A fast-growing AI market can be both real and expensive. That distinction is worth holding onto when headlines turn infrastructure spending into a single question: is this a boom or a bubble? The more useful question is narrower: what would show that the money committed to compute, power, and data centres is earning an adequate return?
Editorial illustration from a Pexels image credited to panumas nikhomkhai. It does not depict any company, facility, financing arrangement, or return outcome discussed below.
There is evidence of genuine use. A Federal Reserve review of surveys through 2025 puts business AI adoption at roughly 18% in one Census Bureau series and work-related generative-AI use at roughly 41% in a separate worker survey. Those numbers are not interchangeable: they measure different populations and definitions. Neither tells us how intensively a firm uses AI, whether it has changed output, or whether a provider earns a profit on the service.
Keep four measurements separate
The first measurement is adoption. A company can deploy an AI tool in a support queue, development workflow, or document process. This shows that the technology has a place in work. It does not tell us whether the deployment is material, recurring, or profitable. Survey wording also matters. The Federal Reserve notes that the Census Bureau broadened its business question in late 2025, so a rise across that break should not be read as a like-for-like growth rate.
The second measurement is investment. The Federal Reserve's accessible data puts combined capital expenditure by Amazon, Alphabet, Meta, Microsoft, and Oracle at $131 billion in the fourth quarter of 2025 and $412 billion across that year. That is a useful scale marker, not an AI-only bill: the companies invest in broader cloud and technology operations too. A large number says the buildout is economically important; it does not identify the cash flows that will repay a particular data hall, accelerator fleet, or power contract.
The third measurement is financing. The Bank for International Settlements describes a shift toward debt, leases, capacity-offtake agreements, and dedicated vehicles that can sit partly outside a hyperscaler's main balance sheet. These arrangements can be sensible ways to match long-lived assets with long-term funding. They also make the relevant question broader than headline capex: who has committed to pay, for how long, on what guarantees, and what happens if expected demand arrives later or at a lower price?
The fourth measurement is return. This is where the story becomes difficult. A provider needs enough paid demand, utilisation, and margin to cover operating costs, asset depreciation, and the cost of capital. A user needs a repeatable benefit that survives the end of a trial, a promotional credit, or a bundled product launch. Strong adoption can coexist with weak unit economics; falling compute prices can expand access while compressing revenue per unit.
Ask for a chain of evidence, not one spectacular number
A durable assessment starts by linking these four measurements rather than treating any one as decisive. For an AI service, ask whether usage is recurring, whether customers are paying separately or receiving it as a bundle, and whether retention holds when pricing changes. For infrastructure, ask how much capacity is contracted, how much is actually used, how long the equipment is expected to remain competitive, and whether power and construction commitments are reversible.
The financing layer deserves equal attention. BIS notes that a special-purpose vehicle may own a facility while a hyperscaler takes a minority stake and commits to long-term leases or capacity purchases. The debt may sit with the vehicle, but economic exposure can still travel through lease payments, guarantees, private-credit lenders, and insurers. That is not proof of a failure. It is a reason to trace obligations rather than assume they disappear because they are reported in a different place.
Treat productivity as an outcome still to be demonstrated
A useful AI workflow can save time without immediately appearing in economy-wide productivity statistics. It can also shift work rather than eliminate it. The Federal Reserve's later buildout framework makes the sequencing clear: lower costs and improved capabilities can precede adoption and investment, while broad productivity and labour-market effects take longer to measure. This is a useful antidote to two equally weak shortcuts: declaring that every deployment already pays for itself, or dismissing all adoption because aggregate data has not transformed overnight.
For readers evaluating AI tools, the practical test is local. Define the task, baseline time or quality, cost of the model and review, error rate, and the point at which the workflow is used often enough to matter. For leaders evaluating infrastructure exposure, apply the same discipline at a larger scale: revenue, utilisation, renewal, depreciation, financing commitments, and downside scenarios should connect.
The evidence supports a substantial AI buildout and meaningful adoption. It does not support a universal verdict on valuations, a forecast of a crash, or an investment recommendation. Separating use, spending, financing, and return makes the discussion more demanding—but much more informative.
AI Tools Radar separates product facts, editorial judgment, and commercial placement. Updated facts retain their verification date.