SemiAnalysis: Trillion-scale AI 'Hidden Economic Output' Underestimated
N.R. Finch
AI is creating roughly $1.5 trillion in economic value that traditional GDP statistics cannot capture; ignoring this blind spot risks mislabeling a real technology boom as a bubble.
What is "dark output," and why does it matter?
As AI replaces white-collar tasks at scale, production shifts from external procurement to in-house AI processing. This means → transaction volumes shrink, GDP looks like it is contracting, but actual output has not declined.
Research firm SemiAnalysis estimates that AI can currently replace or augment labor tasks worth roughly $1.5 trillion, concentrated in traditional services — value that goes uncounted because it cannot be directly measured.
In plain terms = the work gets done and value gets created, but so little money changes hands that the statistical ruler cannot pick it up.
How does a cost collapse make economic data "lie"?
Consider legal documents: drafting a will cost $400 in 1990; in 2026 a large-model API does it for $0.50 — a drop of over 99%. This means → the same will goes from a visible transaction in macro data to a rounding error that effectively "disappears."
Traditional statistics derive service output by dividing total spending by price. Once AI crushes prices, total spending shrinks, and the statistical system records a decline in service output — even as actual volume rises.
This reflects a fundamental mismatch: the measurement framework was built for an era of "pay people for services." Faced with "a few dollars of compute replacing tens of thousands of dollars of labor," it systematically understates real output.
Two types of dark output — which one will be bigger?
SemiAnalysis splits dark output into two categories: substitution (AI takes over tasks humans used to do) and novel (entirely new output created because AI makes it cheap enough to exist).
In plain terms = substitution is "the machine does the old job"; novel is "nobody did this before because it cost too much — now AI makes it almost free."
The firm argues that over time, novel dark output will far exceed substitution. This means → the volume of economic value invisible to statistics will keep growing, not converge.
What are token consumption and jobs data revealing?
Anthropic's March 2026 data shows 37% of token consumption goes to computing tasks, yet official software-investment share of GDP has not broken above its pre-AI-boom trend. This reflects a clear disconnect between token consumption and officially recorded economic output.
The labor market is diverging too: industries most exposed to AI are seeing slowing job growth, but because low-paid junior roles are eliminated first, reported average wages in those sectors are actually rising.
Put simply = fewer people, higher average pay — not because everyone got a raise, but because the cheapest positions vanished first, pulling the average up.
What does this mean for markets and policy?
Incoming Fed Chair Kevin Warsh noted in late 2025 that relying on traditional lagging data will cause policy to fall behind. He argued the economy can sustain faster growth without triggering inflation, and policymakers must act preemptively.
This means → if decision-makers watch only GDP and inflation, they risk tightening during a genuine technology boom and choking off growth.
SemiAnalysis's core warning: the multiple divergences — between token consumption and economic output, between employment and wages — are tangible evidence that dark output is expanding, and the central reason markets are debating whether AI is a bubble.
Content is for reference only, not financial advice.