Everything About AI
NextFin News - The world is pouring trillions into artificial intelligence, yet the economic statistics designed to track that transformation are flying blind. US nonfarm business productivity jumped 4.9% in the third quarter of 2025, the kind of surge that should settle a decades-old debate about whether technology ever shows up in the data. At the same time, national accounts still cannot say how much AI is actually being used, what it produces, or whether the $5.3 trillion that technology companies plan to spend on AI infrastructure through 2030 will pay for itself. The result is a rare and dangerous mismatch: capital markets are pricing a revolution that official data can neither confirm nor deny, and policymakers are steering with instruments built for a factory economy.
The Economy That Exists but Cannot Be Counted
The first thing to understand is that the AI economy is being built in two parallel worlds. In one, the numbers are enormous and precise. Goldman Sachs Research estimates that annual AI-related capital expenditure will reach $765 billion in 2026 and climb to $1.6 trillion by 2031, with cumulative spending of $7.6 trillion across compute, data centers, and power. The same analysts put combined capital spending by the large technology companies leading the buildout at $5.3 trillion between 2025 and 2030. In the other world, the numbers are blank.
National accounts were designed to count things that change hands for money in a manufacturing economy: tons of steel, hours of paid labour, machines sold at invoice prices. Artificial intelligence does not fit that frame. When a firm replaces paid customer-support staff with a free chatbot embedded in its software, measured output falls even though the same work gets done. When a household uses an AI assistant to draft a tax return, unpaid labour substitutes for paid labour, and GDP records the loss of the accountant's fee as a contraction. When a company trains a proprietary model on its own data, the asset it creates has no market transaction and therefore no entry in the accounts.
The statistical system knows this. In 2025, the United Nations Statistical Commission adopted the 2025 System of National Accounts, which for the first time formally recognises "data and databases" as a produced asset alongside software and research-and-development. That is a landmark change, but it is also an admission of how far behind the boundary sits. Organisational capital, human capital, brand equity, customer relationships, and non-R&D intellectual property — the categories that research consistently identifies as the largest drivers of firm-level productivity — remain outside the asset boundary.
The gap is not a rounding error. Jason Furman, a professor at Harvard University's Kennedy School, calculated from Bureau of Economic Analysis data that 92% of US gross domestic product growth in the first half of 2025 came from investment in information-processing equipment and software. That is growth recorded as an input, not as an output. The economy is buying shovels; the accounts cannot see the holes.
"GDP is not doing a very good job of reflecting how the economy is today," Diane Coyle, the Cambridge economist who has spent a career studying economic measurement, has argued. "We're making policy decisions about a technology we can't properly measure, using statistics designed for a manufacturing economy."
The mechanism is straightforward and it matters for every decision that follows. Measured GDP counts the capital expenditure — the servers, the chips, the data-center structures — the moment the money is spent. It does not count the productivity gain until that capital produces a marketable good or service at a recorded price. It does not count consumer surplus from free or bundled AI tools at all. So headline GDP can simultaneously overstate AI's contribution, by crediting the investment before any return, and understate AI's welfare payoff, by missing the value that never passes through a market. A statistic that can err in both directions at once is not merely noisy. It is structurally blind.
The Productivity Paradox Returns — With Higher Stakes
Robert Solow's 1987 observation — that you could see the computer age everywhere but in the productivity statistics — defined the first productivity paradox. The current episode has the same shape but different stakes, because this time the investment is being financed on credit and concentrated in a handful of listed companies.
The optimists have a genuine case. Erik Brynjolfsson, director of Stanford's Digital Economy Lab, has argued that the AI productivity take-off is finally visible in US data. New benchmark revisions from the Bureau of Labor Statistics reduced previously reported payroll growth by about 403,000 jobs, while real GDP held up — including 3.7% growth in the fourth quarter. Output rising while measured labour input falls is the hallmark of productivity growth. His own updated analysis points to US labour productivity growth of roughly 2.7% in 2025, nearly double the 1.4% annual average that characterised the previous decade.
"The updated 2025 U.S. data suggests we are now transitioning out of this investment phase into a harvest phase where those earlier efforts begin to manifest as measurable output," Brynjolfsson has written.
There is micro evidence behind the claim. A study published in the Quarterly Journal of Economics by Brynjolfsson, Danielle Li, and Lindsey R. Raymond examined the staggered rollout of a generative AI conversational assistant to 5,172 customer-support agents at a Fortune 500 software company. Agents who used the tool resolved 14% to 15% more issues per hour on average, with gains of 30% to 35% for the least experienced agents. The distribution matters: highly skilled agents saw minimal gains and, in some cases, slight quality declines. That is augmentation, not replacement — and it is concentrated where the least productive workers are.
The Bureau of Labor Statistics reported nonfarm business productivity up 4.9% in the third quarter of 2025, with the second quarter revised upward to 4.1%, while unit labour costs declined for two consecutive quarters — a pattern not seen since 2019. On its face, that is the harvest arriving.
But the counter-evidence is not trivial. The Yale Budget Lab warned in February 2026 against reading too much into a single strong print: job growth in 2025 was barely above zero at about 15,000 a month even as GDP came in strong, and a quarter that prints hot after years of downward payroll revisions is not a regime change. Alexander Arnon of the Penn Wharton Budget Model estimates that AI will raise productivity and GDP by 1.5% by 2035, nearly 3% by 2055, and 3.7% by 2075 — but with a permanent effect on the annual growth rate of less than 0.04 percentage points, because sectoral shifts dilute the impulse. On that reading, AI lifts the level of output without permanently accelerating the rate. The paradox is not solved; it is deferred.
The mechanism that explains both sides is the J-curve of intangible complementarity. General-purpose technologies do not raise measured productivity when they arrive; they raise it only after firms make the unmeasured investments that let them use the technology — workflow redesign, training, data cleaning, new management practices. Those complements are themselves largely intangible and largely invisible in the accounts. The lag between the capex and the measured output is the paradox. It is also the trap: investors and policymakers who cannot see the complements will mistake the investment phase for the payoff.
The Capex Boom Is Real — and It Is Being Financed on Debt
The scale of the buildout is not in doubt, and neither is how it is being paid for. Goldman Sachs Research expects AI-related capital expenditure to grow from $765 billion in 2026 to $1.6 trillion in 2031. Hyperscalers are increasingly turning to debt markets to fund it. Investment-grade bond issuance by the large technology companies totalled $108 billion in 2025, roughly 26% of their capital expenditure. Through the first half of 2026 they had already issued $194 billion, with full-year issuance expected to reach about $250 billion, or roughly a third of capex. That figure does not include an additional $412 billion of AI-linked debt issued by companies outside the hyperscalers in 2026, spanning software, semiconductors, and data-center financing.
The concentration of the equity side is equally striking. As of July 2026, Nvidia's market capitalisation stood at $4.86 trillion, ahead of Apple at $4.51 trillion and Alphabet at $4.36 trillion, with Microsoft and Amazon at $3.45 trillion and $2.93 trillion. Earlier in the year, analysis showed Nvidia's value alone was more than double the combined worth of every company in the S&P 500 energy sector, and 14% greater than the materials, utilities, and real estate sectors combined. Diane Coyle noted that AI and technology shares accounted for about three-quarters of the S&P 500's gains in 2025, while venture capital investors poured an estimated $200 billion into the AI sector that year.
Here is the second-order problem that the aggregate data hide. The returns on this $5.3 trillion of spending must flow through one of two channels. Either AI raises broad-based productivity, in which case the gain is real but statistically undercounted; or it generates monopoly rents captured by a handful of firms, in which case the gain is counted but concentrated. Either way, aggregate GDP becomes a poorer guide to broad welfare than it was in an economy where investment, employment, and output moved together.
There is also a maturity structure to this bet. A third of capex funded by investment-grade bonds issued in 2026 creates a refinancing wall in the 2028-2030 window — precisely when the productivity payoff must be visible to justify the valuations. If monetisation lags, the debt service does not. The measurement gap therefore cuts both ways for investors: it may hide the upside of genuine productivity gains, but it also hides the point at which the capex cycle has overshot.
The Strongest Case Against the Skeptics — and Where It Breaks
The bullish case deserves to be stated at full strength, because it is not a strawman. The data really do show a decoupling of output from labour input. Payroll revisions cut 403,000 jobs while GDP held. Productivity in 2025 ran at 2.7%, double the decade average. Unit labour costs fell for two quarters. Employment for workers aged 22 to 25 in the most AI-exposed occupations fell about 19% relative to their peers in less exposed fields, while workers using AI tools to augment their skills saw stronger outcomes. Firm-level studies show double-digit productivity gains from generative AI. This is not a narrative built on faith; it is a narrative built on the first hard numbers that have appeared in thirty years.
The rebuttal is not that these numbers are wrong. It is that they are narrow, and that narrowness is exactly what the aggregate statistics cannot reveal. The 92%-of-growth-from-capex figure shows the expansion is still input-led. The Yale Budget Lab's caution stands: one strong print after years of downward revisions is not a regime confirmation. The MIT study's own distribution — gains concentrated among the least skilled, with quality declines among the best — implies augmentation at the margin, not economy-wide transformation. And the Penn Wharton estimate of a sub-0.04-percentage-point permanent effect on the annual growth rate is a reminder that a one-time level shift, however large, is not the same as a permanently higher growth path.
The counter-case proves that AI is working at the firm level. It does not prove that the macro measurement problem is solved, or that the capex will earn its cost of capital. Those are two different questions, and the statistics conflate them.
What to Watch: A Structural Shift With a Cyclical Leg
The central judgment is that this is a structural change with a cyclical overlay, and the two must be separated. The structural leg will not revert on its own: the 2025 System of National Accounts reform is itself evidence that the old measurement boundary is broken; the shift to intangible and data assets is permanent; free and bundled digital goods permanently separate market transactions from welfare; and AI agents will push more activity across the production boundary, into unpaid or unpriced space. The cyclical leg — the capex cycle and the single-quarter productivity spike — will mean-revert if monetisation disappoints or financing conditions tighten.
The beneficiaries are identifiable. Firms that can capture AI rents in priced transactions — the hyperscalers, chip designers, and workflow-software companies with distribution — will show up in the data regardless of the measurement gap. Statistics offices that adopt SNA 2025 and build satellite accounts for AI activity will see the economy sooner. Investors who read firm-level microdata rather than headline GDP will have an informational edge.
The exposed are equally identifiable. Policymakers setting interest rates on headline GDP and inflation prints are navigating with a distorted map. Bond investors underwriting $250 billion or more of annual hyperscaler issuance on GDP-growth assumptions face a refinancing wall in 2028-2030. Workers in entry-level, AI-exposed roles — where employment for the youngest cohort has already fallen about 19% — face the first wave of adjustment. Economies that lag in adopting the 2025 national-accounts standards will be the last to see their own transformation.
The forward look splits by horizon. In the short term, expect volatility around capital-expenditure guidance and the next productivity prints; the market will keep pricing the boom until a financing signal breaks. In the medium term, the 2027-2030 refinancing window meets the monetisation test, and that is when the capex crowd separates into winners and stranded assets. In the long term, the question is institutional: whether national-accounts reform — SNA 2025 adoption, satellite accounts for AI, time-use and household-production measurement — arrives fast enough to make the next general-purpose technology visible while it is still happening.
Three scenarios frame the range. In the base case, productivity grinds up toward 2% annually as complements diffuse, GDP remains a partial picture, capex moderates toward $1 trillion by 2030, and measurement reform proceeds slowly through satellite accounts. In the upside case, agentic AI pushes measured productivity above 3% by 2028, SNA 2025 data-asset capitalisation lifts measured investment and closes part of the gap, and the harvest narrative is vindicated. In the downside case, monetisation lags, refinancing hits at higher rates, capex rolls over, productivity reverts toward the 1.4% decade average, and the boom is re-rated as a 1999-style overbuild — with the difference that the measurement system would not see the bust coming either.
There is one falsifying signal worth naming. If Bureau of Labor Statistics nonfarm business productivity growth averages below 1.5% annualised through 2027 while annual AI capital expenditure exceeds $1 trillion, the structural-productivity thesis is wrong: the boom is a capex cycle, not a transformation.
The AI revolution may be real. What is certain is that the dashboard we use to navigate it is not.
Explore more exclusive insights at nextfin.ai.
