AINEWS
Why the official numbers miss the AI buildout, and what the 2025 international guidelines already say to do about it.
The Q2 2026 BEA advance estimate reads as a familiar, mid-2-percent kind of growth number. For an economy supposedly in the middle of an AI revolution, that headline looks dull. A new analysis from Epoch AI argues the dullness is an artifact: the official accounts may be understating US growth by roughly 0.3 percentage points over the last year, with a projected gap of close to 2 percentage points per year by 2028 if Nvidia's growth pace holds.
The reason sits in a category distinction most readers have never had to think about. GDP tracks value across four export lanes: goods, intellectual property (IP), services, and merchanting. A chip designed in California, fabricated in Taiwan, and sold to a buyer in Singapore is a "factoryless" export: the work happened in the United States, but no physical good crosses a US border, and no foreign buyer pays the US firm explicitly for the design IP. Epoch AI reviewed all four categories, and Nvidia's value-add appears in none of them. Epoch AI's full report confirmed this reading with BEA staff, though any characterization of the agency's full position reflects that analyst outreach, not a BEA endorsement of the conclusions.
The blind spot is not new. It was first written into US national accounts when the Bureau of Economic Analysis began studying factoryless goods producers more than a decade ago, and it applies to any US firm that designs abroad-built goods. Until recently, the missing value was small enough that nobody worried about it. Spending on computers and peripheral equipment has roughly tripled since 2023, and AI compute equipment investment alone now runs near $400 billion a year. At that scale, the unmeasured slice of US growth is no longer rounding error.
This matters for how the cycle gets read. A Federal Reserve weighing whether the economy is running hot or cool leans on GDP, productivity, and trade balance; if productivity looks low and the trade balance looks like the US is selling less than it actually is, the conclusion can tilt toward "the AI boom has not shown up in the data" when the better description is "the AI boom is showing up off the books." Industrial-policy arguments that lean on measured US output also lean on a number that has been undercounting the country it is meant to describe.
The constructive fix is already in the rulebook. According to Epoch AI, international guidelines updated in 2025 call for recording factoryless manufacturers' overseas sales as goods exports, with a narrower alternative: record the markup as an IP export. Either change would put US-designed chips back on the scoreboard, though both depart from current US practice and BEA has not committed to a timeline. BEA's 2014 Q&A on the international guidelines and the companion document on conceptual impacts show the agency has worked through these issues for years. Implementation will take time, and until it lands, the AI buildout will keep showing up smaller than it is.
The practical read for the next GDP release: when the headline number looks weak, ask whether the fabless design economy is being captured. A US-engineered chip that runs in a data center in Singapore does not show up in goods exports, IP exports, services exports, or merchanting. Until BEA implements the 2025 guidelines, the AI boom is showing up off the books, not failing to show up.