The Boom Before the Payoff
The subject of this series is the AI investment boom as an economic transition, not artificial intelligence in the abstract. The immediate question is why an economy can look stronger while becoming harder to interpret: investment, trade and asset prices are already responding to AI, but the productivity gains that would justify those expectations remain uneven and uncertain. By the end of this series, the reader should be able to separate what is happening now from what depends on future diffusion, and to judge claims about the boom by asking which channel is moving, who bears the adjustment and how strong the evidence is.
That distinction matters because AI is no longer confined to a research lab or a narrow group of software firms. The Bank for International Settlements describes a boom large enough to shape the global outlook in real time. The International Monetary Fund treats AI as a “macro-critical transition” that can affect productivity, employment, inequality, fiscal positions, monetary operations and financial stability. Those descriptions do not mean that a transformation is guaranteed. They mean that expectations about the technology have already become economic forces.
A boom that is already visible
The most visible part of the story is investment. Building an AI economy requires data centers, specialized chips, cloud infrastructure, energy supply, networking equipment and software. BIS estimates put spending on data centers and related information-technology manufacturing at around 0.8% of GDP in the United States and above 1% in Australia. Industry projections cited in the same bulletin put global AI-related investment at roughly $500 billion today and $3–4 trillion by 2030, while data-center capital expenditure is projected to triple. The latter figures are projections, not observations, but even the current build-out is large enough to affect national accounts and supply chains.
Investment raises demand before it raises the economy’s productive capacity. A new data center requires construction, power, equipment and financing now. The spending can lift business investment and the suppliers’ revenues even if the models hosted inside the facility later produce less value than expected. This is not a contradiction. Gross domestic product records the activity involved in building the asset; it does not certify that the asset will earn an adequate return over its life.
The financing structure makes the distinction more important. As AI firms’ capital expenditure has risen, their free cash flow has fallen. BIS reports a shift toward private credit and bond markets: private-credit loans to US AI-related firms grew from near zero in 2016 to $200 billion in 2025, and their share of total private-credit loans reached 8%. AI firms issued $243 billion in bonds in 2025, up from $79 billion in 2023. These numbers describe funding channels, not an impending crisis. But they show why an expectation about future productivity can affect today’s leverage, refinancing needs and exposure to a change in interest rates or collateral values.
The boom also travels through trade and wealth. Goods with high AI content have recorded some of the fastest export growth, while strong demand has pushed up their prices. Upstream exporters of semiconductors and related equipment can receive terms-of-trade gains; countries importing the machinery needed to build digital infrastructure can face higher import costs. Equity markets add another channel. The BIS reports that AI firms’ share of US market capitalization rose from 24% in 2022 to 40% in 2025. Higher valuations can support consumption through wealth effects, but those gains are concentrated among asset holders and economies integrated into the AI production chain.
These are demand-side effects: they are visible in spending, prices, financing, trade and valuations. They can support growth even when the supply-side payoff is still a question. The distinction is easiest to see in a simple map.
The upper path in this diagram is already observable. The lower path is conditional. A data center can be built before the organizations that use it have redesigned their processes, before workers have acquired complementary skills, and before regulators and customers trust the systems. The investment is real in the present; its full economic return belongs to the future and depends on what happens between capability and use.
Why the payoff is harder to see
AI can affect supply through at least three routes: automating tasks, improving performance within tasks and accelerating research and development. But the route from a better model to higher economy-wide output is not automatic. Some tasks can be performed virtually; others require physical presence, energy, materials, coordination or human responsibility. A model can be impressive at one task and unreliable at another that appears superficially similar. The relevant question is therefore not only how capable the frontier system is, but how much of the production process can actually be reorganized around it.
The IMF identifies physical and organizational bottlenecks that can limit aggregate gains even when technical progress is rapid. Energy and grid capacity, data-center infrastructure, non-automatable complementary tasks, institutional coordination and capital accumulation can all become binding. Regulation and compliance can slow adoption; excessive complexity can also favor large firms that can absorb the cost. Managerial risk aversion, concerns about reliability and public trust matter because adoption is a sequence of decisions by firms, workers, customers and governments. A lower price for using a model reduces one barrier. It does not remove the others.
This is why the sources distinguish frontier capability from economic diffusion. BIS cites early evidence of productivity gains and reports a median estimate of about 0.5 percentage points of annual total-factor-productivity growth across a wide range of macro estimates. The same discussion warns that micro-level gains may take time to appear in aggregate productivity as firms reorganize and resources move between them. The estimate is useful as a sense of scale, not as a promise. It does not tell us when gains will arrive, how widely they will spread or how much of them will be captured by workers rather than capital owners.
The IMF’s scenario exercise makes the logic especially clear. It holds rapid progress in frontier capability common across two paths and changes the speed and breadth of adoption. In the baseline-diffusion scenario, regulation, infrastructure constraints, skills gaps, social resistance and organizational frictions keep uptake uneven. In the runaway-diffusion scenario, adoption spreads rapidly, automation is broader and displacement arrives earlier. The exercise is not a probability forecast. Its value is that it prevents us from treating technical capability as the only variable that matters.
The same caution applies to the word “productivity.” If it means more output from a given combination of inputs, AI may eventually expand supply. But if the investment boom is financed with debt, draws power away from other uses, raises the price of scarce equipment or produces gains only in a concentrated group of firms, the transition can be difficult even while the long-run supply potential is positive. Aggregate output may rise while household demand weakens, workers lose bargaining power or public revenue falls. A single growth number cannot reveal those differences.
The map we need
The central issue is therefore not whether AI will matter. It is how to read an economy in which the demand impulse, the supply promise and the institutional response are arriving on different schedules. Four kinds of statement should be kept separate.
An observation describes something already measured: current investment, bond issuance, export prices, equity-market weights or employment changes in exposed sectors. An estimate translates incomplete data into a quantitative judgment, such as a productivity effect or an AI-preparedness relationship. A projection extends current assumptions into the future, as with industry forecasts for infrastructure spending. A scenario asks what would follow if a defined set of conditions held. Scenarios are valuable precisely because they are conditional; they should not be repeated as if they were forecasts.
This evidence discipline also clarifies why economic signals can point in opposite directions. Investment and wealth effects can raise demand and prices. Greater productivity can expand supply and restrain prices later. Fear of job displacement can reduce consumption, while a shortage of power or specialized equipment can raise costs in the near term. Strong growth might therefore reflect a temporary construction cycle, a durable increase in productive capacity, an asset-price feedback loop or several of these at once. The data are not meaningless. They are answering different questions.
The practical judgment supported by this series is a form of disciplined interpretation. When a new AI-related figure appears, ask first whether it measures capability, investment, adoption, output, distribution or expectation. Then ask what must be true for the observed movement to persist. If the answer depends on infrastructure, organizational change, labor reallocation or public trust, those conditions belong in the economic analysis rather than in a footnote.
The same approach keeps the human and institutional stakes visible. The IMF argues that states will need better measures of AI deployment, task-level labor effects, concentration, investment flows and financial exposures. Tax administration, social protection, supervision and data governance are not external to the transition; they help determine its path. A country with strong models but weak capacity to deliver adjustment may capture less of the gain and suffer more of the disruption. Conversely, a slower diffusion path may reduce immediate displacement while prolonging uncertainty and investment weakness.
The AI investment boom is thus both an observable event and an unresolved question. Its construction spending, financing, trade and market effects are already moving through the economy. Its productivity payoff depends on diffusion, complements, institutions and time. Reading the boom well means following both paths without confusing one for the other. The next step is to follow the physical build-out more closely: how expectations become capital expenditure, credit, supply-chain prices and wealth before the promised payoff has been fully measured.