AI Versus the Rest of the Economy: Inside America’s 1.5 Percent Problem

The Bureau of Economic Analysis put second-quarter growth at 1.5 percent annualized. That reading sits well below the 2.1 percent the economy managed in the first quarter. It also undershot the 2.1 percent consensus that economists had penciled in beforehand. What Really Drove US Growth

Beneath that soft headline, the composition tells a stranger story. Corporate profits jumped by $400.9 billion in a single quarter. Profit margins reached 14.9 percent of GDP, an all-time record, according to EY-Parthenon’s reading of the release.

So American companies earned more than ever, while the economy they operate in slowed sharply. That gap sits at the center of this year’s argument about artificial intelligence. One question now dominates every macro desk in New York and London. How much of US growth actually belongs to the AI buildout, and how much belongs to everything else?

Why the AI share is a contested figure

Start with an inconvenient fact. The national accounts contain no line item labelled “artificial intelligence”.

Statisticians at the BEA track information-processing equipment, intellectual property products, software, and non-residential structures. Analysts then decide which slices of those categories belong to the AI buildout. Every analyst draws that boundary somewhere different.

Choose a wide boundary and AI looks like the entire economy. Choose a narrow one and AI nearly disappears. Subtract imported chips and the number falls again, because imports reduce measured GDP.

This is not a rounding dispute. The published estimates for AI’s share of US growth range from roughly zero to 92 percent. We think readers deserve the full spread rather than whichever figure suits a given argument.

The 92 percent claim, and what it measured

Harvard economist Jason Furman produced the number that launched the debate. He showed that investment in information-processing equipment and software equalled just 4 percent of US GDP. That same sliver accounted for 92 percent of GDP growth across the first half of 2025.

Strip those categories out, Furman calculated, and the rest of the economy grew at 0.1 percent annualised. That is a standstill in everything but name.

Furman himself refused to push the claim too far. He noted that “absent the AI boom we would probably have lower interest rates [and] electricity prices”. Cheaper money and cheaper power would have produced growth elsewhere. The counterfactual economy is not a frozen one.

Goldman’s objection: the chips come from Taiwan

Goldman Sachs chief economist Jan Hatzius reached the opposite conclusion. He told the Atlantic Council that AI investment contributed “basically zero” to US growth in 2025.

His reasoning turns on geography. American hyperscalers buy enormous volumes of accelerators, servers, and networking gear from abroad. Those purchases lift Taiwanese and Korean output, not American output.

Imports subtract from GDP by construction. A data center filled with foreign silicon therefore adds far less to domestic growth than its price tag suggests. Goldman’s Joseph Briggs argued that the AI-drives-everything narrative felt intuitive enough to escape scrutiny.

The second-quarter trade numbers support the mechanism. Imports surged 12.5 percent while exports rose only 4.5 percent. Net trade alone subtracted 1.1 percentage points from headline growth.

Where the defensible middle sits

Between those poles, several institutions have done the arithmetic carefully. ING’s economics team published the most transparent recent attempt, in August 2026.

Their narrow measure covers computing, peripherals, and software investment only. On that basis, tech capital spending delivered 44 percent of second-quarter GDP growth. The four-quarter average sits at 43.3 percent.

ING then subtracts net imports of computers, peripherals, and semiconductors. That adjustment cuts the contribution to 36 percent for the second quarter. The four-quarter average on the same basis reaches 37 percent. ING calls this its fairest available measure and concludes that AI accounts for roughly a third of 2026 growth.

Bridgewater reached a similar order of magnitude by a different route. Its researchers estimate AI capex adds about 140 basis points to 2026 US growth.

The scale of the capital wave

The dollar figures explain why the argument matters so much. TrendForce raised its 2026 forecast for the world’s nine largest cloud providers to roughly $830 billion in capital spending.

Among the US names, Microsoft guided to about $190 billion. Google guided to $180–190 billion, Meta to $125–145 billion, and AWS to more than $230 billion. Each of those four firms now outspends most national governments on physical infrastructure.

Apollo Global Management sizes 2026 hyperscaler capex at roughly $646 billion. That equals about 2 percent of US GDP flowing through a handful of corporate budgets.

Goldman Sachs projects $1.15 trillion of hyperscaler capex across 2025 to 2027. The comparable figure for 2022 to 2024 was $477 billion. The spending has more than doubled in three years.

A different lens: share of output, not share of growth

Share-of-growth figures swing violently, because growth itself is a small number. Share-of-output figures behave far better, and they deserve more attention than they get.

Epoch AI tracks the buildout using BEA computer and peripheral data, Census construction spending, and its own networking estimates. Its researchers size AI-related data centers, hardware, and networking together. That bundle reached about 0.8 percent of US GDP in the first quarter of 2026.

Computing infrastructure as a whole climbed to roughly 1.5 percent of GDP. The 2015–2022 average was about 0.7 percent. The AI boom has therefore doubled computing’s footprint in national output within three years.

That is large, historically unusual, and still modest in absolute terms. It is not the economy. It is the margin that moves the economy.

The quarter the AI engine idled

Here is the finding that most coverage has missed. The AI capex impulse weakened sharply in the second quarter of 2026.

High-technology equipment investment grew at 39.9 percent annualised in the first quarter. That rate collapsed to 6.9 percent in the second. Information-processing equipment rose only 8.3 percent, against 13.6 percent for equipment overall.

Non-residential structures investment fell again, declining 1.8 percent on EY’s reading of the second estimate. Some data providers record a tenth consecutive quarterly contraction in structures.

We would caution against calling this the end of the buildout. One quarter of deceleration proves very little on its own. But the quarter matters for a different reason. The AI-carries-everything thesis stopped describing reality, and growth slowed anyway.

What picked up the slack

Other parts of the economy quietly took over. Industrial equipment investment surged 27.8 percent annualised. Transportation equipment rose 27.1 percent.

Consumer spending accelerated hard, from 0.5 percent in the first quarter to 3.4 percent in the second. Households bought vehicles, furniture, clothing, and restaurant meals. Services outlays rose 3.1 percent, led by recreation at 6.8 percent.

Residential investment rose 1.2 percent. That marked the first increase since the fourth quarter of 2024.

Real final sales to private domestic purchasers advanced 4.2 percent. That measure strips out trade, inventories, and government, and it looks genuinely strong.

So why did headline growth fall to 1.5 percent? Inventories subtracted 0.7 points. Net trade subtracted 1.1. Federal non-defense spending collapsed 13.1 percent.

The labour market runs on a different clock

Employment tells a harsher story than output does. American employers added 162,000 jobs in August, far above the 53,000 consensus.

That headline flatters the trend badly. Average monthly job growth over the prior twelve months ran at just 31,000. August was an outlier, not a norm.

The unemployment rate held at 4.1 percent. Labor-force participation edged up to 61.6 percent, though it remains 0.5 points below January. Long-term unemployment covers 1.9 million people, or 27 percent of all unemployed Americans.

Vanguard senior economist Adam Schickling captured the split precisely. The labor market, he said, “is showing signs of near-term cyclical strength, even as longer-term structural concerns remain”.

The entry-level door closes first

Aggregate unemployment hides the distributional damage. Researchers at Stanford examined ADP payroll records directly. They found a 13 percent relative employment decline among workers aged 22 to 25. That decline hit the most AI-exposed occupations, while older workers in identical roles held steady.

The August payroll report showed information-sector employment falling, even as bars and restaurants led job creation. Analysts at CNBC linked that pattern directly to AI investment.

Challenger, Gray & Christmas has tracked employer-stated reasons for layoffs since 1989. In March 2026, AI led all cited reasons, accounting for 15,341 announced cuts. That equalled 25 percent of the month’s total.

We read these numbers as a composition shift rather than a jobs apocalypse. The ladder is losing its bottom rungs.

Productivity: the one place AI clearly lands

Output statistics argue about AI. Productivity statistics increasingly do not.

Economists at the Federal Reserve Bank of Dallas published the cleanest evidence in July 2026. US labour productivity has grown at 2.4 percent annualised since the start of 2024. The five pre-pandemic years averaged 1.6 percent.

The gains concentrate sharply. Finance, professional and technical services, and information account for only 16 percent of total hours worked. Those three sectors delivered 3.7 percent annualised productivity growth since early 2024. The rest of the economy managed 1.7 percent.

The Dallas Fed cross-checked the pattern internationally, using variation in national AI adoption. The relationship holds across countries. That consistency makes coincidence a weaker explanation.

A Fed-branded caveat on the productivity story

We should not oversell that finding. A separate NBER paper, based on an Atlanta Fed survey of roughly 750 executives, reached a more cautious verdict.

Those executives report positive productivity gains that vary widely by sector. High-skill services and finance lead. Manufacturing and construction show gains closer to 0.4 percent.

Crucially, the researchers found those gains do not flow mainly from capital deepening. They flow from revenue-based total factor productivity, meaning innovation and demand channels rather than new machines.

Economy-wide productivity growth also slowed in the first quarter of 2026. Robert Solow’s old paradox has not fully retired. Firms see AI working in their own operations, while the aggregate statistics remain ambiguous.

The electricity bill arrives

Every megawatt the buildout consumes has to come from somewhere. American households increasingly suspect they are paying for it.

Residential electricity rates rose 7.3 percent nationally over the past year. Utilities requested $18.6 billion in rate increases during the first half of 2026 alone.

The clearest evidence comes from PJM Interconnection, which operates the grid across thirteen states. Its capacity price climbed from $28.92 per megawatt-day in 2024 to $329.17 today. PJM’s own market monitor named data-centre demand as a primary cause. It attributed $23 billion of customer price increases to that demand, running to at least 2028.

Gasoline compounds the squeeze. Pump prices rose 27.4 percent over the year to August, and fuel oil rose 52 percent.

The case against blaming data centres

Fairness demands the counter-evidence, and it is substantial. The Electric Power Research Institute found that data centres put downward pressure on average electricity prices through 2024.

Researchers at Lawrence Berkeley National Laboratory reached a similar conclusion. They attribute the price increases mainly to fuel costs, purchased power, ageing grid infrastructure, and disaster recovery. Their work explicitly lists data centres among factors not driving national inflation-adjusted electricity prices.

The honest reading splits the difference. Nationally, data centres have not yet moved the average price much. Locally, in PJM territory and in Northern Virginia, the effect looks real and large.

Politics will not wait for the econometrics. Roughly three-quarters of Virginia voters already blame data centres for their bills.

Main Street’s version of the boom

Small business owners describe an economy with no AI tailwind at all. The NFIB Optimism Index slipped 1.1 points to 98.7 in August, barely above its 52-year average of 98.0.

The internals look worse than the headline. The share of owners expecting the economy to improve dropped five points. Earnings trends deteriorated by three points. Net hiring plans fell to 17 percent.

NFIB chief economist Bill Dunkelberg noted that “uncertainty remains elevated among small business owners”. His Uncertainty Index stands at 89, against a historical norm of 68.

TD Economics spotted the crucial divergence. Strong August payrolls alongside weak small-business hiring suggests the job gains concentrated among large firms.

Two consumers, one spending number

Aggregate consumption looks healthy. The distribution underneath it does not.

EY-Parthenon chief economist Gregory Daco identified the fault line explicitly. He points to income erosion across lower- and upper-middle-income households. He sees the same erosion among “smaller firms operating outside the AI investment orbit”.

Credit data confirms the split. Aggregate commercial-bank card delinquency sat at 2.92 percent in the first quarter of 2026. At banks outside the top 100 by assets, where lower-income borrowers concentrate, the rate reached 6.43 percent.

Total household debt slipped $13 billion to $18.8 trillion in the second quarter. Some 4.7 percent of outstanding balances remain delinquent in some stage. The averages conceal two quite different Americas.

The buildout now runs on borrowed money

For most of this cycle, hyperscalers funded capex from operating cash flow. That constraint has now broken.

Alphabet reported negative free cash flow in the second quarter of 2026, its first negative reading in decades. Its long-term debt more than doubled to $98 billion across the first half. Amazon’s long-term debt jumped 81 percent to $119 billion in a single quarter.

Collective hyperscaler dollar debt has more than doubled to over $360 billion. Moody’s flags roughly $662 billion of signed but not yet commenced data-centre leases, sitting off balance sheet.

PIMCO estimates hyperscaler capex will absorb about 94 percent of operating cash flow across 2026 and 2027. Morgan Stanley and J.P. Morgan expect the technology sector to issue around $1.5 trillion of new debt.

The Fed tightens into a one-engine economy

The Federal Open Market Committee opened its two-day meeting on 15 September. Markets have effectively fully priced a quarter-point increase, which would be the first hike since 2023.

The case rests on energy. Headline CPI held at 3.4 percent in August, while gasoline contributed over a third of the monthly increase. Producer prices run at 5.4 percent annually. Brent crude trades above $100 amid the continuing Iran conflict.

Core CPI tells a gentler story at 2.4 percent, the lowest since March 2021. Chair Kevin Warsh has nonetheless insisted the Fed has “work to do” on inflation.

Nationwide chief economist Kathy Bostjancic framed the reaction function bluntly. Rates, she said, “can remain on hold only if disinflation continues”.

Why a hike lands differently this cycle

Consider what higher rates now do to this particular expansion. The marginal growth engine has become a leveraged, capital-intensive construction programme.

Hyperscalers issue investment-grade debt at scale, so their funding cost tracks Treasury yields directly. Neoclouds and private-credit-financed operators sit further out the risk curve, and they feel tightening faster.

Furman’s caveat now cuts the other way. He argued the AI boom had itself raised interest rates and electricity prices, suppressing other sectors. A tightening cycle driven partly by AI-adjacent inflation would compound that crowding-out effect.

Meanwhile the non-AI economy carries the rate increase without the capex offset. Small firms, renters, and subprime borrowers absorb the cost. They will not receive the GPUs.

What we actually conclude

We reject both extreme readings of this economy. AI did not deliver 92 percent of 2026 growth, and it did not deliver zero.

The defensible range for 2026 sits somewhere around a third of growth, once analysts subtract imported hardware. That contribution remains extraordinary by any historical standard. Business investment has not concentrated this narrowly since the late-1990s technology cycle.

The more useful finding concerns direction of travel. AI’s contribution to measured growth peaked in late 2025 and early 2026. High-tech equipment investment then decelerated sharply, from 39.9 percent to 6.9 percent annualised.

Growth fell anyway, because trade, inventories, and federal spending all turned negative. The rest of the economy did not collapse when AI paused. It simply never accelerated.

The risk nobody has priced properly

Here is the asymmetry that concerns us most. The AI buildout supports growth while spending rises, not while spending stays high.

GDP measures flows, not levels. Capex that plateaus at $800 billion contributes nothing further to growth. Capex that falls subtracts directly.

Bank of America expects AI and cloud capital spending to rise 15 percent in 2027. Growth this year runs at 36 percent. Goldman analysts warn that deceleration will force investors to separate winners from losers.

A plateau would therefore remove roughly a third of current growth contribution, without any bubble bursting. Nobody has to go bankrupt. The arithmetic does the damage on its own.

What we are watching next

Four dates matter over the coming eight weeks.

The FOMC announces its decision on 16 September. Watch the dot plot, which in June implied a 3.8 percent funds rate by year-end.

BEA publishes its annual national accounts update on 30 September. That revision covers GDP, industry data, and state accounts simultaneously for the first time. It may change the AI contribution figures materially.

The September employment report arrives on 2 October. We will read information-sector payrolls and the 22-to-25 cohort first.

Third-quarter hyperscaler earnings land in late October. Capex guidance for 2027 will tell us whether the plateau has begun.