The world has never spent this much on a technology this fast — $784 billion in 2025 alone, climbing toward $2.3 trillion annually by 2027. The productivity payoff, however, remains stubbornly out of sight. Despite the largest peacetime corporate technology investment surge in recorded economic history, aggregate productivity growth across advanced economies has barely moved. Now economists are asking a question that cuts to the heart of how modern economies absorb transformative technologies: is this a timing lag, or is something more structural going wrong?
The numbers from Stanford’s Human-Centered AI Institute 2026 AI Index Report are staggering by any historical benchmark. Global corporate investment in artificial intelligence reached an estimated $784 billion in 2025 and is on track to hit $2.3 trillion annually by 2027. To put that in context: $2.3 trillion exceeds the entire GDP of Italy. It surpasses the combined military budgets of the United States, China, and the European Union. It is being deployed faster than any comparable general purpose technology in recorded history — a pace that makes even the late-1990s internet boom look restrained by comparison.
The OECD’s 2026 Economic Outlook flagged AI capital expenditure as the single largest driver of business investment growth across G7 economies, accounting for roughly 40 percent of all increases in corporate capital spending in the United States and 35 percent in the United Kingdom. In South Korea, Taiwan, and Japan — the semiconductor corridor that feeds the AI hardware supply chain — AI-adjacent investment now represents more than half of all business investment. The International Monetary Fund estimates that AI-related investment will contribute 0.9 percentage points to global GDP growth annually through 2030, adding a cumulative $7.2 trillion to the world economy over that period. On paper, this is the largest productivity-enhancing technology deployment in human history.
Yet the productivity numbers tell a different story. The OECD’s Compendium of Productivity Indicators for 2025 showed multi-factor productivity growth in G7 economies averaging just 0.8 percent annually — the same sluggish pace recorded in 2019, before a single major AI deployment had occurred at scale. The United States, despite leading the world in AI investment, recorded labour productivity growth of just 1.2 percent in 2025, well below the 2.5 percent rate that many economists had projected as the AI dividend began flowing through the economy. McKinsey’s Global Institute, in its April 2026 report on AI value creation, gave this phenomenon a name: “the implementation gap.” The finding is stark — while AI model capability has advanced at a compound annual growth rate of 60 percent since 2022, the integration of those capabilities into actual business workflows has advanced at just 12 percent annually. For every five units of AI capability created, only one unit is currently being effectively deployed in a way that displaces labour, reduces input costs, or expands output.
Several interlocking mechanisms explain the lag. The first is the capital reallocation problem. When a company invests $500 million in AI infrastructure — data centres, specialised chips, model licensing, workforce retraining — that expenditure shows up immediately in investment statistics. It does not show up as productivity for three to seven years, because productivity gains flow only after the technology has been integrated into core business processes, workers have been retrained, and the organisation has rebuilt its workflows around the new capability. The capital expenditure is front-loaded; the productivity return is back-loaded. This is not unique to AI — the same dynamic appeared with electricity, which took 40 years to show up in aggregate productivity statistics because factories had to be completely redesigned around electric motors rather than mechanically driven shaft-and-belt systems.
The second mechanism is what economists call skill-biased technological change. AI raises the productivity of workers who can use it effectively while having minimal impact on, or even reducing, the productivity of workers who cannot. Stanford’s 2026 AI Index found that while AI tool usage among knowledge workers at AI-leading firms had reached 67 percent, the cross-economy average was just 23 percent. In most industries and most firms, AI is still a future productivity promise rather than a present reality. The third mechanism is competitive signalling. Firms invest in AI not necessarily because they have a clear path to productivity improvement, but because their competitors are investing, because boards demand an AI strategy, and because the cost of not investing appears higher than the cost of investing speculatively. McKinsey estimates that 25 to 30 percent of current AI investment by large corporations falls into this defensive rather than productive category — investment that inflates capital expenditure figures without a commensurate near-term productivity contribution.
The historical precedents suggest the AI productivity dividend should begin appearing in the 2027 to 2029 data. Electricity took 40 years to show up in national income accounts. The internet took 15 years from commercial adoption in the early 1990s — the famous Solow computer paradox dissolved only in the mid-2000s. The mobile internet took seven to ten years from the iPhone’s 2007 launch to produce measurable productivity effects in service-sector industries. For AI, the OECD’s central projection — based on sector-level adoption curves and historical GPT diffusion patterns — points to 2027 to 2029 as the inflection point.
Several conditions for that acceleration are already in place. Enterprise software vendors are embedding AI capabilities directly into existing workflow tools, substantially reducing integration friction. Workforce AI training programmes have scaled rapidly — the OECD estimates that 43 percent of large corporations now have formal AI reskilling programmes, up from 12 percent in 2023. The capital stock of AI infrastructure is reaching the threshold where marginal additions generate exponentially more capability per dollar invested. And AI model costs are falling: the cost per computation for frontier models has declined by roughly 40 percent annually since 2022, meaning the same dollar buys substantially more productive capability today than it did three years ago.
Yet there is a critical policy dimension that optimistic productivity projections tend to underweight. Historical productivity dividends from general purpose technologies have been distributed highly unevenly. The electrification dividend accrued primarily to capital owners and highly skilled workers in its early decades; wage gains for lower-skilled workers only materialised after unionisation, labour market regulation, and progressive taxation compressed the distribution. The same dynamic is already observable with AI. Stanford’s AI Index shows that 78 percent of the productivity gains from AI have so far accrued to firms and workers in the top income quintile. Whether the eventual aggregate productivity dividend is broadly shared or concentrated at the top will determine whether the AI investment story is ultimately a productivity story — or a productivity-and-inequality story running in parallel.
The $2.3 trillion question — when the productivity numbers arrive and who captures them — is therefore not just an economic measurement question. It is a question about the distribution of economic power in the decade ahead. The investment is real. The productivity lag is real. And the lag is not simply a timing problem waiting to resolve itself automatically. It is a structural challenge that will require deliberate policy choices about education, labour market institutions, competition regulation, and the terms on which AI firms are permitted to capture the value their technology creates. History suggests that technology alone does not guarantee broadly shared prosperity. The infrastructure of institutions around the technology matters at least as much as the technology itself.
Written by Nathan Brooks, Economy Correspondent
