Both AI bulls and bears are watching the wrong number. GDP measures output. AI's first impact is on inputs. The real story is firm-level productivity dispersion that macro data structurally cannot capture.
The IMF's April 2026 World Economic Outlook formally lists "disappointment over AI-driven productivity" as a downside risk to global growth, alongside war and trade tensions. Goldman Sachs chief economist Jan Hatzius said AI contributed "basically zero" to U.S. GDP in 2025. The consensus interpretation: $700 billion in AI spending has produced nothing.
The same week, Snap reported that 65 percent of its code is now AI-generated. The company cut 1,000 employees, reduced its annualized cost base by $500 million, and the stock jumped. Block is targeting $2 million in gross profit per employee in 2026, double last year's level, with a workforce shrunk below 6,000. Meta deployed 50 AI agents across 4,100 code files and produced 59 context documents that reduced agent tool calls by 40 percent. OpenAI hit $25 billion in annualized revenue. Anthropic reached $19 billion.
Both numbers are accurate. Both are measuring different things. The macro data says AI is not working. The firm-level data says AI is working so well that companies are firing the people it replaced and pocketing the savings. The disagreement is not about timing. It is about the instrument.
The Wrong Thermometer
GDP measures the monetary value of goods and services produced. When AI makes a company produce the same output with fewer workers, fewer hours, and lower costs, the immediate GDP effect is zero or negative. The output stays the same. The inputs shrink. The savings appear as higher margins, not higher GDP. A company that replaces 1,000 engineers with AI agents and ships the same product faster has improved productivity by every meaningful definition. GDP registers nothing.
This is not a temporary measurement lag. It is a structural mismatch between what GDP measures and what AI changes. Hatzius identified one channel: AI hardware is imported, so the investment adds to Taiwanese and Korean GDP, not American. But the deeper issue is that GDP was designed to measure an economy where growth meant more output. AI produces growth through less input. The thermometer is calibrated for fever. The patient's temperature is dropping.
Q4 2025 GDP was revised from an advance estimate of 1.4 percent to a final reading of 0.5 percent. Markets reacted to a number nearly three times the actual figure. The revision matters less than what it reveals: macro data does not just miss AI productivity. It actively misleads in real time.
The Dispersion Event
The AI productivity debate has two sides. Bulls say AI will transform everything. Bears say AI is a bubble producing nothing. Both share the same assumption: that AI's impact should be visible in aggregate statistics. It will not be, because AI is creating a dispersion event.
Dispersion means the gap between firms that deploy AI effectively and firms that do not is widening faster than any aggregate can capture. Snap's 65 percent AI-generated code and Block's doubling of per-employee productivity are not representative. They are outliers. But they are outliers in one direction only. No company is reporting that AI made its workforce less productive. The distribution is skewing, not shifting.
Paul David documented this pattern with electricity. The dynamo was introduced in the 1880s. Measurable productivity gains did not appear in manufacturing statistics until the 1920s, roughly four decades later. The delay was not technological. Early factories simply bolted electric motors onto existing steam-driven layouts. Productivity only materialized when a new generation of managers redesigned factories around the capabilities of distributed electric power: individual motors per machine, layouts following material flow, natural light replacing the windowless boxes that belt-drive systems required.
The AI parallel is exact. The 56 percent of CEOs reporting zero financial returns from AI are the factories that bolted a motor onto the belt drive. Snap, Block, and Meta are the factories that redesigned the floor plan. The aggregate statistic averages the two groups and reports nothing interesting.
Here Is What to Do
If you are investing based on macro data saying AI is not working, you are making the Solow mistake for the third time. Solow identified the productivity paradox with computers in 1987. David explained it with electricity in 1990. The pattern is identical: general-purpose technology, aggregate measurement failure, eventual recognition that the instrument was wrong.
Long companies that publish measurable AI productivity metrics. Block reports gross profit per employee. Snap reports AI code generation rates and cost savings. Meta publishes engineering blog posts with quantified efficiency gains. These companies are showing their work. The market is rewarding them: Block surged 24 percent on its layoff announcement. Snap jumped 8 percent.
Short companies spending on AI with no published productivity metrics. The gap between "we are investing in AI" and "here is what AI produced" is the gap between the belt-drive factory and the redesigned one. Fifty-six percent of CEOs report zero returns. Their shareholders have not yet noticed.
The correct metric is not GDP growth. It is the standard deviation of productivity across firms in the same industry. When that number widens, the general-purpose technology is working. The aggregate just cannot see it yet.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)