Hype Index: 'AI Capex Is Carrying the US Economy'

TL;DR: AI capex really is doing most of the work in US GDP growth, and almost every viral version of that fact is arithmetically unreliable. I enumerated the top 20 results for "AI share of GDP growth 2026" on 2026-08-15 and fetched nine in full. Six attach a percentage to Q1 2026 growth, and they report five different values for that one quarter: 49.5%, 50%, 67%, 74% and 75%. Only two say which BEA estimate they used, which matters because BEA published Q1 2026 growth three times at 2.0%, then 1.6%, then 2.1%. The distinction nobody states: AI is a large share of GDP growth and a small share of GDP, roughly 0.8% of it. Score: 35% noise / 65% signal.

What is the claim?

It travels as a headline. TECHi ran "AI Capex Carries U.S. Economy". The most-forwarded version is a line from the Kobeissi Letter, relayed by Benzinga in May: "without this AI-driven tech investment, Q1 GDP growth would have been close to flat."

The origin is real and credited to a named economist. Everything downstream is where the trouble starts.

Is that a share of growth, or a share of the economy?

Before any figure below means anything, one distinction, and it is the entire post.

GDP is a level: the total value of what the country produces. GDP growth is the change in that level between two periods. A category can be tiny as a share of the level and enormous as a share of the change, because the change is a small number. If the economy grows 2% while one slice worth 4% of GDP grows fast and everything else sits flat, that slice can account for nearly all of the 2% and still be 4% of the economy the next morning.

This is exactly what Jason Furman, the Harvard economist and former chair of the Council of Economic Advisers, actually said. Per Fortune, his calculation was that "investment in information-processing equipment and software was only 4% of U.S. GDP for the first half of 2025, yet it also accounted for fully 92% of GDP growth over that period," with GDP excluding those categories growing 0.1% annualized.

Both halves of that sentence are true at once. They are one fact seen from two angles. The 4% is the level. The 92% is the growth. A reader who takes "92% of the economy is AI" out of it has inverted the finding, and that reading is loose in the wild.

Note the second thing about Furman's number: it is first half 2025. It is not a 2026 statistic, and it never was.

What did the count find?

I took the top 20 results for "AI share of GDP growth 2026" on 2026-08-15 and recorded each one's percentage, quarter, BEA estimate vintage and BEA category. Nine I fetched and read in full; the rest I could classify only by headline, so the figures below describe the nine.

Six of them attach a percentage to Q1 2026 growth. Here is every one:

Source	Figure	BEA vintage named	BEA category named
TFTC	74% (1.55pp of 2.1%)	Yes, third estimate	Yes, info processing equipment + IPP, Table 1.5.2
Beta Finch	~50% (1.09pp of 2.0%)	Yes, advance	Yes, computers and peripherals + software
Benzinga	67% (134bp of 2.0%)	No	Loosely, "software and IT equipment"
Reventure	49.5% (1.06pp)	No	Yes, computers and peripheral equipment only
Crypto Briefing	75%	No	No
TECHi	75%	No	No

Five distinct percentages for a single quarter. Four, if you accept Reventure's own headline rounding its 49.5% up to 50%, which is its own small tell: the headline and the body of that piece disagree.

Two of the six name the BEA vintage. Four name a line item at all.

Why do the numbers disagree?

Two reasons, neither a conspiracy. Both are undeclared choices.

The first is which BEA category you count. Reventure used computers and peripheral equipment alone and got 1.06 percentage points. TFTC summed information processing equipment and intellectual property products and got 1.55. Those are different questions, so they have different answers, and neither is wrong until it is presented as the number. TFTC deserves credit for saying so out loud: "The BEA does not publish a single 'AI contribution' line; that figure is analyst-derived."

The second is the denominator, and it moves. I downloaded BEA's vintage history file, last updated 2026-07-30. Q1 2026 real GDP growth was published at 2.0% on April 30, revised to 1.6% on May 28, then to 2.1% on June 25. A share-of-growth percentage calculated against the April print is computing against a denominator that no longer exists. Benzinga's 67% and Crypto Briefing's 75% both sit on the 2.0% figure, published before the revisions landed.

BEA's own Q2 2026 advance release is candid about why early prints move: the software component, the largest single input to these AI-share calculations, is estimated "based primarily on a judgmental trend and Bureau of Labor Statistics Current Employment Statistics." The number everyone divides by is partly a placeholder for its first two months of life.

I have not computed my own percentage, deliberately. Doing it properly requires NIPA Table 1.5.2 rather than the summary release, and adding a sixth number to five unreconciled ones is not a public service.

What is actually true?

The direction is not in dispute. Real GDP grew 2.1% in Q1 2026 and 1.5% in Q2 2026 on the advance estimate. Information processing equipment and software are a large and probably dominant share of that growth, and the spending behind it is auditable one filing at a time: capex ran 96% of operating cash flow across four hyperscalers in Q2 2026. Furman's underlying point, that the non-AI economy in H1 2025 was close to flat, is a serious finding that no counting dispute makes go away.

The level, meanwhile, stays small. Epoch AI puts AI-related data-center investment at about 0.8% of US GDP in Q1 2026, roughly 1.5% if you include all computing infrastructure. Epoch states its method (project the 2015 to 2022 trend in computers-and-peripherals share of GDP and attribute above-trend growth to AI) and flags its own limits: the category includes non-data-center business IT spending, all figures are nominal, and buildouts outside the US are excluded.

Sitting in the same set of search results is Forbes reporting TS Lombard's estimate that the US will "devote approximately 2% of gross domestic product to artificial intelligence and data center infrastructure in 2026." That is a level claim, published nine results away from a 75% growth claim, with nothing marking the difference. This is why readers end up believing AI is three quarters of the economy.

What does this not prove?

That the AI-carries-everything story is wrong. It mostly is not. My count measures the quality of the public arithmetic, not the underlying economy, and a badly sourced number can still point at a real thing. The mirror-image claim decays the same way: "half the US buildout is canceled" started life as one hedged sentence about global delays.

Furman himself supplied the strongest counterargument, and almost nobody repeats it: absent the AI boom, lower interest rates and lower electricity prices would have generated growth elsewhere, potentially offsetting around half of the measured contribution. The counterfactual is not "the same economy minus the data centers."

There is also a live argument that official statistics undercount AI rather than overcount it. A 2026 PIIE policy brief argues the AI economy's "extraordinary growth is largely invisible in conventional GDP statistics" and calls for AI satellite accounts. If PIIE is right, every number in my table is measuring the wrong object with admirable precision.

Verdict

The claim is directionally sound and numerically feral. A real calculation by a named economist, correctly dated to H1 2025, has become a floating percentage that moves 25 points depending on which category definition and which of three BEA vintages a writer grabbed. The signal: AI capex is genuinely carrying most of the measured growth. The noise: almost nobody repeating it can tell you what they divided by. The same pattern runs through the bubble debate and the fight over who pays for the buildout's electricity.

Verdict: 35% noise / 65% signal. Ask anyone quoting a percentage two questions: which quarter, and which BEA line. Most cannot answer either.