A GPU cluster in a data centre produces the analytical output of 10,000 knowledge workers. It does not eat lunch. It does not commute. It does not spend money at the restaurant downstairs, or pay rent, or buy its kids shoes.
GDP measures output. When the GPU cluster replaces the workers, the output is maintained. GDP holds. Productivity soars. The quarterly numbers look excellent.
But the income that those 10,000 people spent — at restaurants, on rent, on shoes — that income is gone. The output exists. The demand it used to fund does not.
We call this Ghost GDP: economic output that appears in the national accounts but never circulates through the real economy as consumer spending.
The arithmetic
This is not theory. It is an accounting identity.
Consumer spending is roughly 68% of US GDP. In most developed economies it runs 55-65%. Consumer spending is a function of consumer income. Consumer income is overwhelmingly derived from wages.
Our dataset tracks 2,860 occupations across 39 countries. Roughly 41% are rated HIGH or CRITICAL risk. These are disproportionately in the upper half of the income distribution. Professional services. Analytical roles. Mid-management. Administrative functions.
The top 20% of earners in the US account for roughly 40% of all consumer spending. These people are employed in exactly the occupations our model flags as elevated risk.
Do the arithmetic: if AI displaces 15% of these roles over three years — a rate consistent with what our employment change data suggests for the most exposed occupations — the direct impact is not 15% of their spending. It is the marginal spending they cut. The dinners out. The renovations. The vacations. The second car.
This is the demand destruction that does not show up in GDP, because GDP measures what is produced, not who gets paid.
The lag
High-income workers have savings buffers. When a senior product manager loses their job, they do not immediately stop spending. They draw down savings for three to six months. They tell themselves it is temporary.
This is why we track wage change and employment change as separate signals. Employment change tells you the stock — who has a job. Wage change tells you the flow — what the people who still have jobs are being paid.
When both decline simultaneously for the same occupation cluster, the consumer impact is imminent. Not hypothetical. Imminent.
In our data: wage change coverage is 96%. Employment change coverage is 99.5%. We see this in near real-time.
Why this time is structurally different
Every automation wave has produced the same reassurance: new industries will emerge, new jobs will be created, it will all work out. And every time, it has. New industries absorbed the displaced workers.
But here is what our data shows that is different: AI is displacing the absorber sector itself.
Previous waves displaced manual and routine workers. They moved into services and cognitive roles. The professional services economy was the sponge.
Now the sponge is the target. A displaced factory worker in 1985 could retrain as an office administrator. A displaced office administrator in 2026 can retrain as... what?
The roles above require decades of institutional trust. The roles below pay less. The roles sideways are being automated by the same technology.
This is not a prediction. It is an observation about the structure of the current displacement pattern, visible in our data.
What Ghost GDP looks like in the numbers
It does not look like a recession. It looks like everything is fine — until it isn't.
- Productivity up, wages flat or falling in the same occupation. We see this in 67% of HIGH-risk occupations.
- Vacancy counts stable, employment declining in the same occupation. Ghost demand.
- Corporate earnings beating expectations while consumer confidence weakens.
- GDP growth positive while real median income stagnates.
The gap between output and income is Ghost GDP. It is growing. Our data measures the occupation-level mechanics of how it grows.
We will keep tracking. You should keep watching.
Data: BLS, Eurostat, OECD, ONS, ABS, Statistics Canada, e-Stat, Adzuna. Risk model updated monthly. This is quantitative analysis, not financial advice.