One employee can do more work.
One developer can write more code.
One marketing team can create a hundred campaign variations instead of ten.
That logic is correct.
It is also incomplete.
AI is creating another kind of inflation at the same time.
Not inflation in consumer prices.
Inflation in the number of economic objects competing to be noticed, trusted, purchased, reviewed, and stored.
More songs.
More software.
More advertisements.
More reports.
More customer-service interactions.
More synthetic experts offering more synthetic answers.
The cost of producing many of these objects is falling sharply.
Their supply is moving in the opposite direction.
And when supply grows faster than attention and demand, the interesting investment question is no longer whether AI creates more output.
It clearly does.
The question is what happens to the game once everyone can produce almost unlimited output.
Start with the strangest number
In April 2026, Deezer said it was receiving almost 75,000 fully AI-generated tracks every day.
That is more than two million per month.
Those tracks represented roughly 44% of the music delivered to the platform each day, up from approximately 10,000 daily uploads in January 2025.
But they generated only 1% to 3% of total streams.
Deezer also reported that up to 85% of streams from fully AI-generated tracks in 2025 were detected as fraudulent and excluded from royalty calculations.
Those are Deezer’s own platform-level measurements, detailed in its April 2026 disclosure. The company reaffirmed the 75,000-upload estimate in June 2026.
Think about what this means.
Production has exploded.
Consumption has barely followed.
The platform is filling with tracks that are inexpensive to produce, attract relatively little genuine listening, and can be used to manipulate the payment system.
This is not merely a music story.
It is a preview.
Code is moving in the same direction
According to GitHub’s 2025 Octoverse report, developers pushed nearly one billion commits in 2025, an increase of 25.1% from the previous year.
As of GitHub’s August 2025 measurement point, more than 1.1 million public repositories imported an LLM software-development kit. Of those, 693,867 had been created during the previous twelve months.
GitHub is careful not to claim that AI caused all of this growth. Its data shows correlation, expanding developer participation, and increasing adoption of AI tools—not a clean causal estimate.
The full figures are in GitHub’s 2025 Octoverse report.
The amount of software being attempted, generated, modified, and shipped is increasing.
The difficult part is that more code does not automatically mean more valuable software.
A randomized study involving 96 full-time Google engineers found that access to three internal AI coding features reduced the time spent on a complex enterprise task by an estimated 21%.
The researchers emphasized that the confidence interval was large and that a controlled task using Google’s internal tools should not automatically be generalized to all software work. The study is available here.
An early-2025 randomized study from METR found almost the opposite.
Sixteen experienced open-source developers completed 246 tasks in mature repositories they knew well. Before the experiment, they predicted AI would reduce completion time by 24%.
Instead, allowing AI increased measured completion time by 19%.
The original study is available through METR.
But the story did not end there.
In February 2026, METR reported that raw results from a later experiment suggested productivity improvements of roughly 4% to 18%, depending on the participant group. METR did not regard those estimates as reliable because developers who benefited most from AI were increasingly unwilling to participate in tasks where AI might be prohibited.
That methodological warning is important. The 2026 update suggests that newer agents may be more useful while also showing how difficult it has become to measure their effect cleanly.
The studies do not establish that AI always accelerates or always slows software development.
They suggest something more useful.
AI appears strongest when tasks are clearly specified, context is accessible, and outputs can be tested quickly.
Its benefits become less certain when work contains tacit knowledge, large codebases, implicit requirements, high quality standards, or expensive errors.
AI makes implementation cheaper.
That can make reviewing, debugging, securing, coordinating, and maintaining the implementation more economically important.
Services are beginning to inflate too
A field study covering 5,179 customer-support agents found that access to a generative-AI assistant increased issues resolved per hour by 14% on average.
The improvement reached 34% among novice and lower-skilled agents, while the most experienced workers saw comparatively little benefit.
The research is available through the National Bureau of Economic Research.
This is normally presented as a productivity story.
And it is one.
But consider the second-order effect.
As companies gain the ability to produce customer responses, translations, reports, product descriptions, advertisements, legal drafts, and software prototypes at much lower cost, the market does not merely become more efficient.
It becomes more crowded.
The supply of acceptable output rises.
Differences between an average producer and a good producer may become harder to observe.
And the cost of checking the work may not fall as quickly as the cost of creating it.
That is where the economics become interesting.
The data tells us that production is changing.
It does not yet tell us who captures the value.
