The argument over artificial-intelligence spending is often framed as a contest between two forecasts: a productivity revolution or another investment bubble. Current data supports neither conclusion cleanly. It supports a narrower, more useful finding: AI infrastructure has become large enough to affect the composition of U.S. investment and some quarterly growth estimates, while the economic return on that capacity remains unsettled.
Business Insider's report highlights Torsten Slok's comparison between the rise of AI capital expenditure and the housing buildout of the 2000s. The comparison is valuable because it asks how rapidly one investment category can reshape an economy. It becomes misleading if it is treated as proof that the financing, balance-sheet exposure and failure mechanism must be the same.
For investors, the central question is not whether the spending line is steep. It is whether installed computing capacity becomes productive, billable infrastructure before depreciation, power costs and financing commitments absorb the expected return.
A boom can be real before its return is known
Federal Reserve researchers have assembled a public-data framework for tracking the AI buildout. It combines software, data centres, power facilities and computer equipment, rather than pretending that national accounts contain a dedicated “AI” line. Their conclusion is appropriately qualified: the selected components contributed meaningfully to growth from 2025 through the first quarter of 2026, but the estimate relies on assumptions about what counts as AI-related activity.
That distinction matters. Spending is a confirmed flow. The classification of that spending and the cash flows it may later produce are analytical judgments. A server purchased today adds to investment; it does not prove that customers will buy enough inference, training or software services to cover the server's full economic cost.
The housing analogy also has limits. The 2000s cycle transmitted through household mortgages, securitisation and leveraged financial intermediaries. Much of today's buildout is being led by large technology companies and infrastructure providers, although debt markets and off-balance-sheet structures can still spread risk. Similar investment curves do not create identical creditor chains.
Imports break the spending-to-GDP shortcut
Gross capital expenditure is not the same as domestic value added. The Bureau of Economic Analysis measures investment within a system that also subtracts imports. If a U.S. company buys equipment produced abroad, the purchase can raise gross private investment while the import adjustment offsets part of the effect on GDP.
The Fed note identifies this as a central measurement problem. Computing equipment has substantial import content, and national accounts do not neatly label which imported components support investment rather than consumption. Any estimate that counts the gross purchase without treating net exports carefully can exaggerate the domestic contribution.
This does not make the buildout economically irrelevant. Data-centre construction, electrical work, software, engineering and operation can create domestic activity around imported hardware. It does mean that a headline capex total cannot be translated directly into GDP points or domestic income. Investors should treat precise economy-wide contribution estimates as ranges built on definitions, not as a single observable fact.
Capacity is arriving ahead of broad use
Infrastructure is concentrated because the firms able to finance frontier computing are concentrated. Adoption is broader, but far from uniform. Census Bureau analysis of its Business Trends and Outlook Survey found that businesses with at least 20 employees were the largest users in its recent survey window. The Fed researchers likewise describe adoption as rising but uneven, with reported use failing to reveal how intensively the tools are used.
That creates a timing gap. Suppliers can recognise revenue as chips, networking equipment and construction are delivered. The buyers must then fill the capacity with paying workloads. Their customers, in turn, must find applications that save labour, improve products or create revenue. Each stage can grow at a different speed.
Micro-level studies may show that particular workers perform tasks faster with AI. The Fed note cautions that those gains have not yet appeared as a clear break in aggregate productivity across the economy. Both statements can be true: useful tools exist, but diffusion, organisational change and measurement take time.
Utilization will settle the capital-cycle argument
The bullish counterargument is credible. Networks are often built ahead of demand, and early excess capacity can be the condition that allows cheaper, more reliable services to spread. Large firms adopting first may simply reflect where the complementary data, skills and budgets already exist.
The skeptical case is also credible. Capacity can be technically impressive and still earn inadequate returns if workloads are less valuable than expected, prices fall faster than unit costs, or new hardware shortens the useful life of existing assets. Associated Press reporting also shows that the buildout is already affecting electricity and equipment markets, so the cost is not confined to technology-company income statements.
Four types of evidence would change the analysis. First, sustained utilization disclosures would show that installed capacity is being used rather than merely completed. Second, AI-linked revenue and cash flow would need to grow without depending indefinitely on internal transfers among a small group of infrastructure partners. Third, adoption would have to deepen among smaller and non-technology firms, not just broaden as occasional experimentation. Fourth, depreciation assumptions would need to match actual replacement cycles.
Until those signals develop, the most defensible conclusion is deliberately asymmetric. AI capex is already a macroeconomic fact. Its future productivity and investor payoff remain scenarios. The spending boom deserves to be measured without being mistaken for its own return.