shift did not overturn the earlier analysis. It completed the first half of it and left the second half hanging over the forecast.
That change matters because the World Economic Outlook is not a technology essay. It is the Fund’s main statement of how the world economy is expected to grow, where the risks sit, and what governments should do. When AI moves from a special feature on data centers to a force that can offset a war-driven oil shock, the policy problem changes with it.
The April, 2025, Starting Point
Chapter 1 of the April, 2025, Outlook was written under a different headline risk. The global economy had begun to stabilize after a run of shocks. Inflation was falling toward targets, labor markets were normalizing, and growth had hovered near three percent. Then a wave of United States tariffs and retaliatory measures pushed trade-policy uncertainty to historic highs. The chapter’s task was to explain a cooling outlook and to argue for policies that would rebuild buffers and raise medium-term growth.
AI entered that argument as part of the structural-reform agenda. Durable reforms in labor markets, education, regulation, competition, and finance could lift productivity and potential output. Technological progress, including digitalization and AI, could do the same. The Fund added a distributional caveat that later reports would repeat. Gains would be shared more widely if countries built digital infrastructure, trained a digitally competent workforce, and kept regulatory frameworks robust. Public investment in infrastructure and digitalization was worth protecting even as fiscal space narrowed.
AI also appeared among the upside risks. Optimism about the technology, falling usage costs, and further technical progress could raise productivity and consumption. Knowledge could spill across industries and regions. Employment need not collapse if governments upgraded regulation and helped workers move. Electricity prices and environmental costs need not explode if countries expanded renewables and more efficient methods of production.
What the April, 2025, chapter did not do was put economy-wide AI productivity into the baseline forecast. The technology was a reason to hope that mediocre medium-term growth, then projected near 3.2 percent five years ahead and below the 2000-to-2019 average of 3.7 percent, might eventually improve. It was not yet a measured contributor to current G.D.P.
The Energy Special Feature that Defined the 2025 Discussion
The most detailed AI analysis in that chapter sat in the commodity special feature on energy demand. Generative models and large language models require data centers. Those facilities consume large amounts of electricity in both training and inference. The Fund cited estimates that worldwide electricity use by data centers and AI had reached 400 to 500 terawatt-hours in 2023, more than double the 2015 level. By 2030, AI-driven global electricity consumption could approach 1,500 terawatt-hours, a figure comparable to India’s total electricity use at the time and about one and a half times projected demand from electric vehicles. In the United States, data-center demand was projected to rise from 178 terawatt-hours in 2024 to 606 terawatt-hours in 2030 under a medium scenario.
The special feature also documented how quickly AI-producing sectors had already grown in the United States. Their value added quadrupled between 2010 and 2023. Their share of GDP rose from 2.4 percent in 2013 to 3.5 percent in 2023. Labor productivity in data processing grew about four times faster than in the rest of the economy, driven more by capital and intermediate inputs than by a sudden leap in economy-wide efficiency.
Using the IMF-ENV model, staff represented the AI shock as a rise in total factor productivity in information-technology sectors in China, the United States, and Europe, calibrated to expected data-center power demand. Under then-current energy policies, that shock raised average annual global GDP growth by 0.5 percentage points between 2025 and 2030, inside an earlier IMF range of 0.1 to 0.8 percentage points. The same shock increased cumulative global greenhouse-gas emissions by about 1.2 percent over the period, or 1.3 to 1.7 gigatons of carbon-dioxide equivalent. The social cost of those extra emissions, using a median estimate of thirty-nine dollars per ton, came to fifty-one billion to sixty-six billion dollars. That was only 1.3 to 1.7 percent of the extra real world GDP attributed to AI.
The conclusion was pragmatic. Higher electricity prices and emissions were real. They were not large enough, on the Fund’s arithmetic, to cancel the growth gains, especially if generation shifted toward renewables. In 2025, AI was therefore a growth technology with an energy bill attached.
What the 2026 Outlook Found in the Data
By April, 2026, the World Economic Outlook carried a darker title, “Global Economy in the Shadow of War,” after conflict in the Middle East closed the Strait of Hormuz and damaged production in a region central to hydrocarbon supply. The July, 2026, Update then framed the year as a contest between that negative supply shock and a positive technology shock. AI was no longer only a paragraph in the medium-term chapter. It was one of the two forces defining the near-term forecast.
The most important empirical claim was narrow and concrete. Technology investment related to AI added an estimated 0.5 percentage point to United States GDP growth in 2025. Because much of that capital equipment is imported, the impulse leaked abroad, especially to Asia. Korea’s growth surprised on the upside on semiconductor and AI-hardware exports. China’s high-tech manufacturing and exports also ran stronger than expected even as domestic consumption stayed soft. Brisk technology-related trade offset weaker flows in other product categories.
That is a different channel from the 2025 special feature. The April, 2025, model shock was a productivity increase in IT sectors that then pulled in electricity. The 2026 finding is a demand shock: firms spending on chips, servers, data centers, and power. Spending shows up in GDP before any user of a chatbot becomes more productive. The Fund was explicit that the 2025 growth contribution did not include general productivity from AI adoption by ordinary businesses.
Kristalina Georgieva later summarized the geography. What started as a United States phenomenon was becoming a global growth engine as other countries ramped up construction of data centers and related infrastructure. Economies plugged into the supply chain benefited. Commodity importers that were not well positioned in the technology value chain did not. The world economy, in the Fund’s telling, was splitting along an AI axis as well as an energy axis.
The Productivity Puzzle That Did Not Ho Away
The continuity between the two vintages is as important as the contrast. The April, 2026, reference forecast still did not include direct effects of AI on aggregate productivity. Adoption rates remained modest in many sectors. Staff noted that AI use correlated with faster productivity growth across sectors but explained little of the aggregate gain. It was too early, they wrote, for the benefits of adoption to show up materially in the data.
That judgment keeps 2026 closer to 2025 than market commentary often admits. The Fund did not declare that the productivity revolution had arrived. It declared that the investment boom had arrived and that the productivity revolution might follow. The July, 2026, Update made the same distinction. The AI-driven technology cycle was assumed to moderate, and the staff put no exogenous boost to productivity growth into the baseline. Global growth was projected at 3.0 percent in 2026 and 3.4 percent in 2027, with the war’s drag partly offset by technology demand rather than by a measured rise in economy-wide efficiency.
The upside arithmetic, if the second stage does arrive, looks familiar. Faster adoption and efficiency gains could lift global growth by as much as 0.3 percentage point in the near term and by 0.1 to 0.8 percentage point a year in the medium term, depending on the speed of adoption and on AI readiness. Low-income countries would need more than enthusiasm. They would need electricity, digital infrastructure, skills, and a shift of labor out of sectors where AI gains are limited. Those conditions echo the 2025 call for digital infrastructure and training. The numbers sit next to the 0.5-percentage-point annual effect in the 2025 energy model. The story is consistent. Only the calendar has moved.
From One-sided Upside to a Two-sided Risk
April, 2025, listed AI mainly as an opportunity. April, 2026, listed it as both an opportunity and a financial vulnerability. A revaluation of profit expectations, or of viable markups under more intense competition, could cut investment and trigger an abrupt market correction even if some productivity gains eventually appear. Equity markets had become more concentrated in AI stocks. Japan, Korea, Taiwan Province of China, and the United States were outperforming in part because of that exposure. The Global Financial Stability Report had already flagged the concentration. Chapter 1 now treated a disappointment on earnings and productivity as a macro risk, not only a sector story.
The Fund built that worry into scenarios. In one, deeper United States adoption triggers another surge in computing and infrastructure investment. United States GDP rises modestly relative to the reference forecast, investment jumps, and a large share of the impulse goes to imports. The divide between the United States and less exposed economies widens. In another, expected productivity gains are reappraised, real investment in the technology sector falls sharply, and the recent run-up unwinds. Those layers sit beside war, tariffs, and fiscal stress. AI is no longer a side bet. It is a scenario driver.
The Foreword to the April, 2026, report added a labor-market warning that 2025 had treated more gently. Recent developments, including agentic AI, raised the prospect of very meaningful productivity gains, “the ultimate driver of standards of living.” The transition could still be hard. Financial-market enthusiasm might have run ahead of fundamentals. A rapid transformation could make many jobs obsolete and slow aggregate demand. The 2025 text had said employment effects need not be large if policy helped reallocation. The 2026 text said the same help would be needed and that demand, not only supply, could suffer in the meantime.
Energy, Rare Earths, and a Different Constraint
Energy did not disappear from the 2026 discussion. It changed role. In 2025, the question was whether AI’s electricity use and emissions would eat the growth gains. In 2026, the questions were whether power and critical inputs would constrain the buildout, and whether the AI boom could offset an oil shock. The Commodity Special Feature in 2026 highlighted rare-earth elements in global supply chains as a point of friction. Directors and staff still argued that renewables and new production methods could let countries harvest AI without a spike in electricity prices. The emissions ledger that occupied several pages in 2025 was no longer the center of Chapter 1.
That shift follows the facts on the ground. Once firms are pouring capital into data centers, the binding constraint is less a social-cost-of-carbon calculation and more whether grids, generation, and minerals can keep up. The war in the Middle East made that constraint geopolitical as well as technical. Georgieva argued in August, 2026, that the world had weathered the Hormuz shock better than feared because of reserve drawdowns, non-Gulf supply, weaker energy demand, more renewables, and higher coal use, and because AI investment supported earnings and demand. The tug-of-war she described was the 2026 outlook in one sentence: a negative energy supply shock against a positive AI demand shock.
What the Two Reports Ask of Policy
Read together, the two vintages ask governments to do three things at once.
-
- First, treat the investment boom as real and uneven. Countries in the hardware and data-center chain are already feeling demand. Countries outside it are not. Policy that ignores that split will misread growth, trade balances, and political pressure.
- Second, do not confuse capital expenditure with productivity. The Fund has now watched a year of heavy spending and still refuses to put broad total-factor-productivity gains in the baseline. Complementary investments remain the condition for the second stage: power supply, skills, management, interoperable infrastructure, competition, data governance, and programs that help workers move. Without those, the boom can fade when the buildings are up.
- Third, prepare for a market correction if profits lag the story investors have bought. High valuations, concentrated indices, and hyperscaler capital plans that outrun free cash flow are not footnotes. They are how a technology shock becomes a financial shock. Credible macro frameworks, including fiscal buffers and central-bank independence, matter more when one sector carries a large share of sentiment.
The April, 2025, chapter closed by urging countries to harness generative AI responsibly. The 2026 outlook says the harnessing has begun as spending. Responsibility now includes keeping the lights on, spreading the gains beyond a handful of firms and countries, and not letting a disappointment in earnings become a global tightening of financial conditions.
A Consistent Story with a New Clock
The IMF did not flip from skeptic to believer, or the reverse, between 2025 and 2026. It moved AI from the future tense to the present tense on one margin and left it in the future tense on another. The present-tense margin is investment, imports, Asian electronics exports, and equity-market concentration. The future-tense margin is the 0.1-to-0.8-percentage-point medium-term growth range that still depends on adoption, readiness, and complementary policy.
That is why the comparison is useful. The 2025 report asked whether AI could raise potential growth without an unacceptable energy bill. The 2026 reports ask whether a visible capital-expenditure boom can carry the world through a war and a hangover of trade barriers long enough for productivity to show up, and what happens if it does not. The second question is harder. It is also the one the data have forced onto the page.
The Chicken and the Egg
This comparison of IMF reports outlines a classic chicken-and-egg problem that is beginning to resolve itself, though not evenly. Capital equipment and infrastructure have taken the lead. Usage of AI is increasing, but not at the same rate. Much of the disparity can be traced to questions that midsize firms keep asking: How do we fold AI into daily work so that the benefits actually arrive? And when will the fire hose of AI development slow enough that we are not spending all of our time trying to keep up instead of putting the tools into the workflow?
Those two questions are not mutually exclusive. There is not, at present, a one-size-fits-all system. The available options tend toward highly specialized models or general systems that still need to be trained. The fire hose of updates and wholly new systems is only increasing.
What Firms Might Usefully Do
If a firm has the staff, assign two people to write, with AI, Friday summaries of what has changed across the AI spectrum and in the systems the firm already uses. On the following Monday, distill that material into a concise note so that each person can judge what is, and is not, a useful upgrade or a new product that would raise productivity.
The bottom line is that 2026 is not the year for enterprises to standardize on a particular AI, because no such system exists yet. Treat the year as a continuation of the experimentation of the past several years, with best practices shared throughout the organization.
For more information about how our work in Applied Artificial Intelligence can help you, please contact us HERE.