From Shubhi K | Product & Market Analysis
Hundreds of Billions Spent and Still No Visible Effect on GDP
On this page
Hundreds of billions of dollars have been deployed into AI since 2022. Multiple economic analyses report no measurable positive impact on US GDP growth from that investment through 2025. That gap is either a measurement problem that resolves with time, or evidence that the return is not arriving. Distinguishing between the two is the most consequential open question in this cycle.
Key takeaways
- The macro signal is currently absent. Economic analyses report no measurable positive AI contribution to US GDP growth through 2025, despite enormous deployed capital.
- This has a precedent and a name. The productivity paradox described the same pattern during the computing build-out of the 1970s and 1980s.
- The lag then was roughly a decade. Measurable productivity gains from computing appeared in the mid 1990s, long after the investment.
- Firm-level evidence points the same way as the macro data. MIT reported around 95% of enterprise generative AI pilots showing no measurable profit and loss impact.
What the data actually says
The claim is narrow and it is worth stating precisely. Multiple economic analyses have reported that AI investment through 2025 has not produced a measurable positive contribution to US GDP growth.
That is not a claim that AI does nothing. It is a claim about what appears in national accounts, which measure a specific set of things in a specific way and lag reality by design.
It is also not a claim that the investment is invisible. Construction of data centres, purchase of hardware and hiring all register as economic activity. What is missing is evidence of the productivity improvement that would justify the investment.
The productivity paradox
This pattern has a name because it happened before. The productivity paradox described the observation, prominent through the 1970s and 1980s, that computers were visible everywhere except in productivity statistics.
Enormous sums went into computing throughout that period with no corresponding measured gain. The explanations offered at the time were the same ones offered now: measurement failure, adjustment costs, the need for complementary organisational change, and the possibility that the technology simply was not as useful as claimed.
Measurable gains eventually appeared in the mid 1990s. The lag from substantial investment to visible productivity was roughly a decade.
The case for measurement lag
Four arguments support the view that the return exists and has not yet appeared in the data.
National accounts measure output and are poor at capturing quality improvements. A report that is better rather than faster to produce registers as nothing. A great deal of what AI currently does is quality improvement.
Complementary investment takes years. The computing gains of the 1990s required process redesign, not just hardware. Organisations are only now beginning the workflow changes that would let AI show up as output.
The deployment is also recent. Meaningful enterprise adoption began in earnest during 2023 and 2024. Expecting macro-level evidence within two years would be unprecedented for any general-purpose technology.
And capital deployment precedes return by construction. Data centres under construction consume resources and produce nothing until they are running, so the early years of any build-out look like pure cost in the statistics.
The case for genuine absence
The opposing view is not merely contrarian and it has firm-level support.
MIT's Project NANDA reported that around 95% of enterprise generative AI pilots showed no measurable profit and loss impact. If firms cannot see a return in their own accounts, expecting one in national accounts is optimistic.
The absence is also consistent across measurement approaches. Separate work found 42% of companies abandoning most AI initiatives in 2025, up from 17% a year earlier. Abandonment is a firm-level judgement that the return was not arriving, made by people with access to their own data.
And the substitution question is genuinely open. If AI accelerates tasks that were never the bottleneck, output does not change. Faster drafting of a document that then waits three weeks for approval produces no measurable gain, and a great deal of current deployment has that shape.
What would count as evidence either way
Both sides of this argument tend to reach for the same data and read it differently, which is a sign the debate needs sharper tests rather than more citations.
A measurement lag predicts that firms report gains before statistics capture them, that the gains concentrate in specific measurable functions, and that abandonment rates fall as organisations learn to deploy. All three are observable within a couple of years.
A genuine absence predicts the opposite: abandonment rates staying elevated, returns remaining unmeasurable at firm level, and the investment case shifting away from productivity toward competitive necessity. That shift in framing is the softest of the three signals and it is already partly visible.
What firm-level data adds
The macro and micro evidence currently point the same way, which is the strongest reason to take the absence seriously rather than dismissing it as a lag.
The consistency matters. A measurement lag would typically show firms reporting gains that national statistics had not yet captured. The current position is the opposite: firms are not seeing it either, which is examined in detail in the analysis of where AI returns have actually been measured.
The bottleneck problem
The most underrated explanation for the absence is neither lag nor failure. It is that AI has been applied to tasks that were never the constraint.
A process improves only at its slowest step. Accelerating document drafting in a workflow gated by legal review changes nothing measurable, because the queue simply forms at a different point. A great deal of current deployment has exactly this shape.
That explanation predicts something specific and testable. Returns should appear first in workflows where the automated step genuinely was the bottleneck, which tends to mean high-volume, repetitive processes rather than knowledge work with human approval gates.
Where a real effect would show up first
If the gains are coming, they will not arrive evenly. Some sectors have the conditions for measurable improvement and most do not.
Three properties predict where it lands. High volumes of repeatable work, output that can be counted without argument, and a process where the automated step was genuinely the constraint rather than an already-fast part of the chain.
Customer support, claims processing, document review and parts of software delivery meet all three. Professional services, healthcare delivery and construction meet almost none, because their bottlenecks are regulatory, physical or interpersonal rather than informational.
That distribution matters for interpreting the aggregate data. A large effect concentrated in a few sectors can be invisible in headline growth for years while being entirely real inside those industries. It also means the sectors most vocal about AI adoption are not necessarily the ones where measurement will find anything.
How this resolves, either way
Both scenarios produce observable consequences within a few years.
If it is a lag, the sequence should be firm-level returns appearing first, in specific functions, with countable metrics. Back-office automation and document-heavy workflows are where the early evidence would show, before anything reaches national accounts.
If the return is genuinely absent, the sequence is capital expenditure moderating, abandonment rates staying elevated, and the investment case shifting from productivity to competitive necessity. That second framing is already appearing in vendor messaging, which is itself a mild signal.
The honest position is that neither is established. Anyone claiming certainty in either direction is describing a preference rather than a finding.
What to watch
| Signal | What it would indicate |
|---|---|
| Abandonment rates in enterprise surveys | Falling rates would suggest firms are finding returns; rising rates suggest the opposite |
| Where returns are reported by function | A lag resolves function by function, starting with countable back-office work |
| Capital expenditure guidance direction | The companies with the best information about demand are the ones spending |
| Labour productivity in AI-intensive sectors | Sector-level data moves before aggregate GDP does and is published quarterly |
The last row is the most useful and the least followed. Sector-level productivity data is published regularly and would show a genuine effect years before it became visible in headline growth figures.
Why this matters beyond economics
The macro question determines how long the capital cycle can run. Investment sustained on the expectation of productivity gains needs those gains to arrive within the period the capital was underwritten for.
The previous paradox resolved over roughly a decade. Much of the current investment is being financed on shorter horizons than that, which creates a mismatch between when returns are expected and when they historically arrive.
That is the specific reason this is not merely an academic argument. It is the same question as whether the build is affordable, arriving from a different direction, and it is examined from the financing side in the analysis of AI debt issuance.
Frequently asked questions
Has AI increased GDP?
Not measurably so far. Multiple economic analyses report no measurable positive impact on US GDP growth from AI investment through 2025, despite hundreds of billions in deployed capital. The investment itself registers as economic activity through construction, hardware purchases and hiring. What is absent is evidence of the productivity improvement that would justify it.
What is the productivity paradox?
It describes the observation, prominent in the 1970s and 1980s, that computers were visible everywhere except in productivity statistics. Enormous computing investment produced no corresponding measured gain for years. Measurable productivity growth eventually accelerated in the mid 1990s, roughly a decade after substantial investment began, once organisations had redesigned processes around the technology.
How long should AI take to show up in economic data?
The previous cycle took roughly a decade from substantial investment to measurable productivity gains, and required complementary organisational change rather than just better hardware. Meaningful enterprise AI adoption began in 2023 and 2024, so expecting macro-level evidence by 2026 would be faster than any comparable general-purpose technology has managed.
Why do firms report no return from AI either?
This is the strongest argument against the measurement lag explanation. MIT reported around 95% of enterprise generative AI pilots showing no measurable profit and loss impact, and separate research found 42% of companies abandoning most AI initiatives in 2025. A lag would typically show firms seeing gains before statistics captured them. Currently neither does.
Could AI be improving quality rather than output?
Yes, and national accounts capture quality improvements poorly. A report that is better rather than faster to produce registers as nothing in output statistics. Much of what AI currently does has that shape, which is a legitimate reason to expect the measured effect to understate the real one, at least for a period.
What would prove AI is working economically?
Firm-level returns appearing function by function, starting with countable back-office work where output is measurable, followed by labour productivity gains in AI-intensive sectors. Sector-level productivity data is published quarterly and would move years before headline GDP growth did, which makes it the most useful indicator available.
Where to start this week
One question worth asking inside your own organisation, because it is the version of this you can actually answer.
Pick a workflow where AI has been deployed for at least six months and ask what the output was before and what it is now. Not effort, not speed, output. Documents shipped, tickets resolved, cases closed.
If the number moved, you have a data point that the macro statistics do not yet contain. If it did not, you have learned that the bottleneck sits somewhere the tool never touched, which is the more common finding and the more useful one.
There is a version of this that scales. Run it across five workflows rather than one, and the pattern that emerges tells you whether your organisation has an AI problem or a process problem. Those require completely different responses and are routinely confused.
The follow-up matters as much as the first answer. If output did not move, find the step that actually gates the process and check whether anything has been deployed there at all. In most organisations the answer is no, and that is where the next deployment should go rather than into another tool for a step that was already fast enough.
References
- Medium, AI bubble 2026: capex, Fed warnings and GPU lifespans, June 2026. Used for the absence of measurable GDP impact through 2025.
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025, as reported by Fortune. Used for the enterprise pilot return figure.
- CIO Dive, AI project failure rates are on the rise, March 2025, reporting S&P Global Market Intelligence. Used for abandonment and proof of concept figures.
- IntuitionLabs, AI bubble vs dot-com bubble: a data-driven comparison. Used for enterprise adoption and spending context.
The MIT figure is preliminary and was not peer reviewed. Macro claims about GDP impact are drawn from published economic analyses rather than from official statistical agency attribution, since national accounts do not isolate AI as a category.
Related reading